Modélisation statistique de la solidification

colonnaire/équiaxe des alliages binaires

Alexandru Ciobanas

To cite this version:

Alexandru Ciobanas. Modélisation statistique de la solidification colonnaire/équiaxe des alliages bi-
naires. Electromagnétisme. INSTITUT NATIONAL POLYTECHNIQUE DE GRENOBLE, 2006.
Français. �NNT : �. �tel-01338711�

HAL Id: tel-01338711

https://hal.science/tel-01338711v1
Submitted on 29 Jun 2016

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 INSTITUT NATIONAL POLYTECHNIQUE DE GRENOBLE

N° attribué par la bibliothèque

|__|__|__|__|__|__|__|__|__|__|

THESE

pour obtenir le grade de

DOCTEUR DE L’INP Grenoble

Spécialité : MECANIQUE DES FLUIDES ET TRANSFERTS

préparée au laboratoire : ELABORATION PAR PROCEDES MAGNETIQUE

dans le cadre de l’Ecole Doctorale : MECANIQUE ET ENERGETIQUE

présentée et soutenue publiquement

par

Alexandru CIOBANAS

le 15 septembre 2006

Modélisation statistique de la solidification colonnaire/équiaxe

des alliages binaires

Directeur de thèse:
Yves FAUTRELLE

JURY

M. H. COMBEAU , Président
M. CH.-A. GANDIN , Rapporteur
M. D. GOBIN , Rapporteur
M. Y. FAUTRELLE , Directeur de thèse
Mme. A.M. BIANCHI , Examinateur
 à ma chère Nicole

« Savoir que l'on sait ce que l'on sait, et savoir que l'on ne sait pas
ce que l'on ne sait pas: voilà la véritable science. »
(Confucius)

« La statistique est la première des sciences inexactes. »

les frères Goncourt
 Remerciements

Aucun manuscrit de thèse ne serait complet sans les mots de remerciements apportés aux
personnes qui ont contribué à son accomplissement. Car une chose est sûre, tout au long
de ces longues années je n’étais pas seul derrière ce projet de thèse. De nombreuses
personnes ont été à mes côtés, m’ont encouragé et m’ont finalement aidées à accomplir
un travail difficile.

Mes plus vifs remerciements vont à Yves Fautrelle pour son soutien
inconditionnel dans tout ce que j’ai entrepris. Vous m’avez toujours fais confiance et
vous avez su diriger mon travail avec une grande compétence. Merci pour tout Yves.

Je remercie Madame Ana-Maria Bianchi qui m’a toujours soutenu dans mes
démarches et qui a toujours cru en moi. Sans votre soutien ce projet de thèse n’aurait
certainement pas existé.

Je remercie aussi Adrian Bejan pour avoir accepté de réviser une partie de mon
travail et pour les très bonnes critiques apportées. Heureusement que la théorie
constructale existe !

Merci à Christian Trassy qui m’a accueilli au sein du laboratoire et ceci depuis
mon DEA puis pour l’ensemble de cette thèse.

Mes remerciements vont aussi à chacun des membres du jury et plus

particulièrement à Charles André Gandin et Dominique Gobin pour avoir accepté d’être
les rapporteurs de ce travail.

Merci à tout celles et ceux qui ont été à mes côtes tout au long de cette thèse. Je
remercie Anne Noeppel pour son aide précieuse et sa grande patience, Ghislain Quillet
pour mes premières leçons de Fluent, Fouad Khaldi pour les agréables moments passés
ensemble, Yves Delannoy pour ses conseils utiles, Yves Duterrail pour sa disponibilité et
sa bonne humeur, Anne et Pascale pour leur aide dans mes démarches administratives. Je
tiens à remercier aussi Monsieur René Moreau pour ses mots d’encouragement et pour
ses conseils scientifiques pertinents.

Mais s’il y a des personnes que je voudrais remercier particulièrement, ce sont

sans doute les membres de ma famille. Je remercie de tout mon cœur Nicole pour les si
beaux mots d’encouragement qu’elle a su m’apporter dans les moments difficiles de ces
dernières années et pour son soutien permanent dans tous mes projets. Je voudrais aussi
remercier mes parents qui, malgré la grande distance qui nous a séparé, ont été à côté de
moi tout au long de ces cinq dernières années.

Merci à tout celles et ceux que j’aurais pu oublier mais qui ont eux aussi apporté
une contribution à l’accomplissement de cette thèse.
 Table des matières

1 CONTEXTE DE L’ÉTUDE.................................................................................................................. 2
1.1 GÉNÉRALITÉS.................................................................................................................................. 2
1.2 LE PROJET MAP MICAST .............................................................................................................. 3
2 PROBLÉMATIQUE.............................................................................................................................. 5
2.1 LA MICROSTRUCTURE : COLONNAIRE/ÉQUIAXE............................................................................... 5
2.1.1 La transition colonnaire-équiaxe .............................................................................................. 6
2.2 LES MACROSÉGRÉGATIONS ............................................................................................................. 7
2.3 L’APPROCHE MATHÉMATIQUE UTILISÉE .......................................................................................... 8
3 OBJECTIFS ......................................................................................................................................... 10
3.1 LES OBJECTIFS PRINCIPAUX DE LA THÈSE ...................................................................................... 10
3.1.1 L’élaboration d’un modèle mathématique pour les phénomènes de solidification.................. 10
3.1.2 L’implémentation du modèle mathématique dans le moteur de calcul CFD « Fluent © »....... 11
3.1.3 Validation du modèle............................................................................................................... 11
4 RÉSULTATS........................................................................................................................................ 13
4.1 LA MORPHOLOGIE DENDRITIQUE REGARDÉ DANS LA PERSPECTIVE DE LA THÉORIE CONSTRUCTALE
13
4.2 MODÈLE À PRISE DE MOYENNE STATISTIQUE POUR LA SOLIDIFICATION COLONNAIRE/EQUIAXE DES
ALLIAGES BINAIRES. .................................................................................................................................... 14
4.3 VALIDATION DU MODÈLE STATISTIQUE D’ENSEMBLE.................................................................... 16
4.4 EXTENSION DU MODÈLE STATISTIQUE POUR LE CAS CONVECTIF ................................................... 17
4.5 IMPLÉMENTATION DU MODÈLE MATHÉMATIQUE DANS FLUENT. VALIDATION. ............................. 18
4.5.1 Modèle mathématique.............................................................................................................. 18
4.5.2 L’implémentation des équations dans Fluent .......................................................................... 21
4.5.3 L’algorithme de résolution ...................................................................................................... 22
4.5.4 Convection naturelle vs. convection forcée ............................................................................. 23
4.5.4.1 Convection naturelle – cas 2D.................................................................................................. 23
4.5.4.2 L’influence de la convection forcée sur la solidification – cas 2D ................................................ 33
4.5.4.3 L’influence de la convection forcée sur la solidification – cas 3D ................................................ 41
4.5.4.4 Conclusions partielles ................................................................................................................... 49
4.5.5 Comparaison avec l’expérience ACCESS ............................................................................... 50
4.5.5.1 Dispositif expérimental ................................................................................................................. 50
4.5.5.2 Paramètres numériques ................................................................................................................. 51
4.5.5.3 Résultats numériques et comparaison............................................................................................ 52
4.5.5.4 Conclusions partielles ................................................................................................................... 60
5 CONCLUSIONS FINALES ................................................................................................................ 61
6 RÉFÉRENCES..................................................................................................................................... 67
7 ANNEXES ............................................................................................................................................ 69
7.1 LES PROPRIÉTÉS PHYSIQUES DES ALLIAGES UTILISÉS .................................................................... 69
7.2 LES RÉFÉRENCES EN ANGLAIS ....................................................................................................... 71

1
1 Contexte de l’étude
1.1 Généralités
Les alliages métalliques sont presque toujours élaborés à l’état liquide, et ils doivent donc
subir une solidification pour prendre les formes solides désirées. Les transformations
subies par l’alliage lors du processus de solidification ont une influence extrêmement
importante sur les propriétés d’usages mécaniques et électriques/électroniques du solide
final.
Les processus de solidification utilisés dans l’industrie pour l’élaboration des
alliages sont dans leur majorité assez rapides, voire très rapides, ce qui favorise les
phénomènes de déstabilisation de l’interface solide-liquide plane. Les conséquences de ces
déstabilisations donnent naissance à des structures solides très spécifiques aux phénomènes
de solidification, nommés structures dendritiques. Elles peuvent prendre la forme de
colonnes orientées le long de la direction de l’extraction du flux de chaleur : dans ce cas on
les appelle dendrites colonnaires ou bien la forme des grains quasi-sphériques distribués de
façon plus ou moins homogène dans le creuset : ce sont les dendrites équiaxes (voir Figure
1).

Figure 1 : Les trois zones caractéristiques qui peuvent apparaître pendant la solidification
d’un alliage binaire: 1-le solide ; 2-la zone colonnaire ; 3-la zone équiaxe.

Prenons par exemple le cas d’un alliage Al-Si très utilisé dans l’industrie
aéronautique là où les propriétés mécaniques ont un enjeu très important. Parce que les
propriétés mécaniques d’un alliage dépendent fortement de sa microstructure (les
microstructures du type équiaxe ont une meilleure résistance et élasticité que celles du type
colonnaire) on comprend l’intérêt de l’obtention d’un alliage à structure équiaxe et en
particulier une structure équiaxe plus affinée (avec des grains dendritiques nombreux et
petits).
En outre, les corps solides finaux présentent aussi des défauts qui se manifestent
sous la forme d’hétérogénéités de la concentration du soluté. Les écarts de concentration
du soluté se manifestent à l’échelle des dendrites - 10 µm ÷ 100 µm - (les « ségrégations
dendritiques »), mais aussi à l’échelle du lingot (des « macroségrégations » ou des
« mesoségrégations »). Si les premiers défauts peuvent être éliminés par un traitement

2
thermique post-solidification, les derniers persistent encore et font que la composition
locale de l’alliage peut sensiblement s’écarter de la composition nominale. En particulier
dans certains produits (lingots de refusion à l’arc sous vide, ou aubes de turbine
unidirectionnelle) les macroségrégations peuvent apparaître sous forme de « canaux ».
Certaines sections du produit présentent des taches isolées dites « freckles » (taches de
rousseur) (voir Figure 2). Les alliages obtenus deviennent de cette manière impropres pour
des applications à hautes caractéristiques.

Canal ségrégé

Freckle

Figure 2 : Les canaux ségrégés et les freckles correspondants dans lingot d’un alliage Pb-
Sn 10% en masse, Sarazin et Hellawel [1]

À cause des différents phénomènes de transport qui apparaissent et qui sont

fortement couplés, la solidification des alliages même binaires se trouve être un problème
très difficile à résoudre. La solidification représente un carrefour pour différentes
disciplines et c’est ici qu’on retrouve à la fois des écoulements laminaires ou turbulents,
des transferts de masse et de chaleur, des changements de phase et des réactions chimiques.
De plus il y a toujours les contraintes d’équilibre thermodynamique aux interfaces (le
diagramme de phase), une forte variation des propriétés physiques des métaux liquides
avec la température et l’opacité des matériaux. On s’aperçoit vite que, étant donné la
complexité des différents processus physiques qui entrent en jeu dans la solidification, les
analyses expérimentales et théoriques deviennent très difficiles à aborder, sauf pour des cas
simples et/ou des matériaux particuliers.
Dans cette optique, l’existence d’un modèle numérique de la solidification
fournissant des prédictions aux macro-échelles du procédé est indispensable. Il peut
apporter des informations nouvelles sur la solidification des matériaux multi-constitués,
notamment sur les effets des phénomènes de transport (soluté, masse) qui entrent en jeu
tout au long du processus de solidification. Ainsi, on peut espérer mieux comprendre les
phénomènes complexes qui donnent naissance à la coexistence de dendrites colonnaires et
équiaxes dans le même lingot, les facteurs qui influencent la transition colonnaire-équiaxe
et les macroségrégations. En effet, toutes les informations utiles obtenues avec cette
analyse auront des conséquences directes sur la qualité des matériaux alliés obtenus.

1.2 Le projet MAP MICAST

Ce travail entre dans le cadre d’un projet européen nommé MAP MICAST (Microgravity
Application Program « Microstructure formation in CASTing »), soutenu par l’Agence
Spatiale Européenne. Ce programme se situe dans la perspective d’améliorer à terme la
maîtrise des microstructures de solidification grâce à un brassage contrôlé obtenu par

3
application de champs magnétiques tournants ou glissants. Son objectif est d’élaborer un
modèle global du processus en s’appuyant sur des démarches expérimentales conjointes en
microgravité et au sol. En effet, alors qu’au sol la convection forcée est couplée de façon
complexe à la convection naturelle (thermique et surtout solutale), ainsi qu’aux
mouvements de décantation des particules et des cristaux dans le cas de la solidification
équiaxe, la microgravité offre une possibilité unique de séparer les mécanismes mis en jeu,
en partant d’une situation contrôlée par la diffusion, puis en appliquant des niveaux de
brassage croissants.
La contribution du laboratoire EPM dans ce programme MICAST concerne la
modélisation de la macroségrégation et de la convection forcée (brassage
électromagnétique) dans des creusets rectangulaires et cylindriques, ainsi que la réalisation
des expériences capables d’apporter des éléments de validation au modèle numérique.
Plus particulièrement, le travail de cette thèse porte sur la création d’un modèle
mathématique multiphasique du type Euler-Euler qui soit capable de modéliser à la fois la
solidification colonnaire (c) et celle équiaxe (e), mais aussi les éventuelles zones mixtes c/e
qui peuvent apparaître lors des solidifications dirigées. Un intérêt particulier sera mis sur la
transition colonnaire-équiaxe (voir références II et III, Annexe 7.2) mais aussi sur l’effet de
la convection forcée sur les macroségrégations. Le modèle mathématique valide pour une
solidification colonnaire a été ensuite implémenté dans le module diphasique du logiciel
commercial Fluent. En utilisant les résultats expérimentaux obtenus dans le cadre du même
projet MICAST on a aussi effectué une première validation semi-quantitative du modèle
mathématique développé.

4
2 Problématique
Avant de passer à l’énumération des objectifs de ce travail je présenterai quelques détails
sur les problèmes auxquels on veut traiter. Je commencerai tout d’abord par une brève
caractérisation des microstructures colonnaires/équiaxe, je poursuivrai ensuite par une
présentation des causes principales des macroségrégations et terminerai enfin par la
description de l’approche utilisée dans la modélisation numérique des phénomènes de
solidification.

2.1 La microstructure : colonnaire/équiaxe

Comme nous avons pu l’observer, lors d’un processus de solidification on peut distinguer
dans le creuset plusieurs zones spécifiques : le solide, la zone colonnaire ainsi que la zone
équiaxe (voir Figure 1). Les deux dernières nous intéressent particulièrement car c’est dans
ces deux zones que les macroségrégations prennent naissance. Les zones colonnaires et
équiaxes se caractérisent par la présence à la fois du liquide et du solide. On a donc à
traiter des milieux diphasiques.
Ces deux zones ont comme point de départ un même phénomène de nucléation. En
effet, en fonction des contraintes géométriques auxquelles les nuclei initiaux sont soumis
juste après leur naissance dans le liquide en surfusion, les germes peuvent évoluer sous
forme de dendrites colonnaires ou équiaxes.

Figure 3 : L’évolution des germes colonnaires (a) et équiaxes (b)

Les dendrites colonnaires qui naissent à partir de petits germes sur la paroi et puis
évoluent sous des contraintes géométriques provenant du fait que le solide reste « collé » à
la paroi. En effet pour les germes colonnaires qui restent collés à des structures solides
stationnaires (les parois) le champ de croissance est implicitement réduit à un angle solide
de 2π (voir Figure 3a). Ainsi les dendrites colonnaires évolueront vers des structures ayant
une dimension beaucoup plus grande que les deux autres, presque unidimensionnelles, si
on les rapporte à la dimension caractéristique du lingot (voir le zoom de la zone 2, Figure
1). Il est important de noter que cette direction caractéristique des dendrites colonnaires
correspond à la direction de l’extraction du flux de chaleur dans le lingot.
D’autre part il existe dans le liquide des germes équiaxes qui, pour des raisons
diverses (germination dans le liquide en surfusion se trouvant devant les dendrites

5
colonnaires ou fragmentation des dendrites colonnaires), arrivent à être emportés dans le
bain liquide et ainsi auront la possibilité de croître sous un angle solide de 4π (voir Figure
3b).
Les caractéristiques de ces deux structures vont déterminer des comportements
hydrodynamiques très différents. Le solide colonnaire est stationnaire, étant bloqué à la
vitesse des parois (dans la plus part des cas nulle), tandis que le solide équiaxe subit un
transport, produit à la fois par une éventuelle sédimentation mais aussi par les courants
convectifs dans le creuset. Dans le cas des structures colonnaires, étant donné qu’elles
restent collées à des structures stationnaires, il ne faudra pas résoudre une équation de
transport pour le solide. Par contre les choses se compliquent bien plus pour le cas des
structures équiaxes. Pour celles-ci, il est obligatoire de prendre en compte le transport de
solide équiaxe. A ce transport de masse il faut aussi coupler l’interaction solutale entre le
solide et le liquide l’entourant.

2.1.1 La transition colonnaire-équiaxe

Dans la plupart des cas de solidification des alliages les deux structures colonnaire et
équiaxe coexistent dans le lingot. En effet deux mécanismes principaux ont été identifiés
pour expliquer la coexistence colonnaire – équiaxe. D’une part on peut supposer que des
grains équiaxes germent devant le front colonnaire dans la zone de surfusion devant les
pointes colonnaires, d’autre part, suivant les indications de Jackson [2] on peut s’attendre
dans certaines conditions, notamment lors d’une évolution non-stationnaire du front
colonnaire, à ce que les grains colonnaires puissent subir des phénomènes de
fragmentation. Cette fragmentation doublée d’une convection au niveau du macrofront
colonnaire peut transporter les fragments solides devant le front colonnaire en les
transformant ainsi en germes équiaxes. Dans ce contexte il est inévitable qu’une zone de
coexistence entre les grains équiaxes et colonnaires apparaîtra dans le lingot. Cette zone de
coexistence colonnaire-équiaxe suppose aussi l’existence d’interactions complexes entre
les deux structures. En effet, dans certaines conditions la zone équiaxe se trouvant devant
le front colonnaire peut arrêter l’avancement des pointes colonnaires déclanchant la
transition colonnaire-équiaxe (CET). Ce phénomène a été étudié dans les dernières années
en essayant de quantifier le mieux possible les interactions entre la zone équiaxe et
colonnaire. Deux types d’interactions ont été identifiés. Le premier type, les interactions
« mécaniques », a été quantifié d’abord par Hunt [3]. En effet, en quantifiant la zone
équiaxe devant les pointes colonnaires par une fraction granulaire, Hunt a pu déterminer
les conditions d’apparition de la CET en supposant l’existence d’une fraction granulaire
équiaxe critique au dessus de la quelle le front colonnaire était stoppé dans son
avancement. Hunt [3] a aussi identifié les conditions pour lesquelles une zone quasi-
stationnaire mixte (colonnaire+équiaxe) pouvait être identifiée dans le lingot. Toutefois, le
grand désavantage de l’approche de Hunt réside dans le choix quasi-empirique des seuils
de la fraction granulaire équiaxe. De plus l’approche de Hunt [3] ne prend pas en compte
les interactions solutales entre les grains équiaxes. Le deuxième mode d’interaction
colonnaire-équiaxe de type solutale, a été quantifié pour la première fois par Martorano et
al. [4]. En effet à cause des interactions solutales entre les grains équiaxes le liquide qui se
trouve entre les grains équiaxe subit un enrichissement en soluté. Comme les pointes
colonnaires avancent à travers la zone équiaxe l’enrichissement en soluté du liquide inter-
granulaire équiaxe aura une très grande influence sur l’avancement des pointes
colonnaires. Dans certaines conditions les interactions solutales entre la zone équiaxe
devant les pointes colonnaires et la zone colonnaire sont si fortes que la zone colonnaire se
voit complètement stoppée. En quantifiant d’une manière rigoureuse les interactions
solutales entre les grains équiaxes et l’enrichissement du liquide inter-granulaire équiaxe

6
Martorano et al. [4] a pu déterminer avec précision les conditions pour lesquelles la
transition colonnaire-équiaxe avait lieu. Toutefois le désavantage majeur de ce modèle est
l’incapacité de modéliser les interactions mécaniques entre les deux structures ainsi que
l’existence d’une éventuelle zone mixte colonnaire+équiaxe.
Dans ce contexte l’existence d’un modèle qui pourrait modéliser rigoureusement les
interactions à la fois mécaniques et solutales entre les structures colonnaires et équiaxes
serait d’un grand intérêt. Une partie de cette thèse (voir références II et III, Annexe 7.2)
sera en effet consacrée à la modélisation de la coexistence des structures colonnaires et
équiaxes ainsi qu’à la modélisation de la transition colonnaire-équiaxe. En effet, en
utilisant la méthode de la prise de moyenne statistique nous sommes parvenus à modéliser
rigoureusement les interactions mécaniques et solutales entre les deux structures et par
conséquent à mieux approfondir le phénomène de la transition colonnaire-équiaxe.

2.2 Les macroségrégations

En ce qui concerne la cause des ces défauts, il est déjà bien connu que le mouvement du
liquide inter-dendritique pendant la solidification est le principal motif pour les grandes
redistributions de soluté dans la zone pâteuse. Dans le cas d’une solidification habituelle
sous champ de gravité (sans brassage magnétique) les canaux apparaissent dans la masse
solidifiée comme des traits étroits alignés parallèlement à la direction de la gravité g , les
« freckles ». Il existe différentes études expérimentales et numériques sur la formation des
« freckles » dans les alliages binaires et les conclusions sont quasi unanimes pour dire que
ce sont des phénomènes d’instabilité hydrodynamique du liquide enrichi en soluté se
trouvant dans la zone pâteuse supérieure qui mènent à l’apparition des canaux ségrégés.

a) b)
Figure 4 : Canaux ségrégés formé dans la zone pâteuse :. a) Le champ des vitesses ; b) La
fraction solide (alliage Pb-10% en masse Sn)

Par contre dans le cas d’une solidification faite sous l’influence d’un brassage forcé
les macroségrégations sont principalement causées par l’écoulement forcé existant dans la
zone pâteuse. En effet, comme on le verra plus loin, le brassage forcé du liquide va
significativement influencer l’écoulement dans la zone pâteuse. Ceci va déterminer
l’apparition des canaux ségrégés dans la zone pâteuse et par conséquent l’apparition des
macroségrégations radiales importantes. Dans la Figure 4 on peut voir la présence de deux
canaux ségrégés dans la zone pâteuse colonnaire et l’écoulement correspondant. Les

7
résultats sont obtenus avec le nouveau modèle mathématique développé et implémenté en
Fluent pour un alliage Pb-10% en masse Sn. Le brassage électromagnétique se fait à l’aide
d’un champ électromagnétique glissant. En effet, dans ce cas le mouvement forcé dans le
bain liquide influence fortement la solidification dans la zone pâteuse et redistribue
spatialement le soluté (Cu) en déterminant l’apparition des macroségrégations radiales.
Comme phénomène de transport, les canaux sont donc un résultat de la convection dans la
zone pâteuse (« the mushy zone »).

2.3 L’approche mathématique utilisée

Notre but est donc d’essayer de modéliser l’évolution hydrodynamique, thermique, et
chimique d’un mélange binaire fondu, pendant sa solidification. Le système à traiter (le
creuset) est composé à la fois du liquide et du solide. On a alors à faire à un milieu
diphasique.
Il serait nécessaire dans l’absolu de résoudre directement le système d’équations
locales décrivant le phénomène de solidification : conservation locale de masse de QDM ;
d’énergie, de soluté. Malheureusement une telle résolution n’est pas envisageable en raison
de la présence des interfaces solide - liquide très complexes, aussi bien dans la zone
pâteuse où la croissance colonnaire a lieu mais également dans le bain liquide où de
nombreux grains équiaxes existent (voir Figure 1).
Une solution avantageuse consiste donc à appliquer une procédure de prise de
moyenne au système d’équations locales. De cette façon, on obtient un système
d’équations moyennées, dont la résolution fournit des champs physiques moyens
dépourvus des informations sur les fluctuations. Par contre, les fluctuations seront prises en
compte par un modèle de fermeture en vue de modéliser leur influence sur les champs
moyens (par exemple l’effet du tenseur de Reynolds sur le mouvement moyen).
Mais quelle prise de moyenne doit-on utiliser ? Parmi les méthodes les plus
utilisées dans la modélisation macroscopique de la solidification nous trouvons : la prise de
moyenne volumique, les méthodes variationnelles, les méthodes spectrales et la prise de
moyenne statistique d’ensemble. Toutefois la méthode utilisée presque exclusivement dans
les processus de solidification est de loin celle de la prise de moyenne volumique. La
méthode de prise de moyenne d’ensemble gagne de plus en plus de terrain, notamment par
rapport à ces avantages et possibilités de généralisation.
La méthode de prise de moyenne volumique est basée sur le concept d’un volume
élémentaire représentatif (VER). Ainsi toutes les grandeurs physiques dépendantes du
temps et de l’espace résultant de l’application de cette méthode sont des valeurs intégrées
spatialement dans ce VER. Il est important de noter que la prise de moyenne spatiale ne
peut être envisagée que s’il y a une séparation d’échelles entre l’échelle microscopique et
celle correspondante au VER. En pratique, cette condition n’est pas toujours vérifiée. Dans
le cas de la solidification colonnaire, par exemple, la condition est fortement mise en
défaut à cause de la présence des structures ségrégées très allongées (les canaux) dans la
zone pâteuse. En effet ces structures ont une dimension caractéristique longitudinale de
l’ordre de celle du lingot.
Lors de cette étude nous avons employé la méthode de la prise de moyenne
statistique d’ensemble en raison de ses nombreux avantages. Celle-ci a été utilisée dans le
développement d’un modèle eulérien multiphasique. Suivant Drew et Passman [5] on peut
définir la moyenne d’ensemble comme étant la moyenne qui permet l’interprétation des
phénomènes en termes de répétitivité des systèmes multiphasiques. Il faut noter que
chaque expérience (réalisation) d’un certain processus physique ne peut être répété parce
que les conditions initiales et limites diffèrent d’une expérience à l’autre. En effet, le

8
concept de la moyenne d’ensemble qu’on considèrera par la suite consiste à dire que
l’évolution du système est déterministe, mais que la caractéristique aléatoire du
mouvement provient de l’incertitude des conditions initiales. Alors chaque nouvelle
expérience donne naissance à un nouveau membre de l’ensemble de toutes les réalisations
possibles du phénomène physique en question. Dans le cas des écoulements
multiphasiques, cet ensemble est constitué par un nombre quasiment infini des réalisations
d’une même expérience.

Prenons à présent le cas de la solidification d’un alliage binaire dans un lingot

rectangulaire. La solidification d’un tel lingot fait intervenir de nombreux facteurs. Si l’on
regarde de façon plus précise les processus physiques qui gèrent la solidification d’un tel
lingot on s’aperçoit de la complexité du phénomène. Premièrement, comme nous l’avons
déjà évoqué au début de ce travail, il faut noter la présence de plusieurs phénomènes de
transport : de la QDM, de masse, d’énergie et du soluté. De plus tous ces phénomènes sont
très fortement couplés, aspect qui rend encore plus difficile le problème. Deuxièmement, il
faut aussi prendre en compte les phénomènes de changements de phase qui compliquent
encore le processus et qui font apparaître de nouveaux termes dans les équations de
transport de l’énergie et de soluté : ce sont les rejets de chaleur latente et de soluté à
l’interface solide - liquide. Enfin, pour compliquer encore plus la résolution du problème,
on est confronté à un système diphasique liquide - solide qui présente d’innombrables
interfaces liquide - solide d’une très grande complexité, voir les dendrites colonnaires et
équiaxes. Il faut aussi noter que le milieu diphasique qu’on veut traiter, présente plusieurs
échelles spatiales. En conséquence la multiplicité des échelles spatiales dans la zone
« pâteuse » rend inenvisageable une modélisation supposant la séparation d’échelles.
La moyenne d’ensemble présente des avantages très intéressants qui recommandent
vivement son utilisation dans l’étude des écoulements multi-constituants :
ƒ la moyenne d’ensemble reste très proche physiquement des aspects aléatoires des
phénomènes physiques décrits. De plus elle représente le point de départ mathématique
du concept de moyennage. Les autres types de prise de moyenne, comme par exemple
la prise de moyenne volumique ou celle temporelle, ne sont, en fait, que des solutions
approchées de la moyenne d’ensemble obtenues sous la condition de la validité des
hypothèses d’ergodicité des phénomènes. Ces hypothèses d’ergodicité nous permettent
de faire le passage de la moyenne d’ensemble aux moyennes volumiques et
temporelles. Ceci peut se faire, si et seulement si, pour les champs étudiés, les
hypothèses d’homogénéité d’une part et de stationnarité d’autre part, restent valides
(hypothèses qui ne sont pas toujours acceptables) ;
ƒ la définition de la moyenne d’ensemble ne suppose pas la définition d’un volume
élémentaire caractéristique (VER). Par contre pour que la prise de volume ait un sens il
faut que le VER contienne un grand nombre de particules c'est-à-dire qu’il y ait une
séparation d’échelle entre l’échelle du VER et l’échelle de la plus petite particule.
Ainsi, la moyenne d’ensemble n’est pas conditionnée par l’existence d’une hypothèse
de séparation d’échelle quelle qu’elle soit : spatiale ou temporelle ;
ƒ le concept de moyennage d’ensemble est très simple. La moyenne statistique d’un
champ physique ψ en x et au temps t n’est plus que la moyenne des valeurs ψ ( x, t ; µ )
obtenues pour chaque réalisation µ du phénomène physique en question. La loi des
grands nombres suggère que si l’on utilise un nombre suffisamment grand
d’expériences indépendantes d’un même phénomène on tend vers la moyenne
d’ensemble qui a une limite réelle.

9
3 Objectifs
Les objectifs principaux fixés en début de thèse se sont affinés au fur et à mesure que le
travail a progressé.

3.1.1 L’élaboration d’un modèle mathématique pour les phénomènes

de solidification
Notre premier objectif consiste à créer un nouveau modèle diphasique eulerien. En effet,
l’ancien travail numérique élaboré par Ghislain Quillet [6] et par moi-même [7] dans le
cadre de la première étape MICAST a à sa base un modèle diphasique simplifié, développé
par Felicelli, Heinrich et Poirier [8]. Celui-ci a quelques limitations extrêmement
pénalisantes. Premièrement, ce modèle ne peut prendre en compte que la solidification du
type colonnaire, celle dans laquelle le solide reste stationnaire. Le solide en mouvement,
spécifique à la solidification équiaxe ne peut pas être traité avec ce modèle. Deuxièmement
ce modèle considère que le liquide dans la zone pâteuse est dans un état de mélange parfait
ce qui implique que ce liquide se trouve à une concentration en soluté égale avec la
concentration du liquide à l’interface solide-liquide, qui s’exprime explicitement en
fonction de la température locale par l’intermédiaire de la courbe du liquidus. Cette
caractéristique ne permet pas la modélisation des phénomènes de non-équilibre solutal
entre les interfaces et le liquide dendritique, phénomènes présents dans les parties
supérieures de la zone colonnaire ou dans les zones équiaxes. Enfin il faut ajouter les
approximations faites au niveau des termes d’advection dans l’équation de conservation de
mouvement pour le liquide ainsi que l’inexistence dans le modèle d’un traitement de la
germination et de la réaction eutectique.
Dans ce contexte nous avons voulu développer un modèle mathématique plus
complexe qui pallie tous ces inconvénients. Ainsi, on établit que le nouveau modèle doit
avoir les caractéristiques suivantes :
i. La formulation mathématique doit prendre la forme d’un modèle diphasique du type
eulerien avec des équations de conservation écrites pour chaque phase et
comprenant des lois d’échange entre les phases.
ii. L’obtention du modèle mathématique macroscopique doit être faite par une prise de
moyenne statistique d’ensemble en évitant ainsi les incertitudes liées à l’hypothèse
de la séparation d’échelles micro-macro.
iii. Le système d’équations doit être valable aussi bien pour la solidification colonnaire
que pour celle équiaxe. Les différences entre les deux traitements se manifestent
seulement par une particularisation des coefficients d’échanges entre les phases ; par
contre les équations garderont la même forme. Cette caractéristique est très
importante car elle nous permettra d’aborder dans le même domaine physique et
avec les mêmes méthodes numériques les deux types de solidification, équiaxe et
colonnaire, rendant envisageable un traitement numérique de la transition
colonnaire-équiaxe (CET).
iv. Le modèle prend en compte le non équilibre solutal entre l’interface solide-liquide
et le liquide dendritique.
v. Le modèle prend en compte la séparation d’échelle entre le liquide inter-dendritique
et le liquide extra-dendritique.
vi. L’existence d’un modèle de germination, rendant possible le traitement des zones
équiaxes.

Les objectifs que l’on s’est fixés au début de cette thèse se sont au fur et à mesure
affinés. Ces modifications se réfèrent principalement à l’idée de l’utilisation du modèle

10
numérique colonnaire/équiaxe dans le but de prédire les phénomènes de transition
colonnaire-équiaxe. En effet, sa formulation statistique en fait un candidat bien placé pour
un tel objectif, mais il y a toutefois des obstacles à surmonter. Les difficultés majeures sont
liées principalement au fait que la prédiction de la transition colonnaire-équiaxe nécessite
une détermination précise de l’évolution de l’interface entre les zones colonnaire et
équiaxe. Ceci nécessiterait l’utilisation d’une technique de suivi de front très coûteuse en
temps de calcul. En essayant de contourner ce problème on a commencé à travailler sur un
modèle multiphasique qui suppose l’existence dans un même endroit des structures
équiaxes et colonnaires. Ce concept part du concept statistique de l’existence d’un
ensemble contenant toutes les réalisations possibles d’un même processus de solidification.
En effet il ne peut y avoir que pour un point P du domaine une partie des réalisations du
processus telle que le point P se trouve au sein d’une structure colonnaire. L’autre partie
des réalisations possibles du processus est telle que le point P se trouve au sein d’une
structure équiaxe. Ainsi, on écrit des équations moyennes (dans le sens statistique
d’ensemble) qui prennent en compte ce phénomène de coexistence des structures
colonnaires et équiaxes.

3.1.2 L’implémentation du modèle mathématique dans le moteur de

calcul CFD « Fluent © »

« Fluent © » est un logiciel commercial spécialisé dans la modélisation numérique en

mécanique des fluides. Il a à sa base une méthode à volumes finis, méthode qui fait de
celui-ci un très bon et robuste outil pour la prédiction des écoulements bi et
tridimensionnels que ce soit en régime laminaire ou turbulent.
Il dispose en même temps d’un nombre considérable de modèles pour la
turbulence, mais également de modèles pour la combustion, pour les écoulements
diphasiques et même d’un modèle pour la solidification des corps purs. En outre, ce
logiciel comprend trois modules de résolution numérique spécialisés dans le domaine des
écoulements multi-phasiques : un modèle « volume du fluide », VOF ; un modèle
diphasique de « mélange » et enfin, le modèle le plus complet, le modèle diphasique
eulérien. En considérant ses possibilités à la fois de généralisation et de particularisation on
a décidé d’implémenter le modèle mathématique dans ce dernier module.
Un des autres points forts de « Fluent » est la présence d’un logiciel auxiliaire –
« Gambit » - qui nous aide à définir la géométrie du problème et facilite le maillage de ce
domaine. On peut également mailler des couches limites et définir des maillages structurés
ou non-structurés.

3.1.3 Validation du modèle

Tout modèle numérique doit être vérifié en comparant ses résultats avec les résultats
expérimentaux. Dans cette direction nous avons envisagé dans un premier temps une
comparaison avec les résultats obtenus par Ghislain Quillet [6]. Ainsi, nous avons pu
vérifier le modèle dans ces aptitudes à modéliser la solidification colonnaire.
Dans le cadre du même projet MICAST il y aura une continuation et une
amélioration de l’expérience benchmark initialement développée par Ghislain Quillet [6].
En effet le travail expérimental est déjà commencé au sein du laboratoire EPM et il
aboutira au début de 2006 à une manipulation plus complexe que la précédente. Elle sera
plus complète sous l’aspect thermique mais elle comprendra aussi un brasseur
électromagnétique linéaire. On pense ainsi pouvoir obtenir des régimes de solidification
équiaxes et des zones de transition colonnaire – équiaxe. Malheureusement le manque de
résultats ne nous a pas encore permis de faire une comparaison directe entre cette

11
expérience et le modèle numérique. Toutefois nous avons pu comparer le modèle
numérique avec une autre expérience de solidification, celle développée par l’équipe
ACCESS à Aachen, Allemagne. La comparaison avec les résultats expérimentaux sera
présentée plus tard dans ce manuscrit.

12
4 Résultats
Nous présentons maintenant les résultats actuels. Dans un premier temps je présenterai le
résumé des quatre publications en anglais élaborées tout au long de la thèse. Ces
publications se trouvent attachées à ce document et portent sur :
ƒ la morphologie dendritique (espacement dendritique secondaire et primaire) vue dans
la perspective de la théorie constructale ;
ƒ l’élaboration du nouveau modèle mathématique capable de modéliser la coexistence
des structures équiaxes et colonnaires dans un cas purement diffusif ;
ƒ la modélisation unidimensionnelle de la transition colonaire-équiaxe et la validation du
modèle mathématique précédent ;
ƒ l’élaboration du modèle mathématique capable de modéliser la coexistence des
structures équiaxes et colonnaires dans un cas où la convection est prise en compte.
Dans un deuxième temps je présenterai les résultats numériques obtenus avec le modèle
purement colonnaire et la validation de celui-ci avec des résultats expérimentaux.

4.1 La morphologie dendritique regardé dans la perspective de la

théorie constructale
Dans ce premier papier (référence I, Annexe 7.2) on se concentre sur l’application de la
théorie constructale dans la prédiction de la structure dendritique solide. Premièrement on
analyse le critère de stabilité marginale vu dans la perspective de la théorie constructale.
En ayant comme principe de base la loi constructale on conclu que la forme de la dendrite
s’adapte de telle manière que la vitesse de croissance de la pointe dendritique soit
maximale. Dans ce contexte on a montré que, parmi le domaine des rayons possibles de la
pointe dendritique prédit par l’analyse de stabilité, la pointe dendritique va choisir le plus
petit rayon ça veut dire un rayon égal avec la plus petite longueur d’onde des perturbations
donnant naissance à des instabilités. On identifie aussi l’existence d’une compétition entre
la croissance contrôlée par la diffusion moléculaire et la croissance dendritique. En
analysant cette compétition on a été capable de déterminer le temps critique et le rayon
critique pour lesquels la déstabilisation de l’interface solide-liquide se déclanche. Dans un
deuxième temps on développe un modèle pour l’espacement dendritique secondaire. En
analysant la croissance d’une seule pointe dendritique on conclu que la croissance latérale
de la pointe dendritique est proche d’une croissance contrôlée par la diffusion moléculaire
et par conséquent est caractérisé par une vitesse de croissance qui diminue dans le temps.
Dans ce contexte la théorie constructale prédit une optimisation de la croissance latérale de
la dendrite. En faite on identifie encore une fois une compétition entre la croissance
latérale contrôlée par la diffusion moléculaire et la croissance dendritique. En analysant
cette compétition on a été capable de caractérisé le mécanisme de la croissance latérale des
branches dendritiques secondaires et de calculer finalement l’espacement dendritique
secondaire. Le résultat est en bon accord avec le modèle de J.S. Langer and H. Muler-
Krumbhaar, Acta Metallurgica 26, 1681 (1978) mais aussi avec des différents résultats
expérimentaux. Finalement, l’espacement dendritique primaire est analysé dans la
perspective de la loi constructale. Quand on l’applique à un front colonnaire la loi
constructale prédit que le seul moyen dont le pointes dendritique dispose pour
l’optimisation du processus de solidification est la minimisation de l’espace entre deux
pointes adjacentes, c'est-à-dire la minimisation du λ1 . Le mécanisme responsable pour la
diminution du λ1 a été identifié dans le mécanisme de division dendritique. Toutefois, à

13
cause des interactions solutales entre les pointes adjacentes il a été trouvé qu’il existe un
espacement critique λcr en dessous duquel la division des pointes colonnaire ne peut plus
se faire. Dans ce sens la loi constructale prédit que l’espacement primaire caractérisant
l’état quasi-stationnaire du front colonnaire sera proche de l’espacement stable minimum
λcr . En quantifiant la division dendritique et les interactions solutales entre les pointes
colonnaire on a finalement été capable d’obtenir un modèle pour l’espacement dendritique
primaire. En effet, on obtient que λ1 est de même ordre de grandeur avec λcr et plus
précisément que λ1 ≈ 1.5λcr . Le résultat est par la suite validé avec plusieurs donnés
expérimentales et on obtient que le nouvel modèle est en très bon accord avec les
expériences.

4.2 Modèle à prise de moyenne statistique pour la solidification

colonnaire/equiaxe des alliages binaires.
Au cours de la deuxième publication (référence II, Annexe 7.2) nous avons développé un
nouveau modèle diphasique eulérien valide pour la solidification colonnaire et équiaxe et
pour un cas purement diffusif (sans convection). Les équations moyennes ont été obtenues
à l’aide d’un moyennage statistique de phase et la formulation finale du modèle est valable
aussi bien pour une structure équiaxe que colonnaire. L’utilisation d’une telle prise de
moyenne a plusieurs avantages devant la prise de moyenne volumique, largement utilisée
durant les vingt dernières années. Premièrement l’approche statistique n’a pas besoin de la
définition d’un volume élémentaire représentatif (VER). Deuxièmement, cette approche est
cohérente avec l’aspect aléatoire de la germination des grains équiaxes et avec l’existence
possible d’un régime turbulent de l’écoulement fluide. Le facteur le plus important est que
nous sommes finalement capable, grâce à la nature statistique du modèle, de traiter
rigoureusement la coexistence des grains équiaxes et colonnaires et par conséquent la
transition colonnaire-équiaxe.
Nous nous proposons tout d’abord d’analyser le cas d’une structure colonnaire pure
ou équiaxe pure. A l’aide des fonctions de densité de probabilité des particules les plus
proches, f
(1)
( t , x; z ) , on est capable de calculer rigoureusement les termes et les tenseurs
1

inconnus existant dans les équations de conservation moyennes non-fermées. En effet, en

utilisant « l’approximation de la cellule » [5] pour ces fonctions de densité de probabilité,
les champs moyens peuvent être facilement calculés à l’aide de simples intégrations
volumiques dans une cellule sphérique caractéristique. La principale hypothèse physique
du modèle se réfère à la séparation d’échelle entre le liquide inter- et extra-dendritique.
Dans une approche similaire avec celui de Rappaz et Thévoz [9] et Wang et Beckermann
[10] le modèle de l’enveloppe dendritique est utilisé pour paramétrer les échelles de
longueur plus petites que la dimension de la dendrite, c'est-à-dire l’espacement inter-
dendritique. L’enveloppe dendritique est définie comme la surface virtuelle enveloppant la
dendrite qui connecte toutes les pointes dendritiques de la dendrite. Ce procédé nous
conduit à distinguer deux types de liquide, le liquide inter-dendritique et le liquide extra-
dendritique. Le liquide inter-dendritique est supposé à l’état d’équilibre avec l’interface, sa
concentration étant uniquement déterminée par la température locale et le diagramme de
phase. Par contre le liquide extra-dendritique peut avoir une concentration différente de
celle à l’équilibre. Le modèle de la maille couplé avec les hypothèses valides à la échelle
du grain nous permet de fermer le système des équations de conservation moyennes. Le
modèle final valide soit pour une structure colonnaire soit pour une structure équiaxe est
similaire à celui obtenu par Wang et Beckermann [10]. L’équivalence entre les deux

14
modèles est une conséquence directe de l’utilisation de l’hypothèse de l’équilibre
thermique local. En effet, si l’hypothèse d’homogénéité thermique locale est prise en
compte alors la moyenne statistique d’ensemble reproduit l’approche de la prise de
moyenne volumique. Toutefois le moyennage statistique est fondamentalement différent
du moyennage volumique. En effet, si l’hypothèse d’homogénéité thermique locale n’est
plus valide, l’utilisation de la prise de moyenne statistique donne naissance à des
expressions de fermetures différentes de celles obtenues par l’application du moyennage
volumique. De plus, les résultats obtenus avec le moyennage volumique ne sont pas
physiquement corrects. Dans ce contexte il est montré en référence II (Annexe 7.2) que
l’utilisation de la prise de moyenne statistique couplé avec le modèle de la cellule permet
avec succès l’intégration des effets des inhomogénéités locales (gradients locaux des
champs physiques) sur les expressions de fermeture.
Deuxièmement, le cas de la coexistence entre les structures colonnaires et équiaxes
(c/e) est analysé du point de vue du moyennage statistique. Plus précisément, le cas de la
pénétration du front colonnaire dans une zone équiaxe développée est analysé. Ce type de
couplage peut être fréquemment rencontré pendant les phénomènes de solidification et
spécialement dans le cas des phénomènes de transition colonnaire-équiaxe (CET). Le
principal avantage de la prise de moyenne statistique réside dans le fait qu’à l’aide des
( )
distributions de probabilité des particules les plus proches, f c / e t , x; z1 , on peut quantifier
d’une manière rigoureuse les influences des particules colonnaires/équiaxes (c/e) en x. En
utilisant l’approximation de la cellule pour f c / e deux cellules sphériques centrées en x et
de rayon R1c* et R1e* ont été identifiées. Chacune de ces cellules caractérise d’une manière
simple la statistique des deux structures (c/e) en x. En utilisant ces deux cellules il est
possible de séparer l’influence des structures c/e dans le point x en définissant des
moyennes d’ensemble correspondantes à chaque structure. Ces moyennes vont ainsi
quantifier les effets moyens en x produit par chaque structure. De plus, toute valeur
moyenne (ex. fractions c/e, fractions solide/liquide, moyennes de la concentration,
moyennes interfaciales, etc.) peut être facilement calculée à l’aide des simples intégrations
sur les positions possibles des grains c/e dans les cellules c/e. Ainsi, suivant le cas où la
particule la plus proche de x est colonnaire ou équiaxe, le modèle nous conduit à
différentier deux types principaux de valeurs moyennes : respectivement colonnaires et
équiaxes. Dans ce contexte deux fractions importantes sont identifiées : la fraction
colonnaire et celle équiaxe, ε c et ε e . Ces deux fractions quantifient la manière dont les
deux structures partagent l’espace en x. Comme les grains colonnaires et équiaxes
partagent le même espace en x la somme des ces deux fractions fait 1, ε c + ε e = 1 . On note
aussi que chaque structure sera caractérisée par sa propre fraction solide et liquide et par
conséquent la fraction c/e sera une somme entre les fractions solides et liquides
correspondantes, ε c / e = ε se / c + ε fe / c . Une autre caractéristique importante de la moyenne
statistique réside dans le fait qu’elle nous permet de quantifier le blocage mécanique
produit par la structure équiaxe sur les pointes colonnaires. On a démontré que les
structures équiaxes et colonnaires interagissent suivant un double mécanisme. Le premier
effet de blocage mécanique déterminera une raréfaction de la zone colonnaire au moment
où cette dernière pénètre dans la zone équiaxe. En effet, à cause de la dimension finie des
grains équiaxes les pointes colonnaires disposent seulement d’un espace limité (1 − ε ge )
dans lequel elles peuvent germer. Ainsi seulement une fraction, 1 − ε ge , du nombre initial
des colonnes, nc0 , est capable de pénétrer dans la zone équiaxe. Le deuxième effet de
blocage mécanique reflète le fait que pour une zone mixte (colonnaire+équiaxe), les deux

15
structures sont forcées de partager localement un même espace. Par conséquent, l’espace
disponible pour chaque structure sera différent et plus petit par rapport à l’espace
correspondant disponible juste avant l’état de coexistence. En fait, le deuxième blocage
mécanique dit que les rayons des cellules des deux structures, R1*c / e , caractérisant l’état de
coexistence sont plus petits que les rayons correspondants caractérisant les structures c/e
juste avant leur coexistence, R1c / e . Cet effet est d’une très grande importance parce que les
rayons R1*c / e peuvent être très différents des rayons originaux R1c / e et par conséquent
l’évolution de la zone colonnaire+équiaxe sera très différente de celle caractérisant une
zone purement colonnaire ou purement équiaxe.
Finalement, en utilisant le modèle de la cellule couplé avec les hypothèses physiques
valides à l’échelle du grain on a été capable de fermer rigoureusement le système des
équations moyennes. Un modèle complet décrivant la coexistence des grains colonnaires et
équiaxes est finalement obtenu.

4.3 Validation du modèle statistique d’ensemble.

Dans la troisième publication (référence III, Annexe 7.2) le nouveau modèle mathématique
qui a été développé au cours de la référence II (Annexe 7.2) est validé contre plusieurs
expériences de solidification dirigés. En effet, en raison de la nature statistique de la prise
de moyenne, le nouveau modèle est capable de modéliser rigoureusement la coexistence
des structures équiaxes et colonnaire et par conséquent de mieux approcher le phénomène
de la transition colonnaire-équiaxe (CET). Dans cette troisième publication on se concentre
premièrement sur la simulation numérique des solidifications quasi-stationnaires dans le
but d’obtenir les cartes de la CET. On a trouvé qu’en fonction du choix de la vitesse de
tirage par rapport à deux vitesses critiques caractérisant le zone équiaxe devant les pointes
colonnaires on peut identifier sur la carte CET trois zones : une zone purement colonnaire,
une zone purement équiaxe et finalement une zone mixte colonnaire+équiaxe. En
quantifiant le mécanisme de division des dendrites colonnaires on a pu déterminer si la
zone mixte colonnaire+équiaxe correspondait à un état quasi-stationnaire aux pointes
colonnaires. Ce mécanisme de division des colonnes (mécanisme de redensification de la
zone colonnaire) a été quantifié à l’aide du modèle de l’espacement dendritique primaire
(voir référence I, Annexe 7.2). La zone mixte colonnaire+équiaxe a été quantifiée à l’aide
de la fraction colonnaire ε c , qui à son tour quantifie rigoureusement l’état de coexistence
des deux structures. Le nouveau modèle nous permet aussi de quantifier aussi bien le
blocage mécanique que celui solutal produit par la zone équiaxe devant les pointes
colonnaires. En fait, parce que les effets de blocage solutal et mécanique sont
intrinsèquement inclus, le nouveau modèle unifie l’approche semi-empirique de Hunt [3]
(le blocage purement mécanique) et celui de Martorano et al. [4] (le blocage purement
solutal).
Dans un deuxième partie de la publication, le nouveau modèle est utilisé pour la
simulation des expériences unidirectionnelles de solidification en [11, 12]. On a ainsi
analysé le couplage complexe entre la zone colonnaire et la zone équiaxe devant les
pointes colonnaires. La stationnarité du front colonnaire par rapport aux conditions locales
de refroidissement a aussi été analysée. Cette analyse a montré que le front colonnaire
évolue dans un état de quasi-stationnarité jusqu’à très proche du moment de la transition
colonnaire-équiaxe. Elle a également montré que la surfusion de nucléation des grains
équiaxes est très proche de la surfusion maximale des pointes colonnaires et qu’ainsi la
CET est virtuellement indépendante de la zone équiaxe devant le front colonnaire. Ceci est
en concordance avec les résultats de Martorano et al. [4] et vient appuyer l’idée que la CET

16
est principalement causée par la fragmentation des dendrites colonnaires. Si la zone
équiaxe n’est pas prise en compte par le modèle on observe que la vitesse du front
colonnaire subit une croissance brusque au début de la solidification suivi par un quasi-
plateau correspondant à l’état quasi-stationnaire aux pointes colonnaires et finalement,
après un temps critique, une évolution oscillatoire. Il est important de noter le fait que le
début de l’évolution oscillatoire du front colonnaire a pu être corrélé avec la position de la
CET mesurée dans les expériences. L’évolution oscillatoire du front colonnaire correspond
en fait à un état non stationnaire de pointes colonnaires par rapport aux conditions locales
de refroidissement. Cette évolution oscillatoire du front colonnaire pourrait aussi être très
favorable à la fragmentation des dendrites colonnaires et par conséquent très favorable à la
CET. On conclut que le régime non stationnaire du front colonnaire par rapport aux
conditions locales de refroidissement représente la cause principale de la CET, au moins
pour les alliages non raffinés.

4.4 Extension du modèle statistique pour le cas convectif

Le modèle à prise de moyenne statistique développé dans la deuxième publication
(référence II, Annexe 7.2) est, dans la quatrième publication (référence IV, Annexe 7.2),
étendu pour prendre en compte les effets de la convection du fluide. Comme pour le
modèle initial le nouveau modèle approche la moyenne d’ensemble à l’aide des fonctions
( )
de probabilité les plus proches, f c / e t , x; z1 . De cette manière les interactions solutales et
mécaniques entre les deux structures (c/e) sont bien prises en compte par le nouveau
modèle. Toutefois, les hypothèses physiques valides à l’échelle du grain sont réévaluées
par rapport à celles initialement décrites en référence II, Annexe 7.2. Tour à tour on
analyse l’écoulement fluide autour des grains perméables colonnaires et équiaxes puis la
croissance du grain en présence de la convection et enfin le transfert de soluté autour des
grains c/e en tenant compte de la convection. On trouve ainsi que le transfert de masse à
l’interface du grain c/e en présence de la convection est plus intense par rapport au cas sans
convection. On propose des corrélations pour le calcul du coefficient de transfert de masse
global dans la limite des très petites et de très grandes fractions granulaires. Les hypothèses
physiques valides à l’échelle micro couplées avec le modèle de la cellule nous permettent
de fermer le système complexe formé par les équations de conservation. Tout d’abord les
équations de conservation des grains sont dérivées. La nucléation équiaxe est modélisée
comme un événement instantané tandis que la nucléation colonnaire (l’apparition soudaine
des pointes colonnaires dans un point x) est modélisé à l’aide d’un algorithme de suivi du
front. On propose ainsi une technique simple de suivi du front ayant à la base le modèle du
volume fluide (volume of fluid (VOF) model). De plus, le modèle de coexistence proposé
en référence II (Annexe 7.2) est ici étendu pour prendre en compte la convection des grains
équiaxes. En effet, les dimensions des cellules R1c* , R1e* et le calcul des fractions
colonnaires et équiaxes ( ε c et ε e ) sont ici déterminées à l’aide d’un nouveau modèle de
coexistence. Les équations de conservation de masse, de quantité de mouvement et de
l’énergie sont dérivées par la suite. Il faut remarquer que, pour des raisons de simplicité,
l’équation de conservation de quantité de mouvement pour le liquide est écrite pour le
fluide total et non pas séparément pour le fluide colonnaire et équiaxe. Pour l’équation de
l’énergie une seule équation de mélange est écrite. Finalement, les équations de
conservation de soluté sont dérivées. Ainsi deux équations pour le solide colonnaire et
équiaxe ainsi qu’une équation pour la conservation de soluté dans le fluide total sont prises
en compte. De plus, les deux équations de conservation de soluté dans le liquide inter-
dendritique sont utilisées pour le calcul du transfert de masse liquide-solide. On a aussi

17
proposé un algorithme pour le traitement implicite du calcul de la concentration dans le
liquide extra-dendritique. Ceci est fait par le couplage à chaque pas de temps de l’équation
de conservation de soluté dans le liquide extra-dendritique avec l’équation de conservation
de soluté dans le fluide total. Finalement un modèle complet à prise de moyenne statistique
d’ensemble est proposé, celui-ci étant capable de modéliser le couplage complexe entre la
structure colonnaire et équiaxe en présence de la convection.

4.5 Implémentation du modèle mathématique dans Fluent.

Validation.
Dans un but de validation, le nouveau modèle (référence IV, Annexe 7.2) doit d’abord être
implémenté dans un code de résolution spécialisé dans le calcul des écoulements et
transferts de chaleur et de masse en milieux multiphasiques. Le logiciel FLUENT a été
choisi car il a déjà été utilisé avec succès dans la résolution du modèle diphasique de
Felliceli [6]. FLUENT est un logiciel commercial spécialisé dans la modélisation des
écoulements fluides et de transfert de chaleur. Il utilise comme méthode de discrétisation la
méthode des volumes finis. Les avantages de ce code sont nombreux :
ƒ Il est associé à un logiciel très performant de maillage (Gambit) permettant ainsi de
varier facilement la géométrie étudiée ;
ƒ Le passage de 2D à un problème, 2D axisymétrique ou 3D se fait
automatiquement : il faut juste adapter le maillage ;
ƒ La robustesse et la fiabilité du solveur ont déjà été confirmées dans la littérature ;
ƒ Très important, FLUENT propose un module multiphasique qui utilise une
formulation eulérienne, donc très proche de la formulation du modèle développé en
référence IV (Annexe 7.2).

4.5.1 Modèle mathématique

Il est important de noter que pour l’instant seulement le modèle colonnaire pur a été
complètement implémenté et testé en Fluent. Une version du modèle équiaxe pur est en
phase de développement mais les résultats actuels ne permettent pas encore une validation
judicieuse de celui-ci. Il reste que dans la suite de ce travail, une fois que le modèle
équiaxe pur est testé, le modèle complet de la référence IV, permettant la simulation avec
couplage équiaxe/colonnaire, sera implémenté dans Fluent. Ceci se fera relativement
facilement dans le sens ou il faudra juste coupler les deux modèles colonnaires purs et
équiaxes purs, en tenant compte de la formulation déjà établie dans la référence IV.
Donc, par la suite, seuls les résultats obtenus avec le modèle colonnaire seront
présentés. Ce modèle mathématique s’obtient facilement à partir du modèle complet
présenté en référence IV en considérant uniquement l’existence d’une structure colonnaire
dans le lingot. Les équations formant ce modèle colonnaire sont résumées dans le Tableau
1.

Tableau 1: Équations constitutives du modèle purement colonnaire

Equations de Solide: Liquide:
conservation :
Grain ∂n ⎡4 3⎤
−1

= nc0 δ (T − Tliq ) , n = ⎢ π ( λ1 / 2 ) ⎥
0
c
∂t ⎣3 ⎦
Masse ∂ ( ρ sε s ) ∂(ρ fε f ) + div
∂t
= Γs
∂t
(ρ ε
f f v f ) = −Γ s

18
 QDM ∂ (ε f ρ f v f )
∂t ⎢⎣ (
+ ∇ ( ε f ρ f v f v f ) = −ε f ∇p + ∇ ⎡ε f µ f ∇v f + ( ∇v f )
T
)⎤⎥⎦ +
+ε f ρ f g − K f ε f v f + ε f Fé .m.
Energie ∂ ⎡⎣( ε s ρ s cs + ε f ρ f c f ) T ⎤⎦
+ div ( ε f ρ f c f v f ) =
∂t
= div ⎡⎣( ε s λs + ε f λ f ) ∇T ⎤⎦ + Γ s L
Soluté ∂ ( ε s ρ s Cs ) ∂ (ε f ρ f C f ) + div
= div ( ε s ρ s Ds ∇Cs ) + (ε ρ f v f Cf ) =
∂t ∂t
f

+Cs∗Γ s = div ( ε f ρ f D f ∇C f ) − Cs∗Γ s

Equations de
conservation
auxiliaires :.
C. masse dans le ∂ (ρ f εd )
liquide inter- = −Γ s + Γ g , Γ g = ρ f S g wg
dendritique ∂t
C. soluté dans le ∂Cl ⎛ ρ f S g Dl ⎞
liquide extra- εl ρ f + ε l ρ f v f ∇Cl = ∇ ⋅ (ε l ρ f D f ∇Cl ) + ⎜ − Γ g ⎟ ( Cl∗ − Cl )
dendritique
∂t ⎝ δ l −d ⎠
C. soluté dans le ∂Cl* ρ f S g Dl ∗
liquide inter- Γ s ( Cl* − Cs∗ ) = ε d ρ f + ( Cl − Cl ) + ε d ρ f v f ∇Cl*
∂t δ l −d
dendritique
Expressions
auxiliaires :
Contraintes: εl + ε d + ε s = 1 , εd + εs = ε g , εd + εl = ε f
Concentrations ⎧kC ∗ if Γs ≥ 0 T − Tf
interfaciales: Cs∗ = ⎨ l Cl∗ =
⎩Cs if Γs < 0 m
Rayon de la R1 = ( 4 / 3π n )
−1/ 3

cellule :
Rayon du grain : a = R1ε g1/ 3
Densité d’air de S g = 4π a 2 n
l’enveloppe du
grain :
Longueur de ⎡ 10a 2δ + 10aδ 2 + 4δ 3 ⎤
diffusion : δ l − d = δ ⎢1 − ⎥ où
⎢⎣ 5 ( R1 3 − a 3 ) ⎥⎦
⎧ Dl / w g
⎪ si (R 1 ( )
− a ) ≥ 2 Dl / w g
δ =⎨
⎪⎩( R1 − a ) / 2 si (R 1 − a) < 2( D / w ) l g

La croissance du Dl m ( k − 1) Cl∗
( Pet )
2
grain : Γ g = ρ f S g wg , wg =
π Γ2

Toutefois quelques simplifications ont été apportées par rapport à la formulation de

la référence IV, à savoir :

19
 ƒ Le coefficient de transfert e QDM à l’interface, K f , n’est pas calculé avec la
formule énoncée en référence IV. Ici une expression simplifiée pour K f est utilisée
qui fait intervenir la perméabilité au fluide isotrope de type Carman-Kozeny [13],
K *f :

µε f λ12ε f 3
Kf = , K *f = (1)
K *f π 2 K c t 2 (1 − ε f ) 2

où K c = 5 et t = 2 sont deux paramètres adimensionnels, λ1 est l’espacement

primaire de la structure colonnaire et ε f la fraction fluide. Cette loi a déjà été
utilisée par Ghislain Quillet [6] et a été choisie par la suite pour avoir une même
base de comparaison avec les résultats obtenus déjà en [6].
ƒ Comme dans l’approche de Ghislain Quillet [6], la surfusion de la pointe
colonnaire est considérée comme étant négligeable. Ainsi, on contourne le
problème de suivi du front colonnaire car la nucléation colonnaire est de cette façon
approchée par un événement local dépendant seulement des paramètres locaux
comme la température et la concentration fluide locale. En effet, la nucléation
colonnaire se déclenche instantanément dès que la température T est inférieure à la
température liquidus locale, Tliq ( = T fusion + mC f ) . Comme les cas de solidification
analysés par la suite sont caractérisés par des forts gradients thermiques
(G>1000K/m), la zone de surfusion devant les pointes colonnaires peut être
supposée négligeable.
ƒ L’effet de convection solutale et thermique est prise en compte par la suite par
l’approximation de Boussinesq : seule la densité qui apparaît dans la force
volumique d’origine gravitaire dépend linéairement de la température locale et de
la concentration locale :

ρ f = ρ 0f ⎡⎣1 + βT (T − T0 ) + β C ( C f − C0 ) ⎤⎦ (2)

où βT et β C sont respectivement les coefficients d’expansion thermique et

solutale.

Il faut noter que l’équation de conservation de soluté dans le liquide extra-

dendritique est utilisée en couplage avec l’équation de conservation de soluté dans le fluide
total. Comme nous l’avons expliqué dans la référence IV (Annexe 7.2), ceci est nécessaire
pour pouvoir traiter d’une manière implicite le rejet de soluté dans le liquide extra-
dendritique en obtenant ainsi un algorithme de résolution plus robuste. L’équation de
conservation de soluté dans le liquide extra-dendritique est discrétisée de la manière
suivante (référence IV) :

( Cl )i − ( Cl )
n n −1

Vcell ( ε l ρ f ) + (ε l ρ f v f ) ∑ (C ) n Sσ = ∑ ( ε l ρ f D f ∇Cl ) ⋅ nσ Sσ +
n n n n
l σ σ
i ∆t i
σ σ
σ

n
(3)
⎛ ρ f S g Dl ⎞
− Γ g ⎟ ⎡⎢( Cl∗ ) − ( Cl )i ⎤⎥
n n
+Vcell ⎜
⎝ δ l −d ⎠i ⎣ ⎦
i

20
où Vcell est le volume de la maille de discrétisation, σ la σ ème face la maille, nσ la
normale extérieure à la σ ème face, Sσ la surface de la σ ème face, n la nème itération
temporelle et i la i ème itération de convergence. On met facilement en évidence le
traitement implicite du rejet de soluté dans le liquide extra-dendritique, c’est-à-dire le
dernier terme du membre de droite de l’équation (3). Comme nous l’avons expliqué dans la
référence IV le couplage avec l’équation de conservation de soluté dans le fluide se fait à
chaque pas de temps par l’intermédiaire de ( Cl ) :
n −1

n −1
⎛ ε f C f − ε d Cl* ⎞
( Cl )
n −1
=⎜ ⎟⎟ (4)
⎜ εl
⎝ ⎠

Ainsi le taux de solidification Γ s se calculera de la manière suivante :

n n
⎛ ∂Cl* ⎞ ⎛ ρ f S g Dl ⎞ ⎡ ∗ n
( Γ s )i ( C )V ⎟ ⎢⎣( Cl )i − ( Cl )i ⎥⎦ Vcell +
n ∗ n n⎤
*
−C = ⎜εd ρ f ⎟ Vcell + ⎜
⎝ δ l − d ⎠i
l s i cell
⎝ ∂t ⎠i (5)
+ε d ρ f v f ∑ ( C * n
l i )n σ Sσ
σ

où ( Cl )i est calculé à l’aide de l’équation (3).

n

4.5.2 L’implémentation des équations dans Fluent

La formulation des équations diphasiques proposée par Fluent est très semblable à la
formulation des équations du Tableau 1. Ainsi, les équations de conservation de masse et
de QDM pour la phase solide et fluide peuvent être rentrées directement dans Fluent telles
quelles sont décrites en Tableau 1. Par contre il faut spécifier dans le logiciel Fluent, par
l’intermédiaire des fonctions utilisatrices (« User Defined Functions »), les sources et les
coefficients de transfert correspondants : la source de masse Γ s , les coefficients de
transfert de QDM à l’interface K f et les forces volumiques électromagnétiques ( Fé .m. ) ou
de gravité. En ce qui concerne les équations de conservation de l’énergie et de soluté,
celles-ci sont introduites dans Fluent à l’aide des équations de transport scalaires. En effet,
dans Fluent on a la possibilité de définir en dehors des équations de transport classiques
(masse, QDM) des équations de transport scalaires supplémentaires (les « User Defined
Scalars »). La forme générale d’une équation scalaire est la suivante :

∂ ( mψ )
+ div ( φψ ) = div ( D*∇ψ ) + S (6)
∂t

où ψ représente le champ physique transporté, m le terme de l’accumulation, φ le flux

massique, D* la diffusivité et S les éventuelles sources. On note que m, φ , D* et S sont
des paramètres à définir par l’utilisateur. Ceux-ci peuvent être facilement particularisés à
l’aide des fonctions utilisatrices (« User Defined Functions »). Ainsi l’équation de

21
l’énergie peut être facilement implémentée dans Fluent en particularisant ψ , m, φ , Dk* et
Sk comme suit :

ψ =T
m = ρ sε s cs + ρ f ε f c f
φ = ρ s ε s cs v s + ρ f ε f c f v f (7)
D* = ε s λs + ε f λ f
S = L Γs

L’équation de conservation de soluté en solide est introduite dans Fluent en

prenant :

ψ = Cs
m = ρ sε s
φ=0 (8)
D = ε s ρ s Ds
*

S = Cs* Γ s

Finalement l’équation de conservation de soluté dans le fluide est implémentée en

Fluent en considérant:

ψ = Cf
m = ρfε f
φ = ρfε f vf (9)
D* = ε f ρ f D f
S = −Cs* Γ s

4.5.3 L’algorithme de résolution

On rappelle que la méthode de discrétisation est celle des volumes finis. Ainsi toutes les
équations aux dérivées partielles du modèle sont résolues en les intégrant sur les cellules
du domaine de discrétisation. La méthode de discrétisation spatiale utilisée est celle
UPWIND du 2ème ordre avec un traitement intégralement implicite. La discrétisation
temporelle est celle d’Euler de premier ordre, implicite. Le système d’équations du
Tableau 1 est résolu selon un algorithme ségrégé ce qui signifie que toutes les équations
sont résolues d’une manière séquentielle et itérative. L’algorithme de résolution peut être
synthétisé de la manière suivante :

ƒ Etape I : Toutes les valeurs au pas de temps t n sont supposées connues ;

ƒ Etape II : L’incrémentation du pas de temps : t n +1 = t n + ∆t ;

22
 ( ε s )i = (ε s ) ,
n +1 n
ƒ Etape III : Entrée dans la boucle d’itération : i=0 ;

(ε ) = (ε f
) , p =p , v = ( Cs ) , ( C f ) = (C f )
n +1 n +1
= v n , Ti n +1 = T n , ( Cs )i
n n +1 n n +1 n +1 n n
f i i ,
i i

(ε ) = (ε ) , ( Γ ) = ( Γ )
n +1 n n +1 n
g i g s i s ;

ƒ Etape IV : Itération i + 1 :

o Calcul en utilisant la méthode de couplage pression-vitesses SIMPLEC de

(ε s )i +1 , (ε f )i +1 ,
n +1 n +1
pin++11 et v in++11 ;
o Résolution de l’équation scalaire UDS_0 représentant l’équation de
l’énergie. Ainsi on obtient : Ti +n1+1 ;
o Résolution de l’équation scalaire UDS_1 représentant l’équation de
conservation de soluté dans le fluide. Ainsi on obtient : ( C f )
n +1
;
i +1

o Résolution de l’équation scalaire UDS_2 représentant l’équation de

conservation de soluté dans le solide. Ainsi on obtient : ( Cs )i +1 ;
n +1

o Calcul de la croissance du grain : ( Γ g ) , (ε g )

n +1 n +1
;
i +1 i +1

o Calcul de la nouvelle concentration dans le liquide extradendritique ( Cl )i +1

n +1

en utilisant l’équation (3). On note le traitement implicite du rejet de soluté

dans l’équation (3) ;
o **Calcul de la rate de solidification ( Γ s )i +1 en utilisant l’équation (5) ;
n +1

Etape V : On vérifie si les valeurs ( ε s )i +1 , ( ε f ) , pin++11 , v in++11 , Ti +n1+1 , ( C f )

n +1 n +1 n +1
ƒ ,
i +1 i +1

** ( Cs )i +1 , ( Γ g ) , (ε g )
n +1 n +1
et ( Γ s )i +1 sont convergées. Si oui, on passe au pas de
n +1 n +1
i +1 i +1

temps suivant donc à l’étape II. Si non on continue les itérations en revenant à
l’étape IV.

4.5.4 Convection naturelle vs. convection forcée

Dans la suite du mémoire nous analyserons l’influence de la convection naturelle et forcée
sur la solidification des alliages binaires sera analysée. Le modèle mathématique présenté
ci-dessus sera résolu à l’aide du logiciel Fluent pour les différents cas 2D et 3D de
solidification colonnaire dirigée effectués sous l’influence de la convection naturelle et/ou
forcée. La convection forcée est le résultat des forces électromagnétiques produites par
l’application d’un champ électromagnétique alternatif. Les champs électromagnétiques
tournants et glissants seront analysés par la suite. Même si les cas traités par la suite ne
permettent pas une comparaison directe avec des expériences réelles, les résultats obtenus
permettront une validation partielle du modèle par une comparaison directe avec d’ autres
codes numériques mais aussi avec des ordres de grandeur théoriques. L’ensemble des
expériences numériques suivantes prend en compte un alliage binaire Pb-10% en masse
Sn.

4.5.4.1 Convection naturelle – cas 2D

Le cas analysé
Dans un souci de validation du code, le cas bidimensionnel analysé sera identique à celui
utilisé lors de plusieurs expériences numériques effectuées par Ghislain Quillet [6] et

23
Mabel Medina [14] en ayant à la base des modèles numériques différents du présent
modèle. On considérera par la suite un lingot bidimensionnel rectangulaire ayant 5 mm de
largeur et 10 mm de hauteur (voir Figure 5). En essayant d’approcher le plus possible le
cas d’une solidification dirigée, les conditions limites sont les suivantes :
ƒ Les parois latérales sont adiabatiques ;
ƒ Une température linéairement décroissante est imposée sur la paroi inférieure :

Tinf ( t ) = T0 − CR ⋅ t (10)

où CR est la vitesse de refroidissement [K/s] ;

ƒ Un flux entrant est imposé sur la paroi supérieure, flux correspondant au gradient de
température vertical initial, G0 :

ϕsup = λliq G0 ⎡⎣W .m −2 ⎤⎦ (11)

ƒ Le flux de masse à travers les parois est nul, ainsi que les flux de soluté.

Les conditions initiales quant à elles peuvent être résumées comme suit :
ƒ La température initiale sur la paroi inférieure, T0 , est égale à la température de liquidus
correspondante à la concentration initiale C0 ;
ƒ Le gradient de température initial est fixé à la valeur G0 ;
ƒ Vitesses nulles dans le liquide ;
ƒ La concentration de l’étain initiale dans le liquide est considérée comme étant très
proche de C0 = 10% , mais légèrement et aléatoirement perturbé par rapport à C0 :

C0 = C0 ± 0.005C0 (12)

On note que la configuration étudiée est thermiquement stable. Par contre, la

distribution de soluté dans la zone pâteuse est fortement instable à cause de la différence de
masse volumique entre étain et plomb. Ainsi, la distribution aléatoire de soluté autour de la
valeur C0 aura comme effet la génération d’un léger mouvement aléatoire dans le ligot qui
permettra la déstabilisation du macrofront et la création des éventuels canaux ségrégés.

24
Figure 5: Géométrie du cas 2D et conditions aux limites

Les conditions de solidification sont résumées dans le Tableau 2. Ces conditions de

solidification ont été choisies pour avoir une base cohérente de comparaison avec les
calculs de Ghislain Quillet [6]. D’ailleurs les paramètres numériques utilisés par la suite
(voir Tableau 3) sont eux aussi quasi-identiques à ceux utilisés en [6].

Tableau 2: Conditions de solidification - cas 2D

C0 = 10 wt% Sn
G = 1000 K / m
CR = 1 K / min
λ1 = 300 µ m

Tableau 3: Paramètres numériques - cas 2D

Maillage : 20 x 40
Pas de temps : 0,02 s
Schéma temporel : 1er ordre Euler
Couplage pression –vitesse : SIMPLE C
Schéma spatial : 2ème ordre UPWIND
Durée de simulation : 20 minutes

On remarque encore une fois que la configuration étudiée est thermiquement stable
et solutalement instable au regard de la distribution de soluté dans la zone pâteuse. En
effet, la zone pâteuse est enrichie en soluté par le rejet de soluté à l’interface solide-liquide
tel que rapidement derrière les pointes colonnaires, le liquide est proche du mélange
parfait, c'est-à-dire à une concentration très voisine de la concentration à l’équilibre Cl* .
Ainsi le gradient de soluté dans la zone pâteuse est négatif. Ceci ajouté au fait que l’étain
est plus léger que le plomb fait que la configuration est fortement instable. Toutefois pour
qu’une déstabilisation du macrofront plan apparaisse et s’amplifie il faut induire dans le
système une faible perturbation solutale qui, par la suite, sera amplifiée et donnera

25
naissance à un canal ségrégé. Cette perturbation n’est plus que la distribution initiale
aléatoire de soluté (équation (12)). Il faut remarquer aussi que la convection naturelle dans
le cas analysé est essentiellement de nature solutale et non thermique. Pour quantifier ceci
on peut déjà regarder le rapport des forces d’Archimède d’origine solutale et thermique :

∆ρC βC ∆C
N= =
∆ρT βT ∆T
(13)
β
= C = 19.5
βT m

en tenant compte des paramètres physiques de l’alliage Pb-Sn, voir Annexe 7.1. Le nombre
N s’appelle aussi le nombre de flottaison. On remarque ainsi l’importance des forces
d’Archimède de nature solutale face à celles d’origine thermique.
A présent si l’on admet l’existence d’un canal ségrégé dans la zone pâteuse on peut
aussi comparer les nombres de Rayleigh thermique et solutal. En sachant que dans le canal
la concentration est proche de celle de l’équilibre on peut écrire :

RaC ⎛ g β C ∆C l 3 ⎞ ⎛ g βT ∆T l 3 ⎞
=⎜ ⎟ ⎜ ⎟
RaT ⎝ ρ 0ν D ⎠ ⎝ ρ0να ⎠
(14)
β α
= C × = 3.4 ×105
βT m D

au regard des paramètres physiques de l’alliage Pb-Sn, voir Annexe 7.1. On remarque donc
une nouvelle fois une prépondérance de la convection solutale dans les éventuelles poches
liquides dans la zone pâteuse. Ce résultat a une conséquence très importante, car si une
déstabilisation du macrofront apparaît, créant une poche liquide dans la zone pâteuse,
celle-ci sera encore amplifiée par la forte convection solutale qui se forme. Dans ces
conditions, le canal ségrégé, une fois formé, s’auto entretient dans le temps par
l’intermédiaire de la convection naturelle qui lui est associée. Une nouvelle question se
pose alors : est-ce que la déstabilisation du macrofront colonnaire peut apparaître ou non ?
Pour répondre à cette question une analyse de l’écoulement d’origine solutale et dans le
voisinage du macrofront doit être faite. Ici, évidemment entre en jeu la perméabilité au
fluide de la zone pâteuse derrière les pointes colonnaires. L’analyse de Ghislain Quillet [6]
basée sur le modèle de Worster [15] a pu confirmer le fait que pour le cas analysé par la
suite il existe les conditions étaient remplies pour que l’écoulement au niveau du
macrofront devienne instable.

Résultats numériques
En regardant les résultats obtenus au niveau de la distribution de concentration dans le
liquide et de la carte de fraction solide (Figure 6a,b) à t=50s du début de la solidification,
on peut déjà observer deux instabilités au niveau du macrofront. Le mouvement ascendant
correspondant, les panaches de liquide enrichi, (Figure 6c) apporte du liquide riche en
soluté de la zone pâteuse au niveau du macrofront et y retarde la solidification formant
deux poches liquides (Figure 6b).

26
 a) b)

c) d)
Figure 6: La distribution a) de la concentration dans le fluide, C f ; b) de la fraction solide,
ε s ; c) de la vitesse ; d) de la température, à t = 50 s du début de la solidification.

a) b)

27
 c) d)

e) f)
Figure 7: La fraction liquide (a, b), la distribution de la concentration dans le fluide (c, d) et
la vitesse du liquide (e, f) ) à t = 200 s (a, c, e) et t = 600 s (b, d, f).

Au fur et à mesure que le macrofront avance, les zones correspondantes aux

panaches enrichies en soluté restent liquides et les poches liquides se transforment
progressivement en canaux ségrégés (Figure 7a, b). Il est donc important de remarquer que
le mouvement solutal ascendant est assez intense ( v 9 × 10−4 m / s ) et ne fait qu’ auto-
entretenir le canal ségrégé car il apporte du liquide riche en soluté de la zone pâteuse
inférieure vers les zones supérieures plus chaudes en empêchant ainsi la solidification dans
le canal. En effet le mouvement solutal dans le canal prélève du liquide à l’équilibre de la
zone pâteuse environnante et la concentre dans le canal. Ainsi le canal se retrouvera
toujours à une concentration proche de celle de l’équilibre, mais légèrement plus grande
(Figure 7c, d). Ceci est possible justement à cause de l’écoulement ascendant dans le canal
qui, même s’il est caractérisé par des vitesses faibles, est bien capable de déformer
localement les lignes iso-concentrations (Figure 7c, d). En effet, le nombre de Péclet
chimique dans le canal pour une vitesse moyenne dans le canal de l’ordre de 10−4 m / s
soit :

28
 Vc h
Pec =
Dl
(15)
10−4 × 0.005
= = 270
1.8 × 10−9

est grand. Donc, la convection dans le canal reste prédominante face à la diffusion
moléculaire. On peut ainsi assimiler le canal ségrégé à une pompe de soluté qui draine en
soluté la zone pâteuse environnante en l’appauvrissant. L’énergie motrice de cette pompe
est évidemment la convection naturelle d’origine solutale dans le canal même.
Effectivement cet effet de pompage de soluté a comme principale conséquence la déplétion
en soluté de la zone pâteuse environnante. En regardant la concentration moyenne à deux
instants de temps t = 600 s et t = 1200 s (Figure 8) on peut observer l’appauvrissement
important en soluté de la zone pâteuse adjacente au canal et l’enrichissement de ce dernier.

a) b)
Figure 8: La concentration moyenne à t = 600 s (a) et t = 1200 s (b)

Il serait intéressant d’essayer d’expliquer et de quantifier ce phénomène. En effet, la

déplétion en soluté de la zone pâteuse s’explique par le fait que le soluté apporté dans le
canal par le mouvement inter-dendritique à partir de la zone pâteuse adjacente n’est pas
compensé par la quantité de soluté qui entre dans la zone pâteuse par la partie supérieure.
Pour mieux comprendre ce phénomène il suffit de regarder le bilan de soluté d’une tranche
de zone pâteuse situé entre deux canaux adjacents et ayant une profondeur H assez grande
pour que l’écoulement inter-dendritique au niveau H soit négligeable (Figure 9).
L’équation de conservation de la concentration moyenne C = ε s Cs + ε f C f est :

∂ρ C
+ div ( ρε f C f v f ) = div ( ρε f Dl ∇C f ) (16)
∂t

En tenant compte de la très faible diffusivité chimique Dl et des vitesses liquides

relativement importantes dans la zone pâteuse derrière le macrofront ( v 10−4 m / s ) on

29
pourra négliger le terme de diffusion macroscopique de soluté par rapport aux effets
convectifs (Pe grand).

Figure 9: Schéma de l’écoulement dans la zone pâteuse environnant le canal ségrégé

Le bilan total de soluté dans le volume de contrôle considéré dans la Figure 9 peut
s’obtenir par une intégration de l’équation ci-dessus dans le volume de contrôle V (voir
Figure 9). Avant même de passer à l’intégration proprement dite, identifions d’abord les
flux de soluté entrant et sortant du volume de contrôle V :
ƒ Le flux liquide entrant dans le volume au niveau du macrofront (surface S1 = AB ) :

min = ∫ ε f v f dS (17)
S1

En tenant compte du fait qu’au niveau du macrofront la fraction liquide est proche
de l’unité on obtient:

min = ρ S1 v f 1 (18)

où S1 est la surface du macrofront entre deux canaux adjacents (AB) et v f 1 la

vitesse moyenne du flux de masse pénétrant dans la zone pâteuse. Il est important
de noter que ce flux de masse dépend principalement de la distribution de pression
au niveau du macrofront. En effet, en raison de sa très faible perméabilité au fluide
les effets inertiels dans la zone pâteuse sont négligeables par rapport au gradient
local de pression. Ainsi l’équation de conservation de QDM dans la zone pâteuse
peut être approchée par la loi de Darcy caractérisant les écoulements dans les
milieux poreux à faible perméabilité. Ainsi l’écoulement dans cette zone pâteuse
sera dicté principalement par le gradient de pression local et la perméabilité locale.
Cette distribution particulière de pression au niveau du macrofront créé par
l’écoulement ascendant associé à chaque canal (basses pressions au niveau du canal
et hautes pressions entres les deux canaux) fait qu’une partie du flux de masse
venant impacter le macrofront pénétrera également dans la zone pâteuse, c’est le
flux min .

30
 Il est aussi important de remarquer que la concentration caractérisant le flux entrant
min est très proche de celle du bain liquide C0 .
ƒ Le flux de liquide latéral sortant du volume V et débouchant dans le canal :

mout = ∫ ε f v f dS (19)
S2

où S 2 ( = AD ) est la surface latérale du canal. A la différence du flux entrant, le

liquide débouchant dans le canal est proche de la concentration d’équilibre Cl*
puisqu’il vient de sortir de la zone pâteuse.

Comme le volume V est lié au macrofront, pour faire l’intégration sur V il faut
aussi tenir compte de la vitesse d’avancement du macrofront Vmf . Dans le cas ou le bain
liquide n’est pas encore enrichi en soluté celle-ci est proche de la vitesse des isothermes
dans le lingot, c'est-à-dire de la vitesse de tirage Vt . En appliquant la règle d’intégration de
Leibnitz on obtient :

∂
∫
∂t V
CdV − C0Vt S1 + C H Vt S1 − ∫ ε f v f C0 ndS + ∫ ε f v f Cl* ndS = 0
S1 S2
(20)

min C0 *
mout C l

* 1
où C l =
mout ∫ε
S2
f v f Cl* ndS et la concentration moyenne du liquide débouchant dans le

1
S1 S∫1
canal et C H = C dS la concentration moyenne dans la zone pâteuse au niveau

y = − H . Le problème peut être sensiblement simplifié si l’on considère l’évolution du

processus de solidification comme étant quasi-stationnaire. En effet, à l’exception du début
de la solidification quand les panaches solutaux sont instationnaires (essentiellement dans
les premières minutes) l’évolution des deux canaux ségrégés latéraux et le champ de
vitesse associé sont proches d’une évolution quasi-stationnaire par rapport à l’évolution du
macrofront. L’hypothèse de quasi-stationnarité nous permet de faire les simplifications
suivantes :
ƒ la variation de soluté dans le volume V, le premier terme de l’équation (20), peut être
négligé ;
ƒ l’équation de conservation de masse pour le liquide écrite sous l’hypothèse de quasi-
stationnarité devient :

min = mout = m (21)

Dans ces conditions on obtient aisément que :

C H = C0 −
( *
m C l − C0 ) (22)
S1 Vt

31
C'est-à-dire que la concentration moyenne au niveau y = − H , C H , est toujours plus petite
que la concentration nominale de l’alliage C0 . Le canal agit donc comme une pompe de
soluté, en appauvrit toujours la zone pâteuse.

a) b)

c) d)
Figure 10: La fraction solide (a) et la concentration moyenne (b) ( C max = 16.3% ) obtenues
avec le modèle en [6] et la fraction solide (c) et la concentration moyenne (d)
( C max = 17% ) obtenues avec le modèle présent ; t = 1200 s

Il faut aussi remarquer que, à terme, le liquide dans le bain liquide s’enrichit en
soluté. Ce phénomène est évidemment un effet direct de l’apport net de liquide enrichi
dans le bain liquide dû aux panaches de liquide enrichi correspondants aux canaux
ségrégés. L’effet de l’enrichissement du bain liquide est bien prononcé dans ce cas
particulier (à t=20 min, Cbain ≈ 15% , Figure 8b) et il est dû aux petites dimensions du
lingot. Ceci implique aussi un important ralentissement de l’avancement du macrofront.
Aussi l’enrichissement du bain liquide détermine d’importantes ségrégations normales
dans la zone pâteuse (voir Figure 8). A ce stade il serait intéressant de comparer les
résultats présents avec les résultats obtenus par Ghislain Quillet [6] en utilisant le modèle

32
mathématique simplifié de Felliceli [8]. En Figure 10 illustre la comparaison entre la
distribution de la concentration moyenne et de la fraction liquide au temps t = 1200 s . On
peut observer un bon accord entre les deux modèles tant au niveau des macroségrégations
que de la structure de la zone pâteuse (les présences des deux canaux latéraux). On
s’attendait à ce que le nouveau modèle donne des résultats semblables à ceux en [6] car
d’un côté les paramètres numériques sont similaires et de l’autre côté la distance entre les
bras primaires choisie est petite ( λ1 = 300 µ m ). En effet, pour cette valeur de λ1 il
correspond une grande densité volumique des grains colonnaires ( nc0 ). Ceci fait que
l’enrichissement du liquide extra-dendritique caractérisant les zones derrière les pointes
colonnaires se fait très rapidement. Ainsi, la zone pâteuse sera, à l’exception d’une très
mince zone proche de Tliq , proche de la concentration de l’équilibre Cl* . Le chemin de
solidification d’un point dans le lingot sera par conséquent très proche de celui obtenu en
référence [6] où l’équilibre thermodynamique est considéré dans toute la zone pâteuse. Le
bon accord entre les deux codes dans la limite où λ1 petit ne peut donc que renforcer la
confiance dans les calculs présents. Toutefois l’utilisation du présent modèle comporte par
rapport au modèle décrit en [6] certains avantages :
ƒ une approche rigoureuse de l’équation du bilan de QDM (en [6] la l’équation de
conservation de QDM provient d’une approche simplifiée).
ƒ la prise en compte des situations de non-équilibre local ( Cl peut être en déséquilibre
avec l’interface) ;
ƒ la prise en compte de la cinétique de croissance des grains ;
ƒ la modélisation de la réaction eutectique.

4.5.4.2 L’influence de la convection forcée sur la solidification – cas 2D

Le cas analysé
On s’attachera par la suite à analyser l’influence de la convection forcée produite par le
brassage électromagnétique glissant sur la solidification. Les cas analysés par la suite sont
identiques au cas précédant c'est-à-dire au cas de la solidification dirigée d’un alliage Pb-
10wt% Sn dans un lingot bidimensionnel de dimensions 5 ×10 mm . Les conditions de
refroidissement restent elles aussi les mêmes : G = 10 K / cm , CR = 1 K / min . La
différence réside dans la prise en compte cette fois d’une force électromagnétique non-
nulle sera considérée.

En effet, l’application d’un champ électromagnétique dans le lingot génère une

force électromagnétique qui à son tour modifie significativement l’écoulement dans le bain
liquide et la distribution de soluté dans la zone pâteuse. Par la suite on appliquera un
champ magnétique B alternatif de module B = B0 , glissant dont le déplacement est
parallèle à l’axe vertical du lingot. Ce champ peut être engendré par un inducteur linéaire
dont l’axe est parallèle à l’axe vertical du lingot (Figure 11). Celui-ci donne naissance à
une force électromagnétique :

1
F= B × (∇ × B ) (23)
µ0

où µ0 représente la perméabilité magnétique du milieu.

33
Figure 11: La force électromagnétique produite par un inducteur linéaire

Si l’on admet l’existence d’un inducteur parfait et très long par rapport à la hauteur
du lingot on peut supposer que la force motrice n’a qu’une composante verticale. On peut
alors calculer la force électromagnétique moyenne agissant sur le métal liquide :

−
2x
σωλm 2
Fmz = F0 e δ
; F0 = B0 (24)
2π

où ω est la pulsation du champ magnétique, σ la conductivité électrique du milieu, λm le

pas polaire et δ est l’épaisseur de peau magnétique :

{
δ = Re ( iµ0σω + k 2 )
2
} (25)

Deux configurations de brassage électromagnétique seront étudiées par la suite :

ƒ brassage électromagnétique engendrant une convection forcée à double vortex,
ascendant sur les côtés. Cette configuration est susceptible d’être créée par deux
brasseurs linéaires latéraux produisant une force verticale résultante dirigée vers le
haut :

⎛ − 2 x − 2( l − x ) ⎞
Fmz = F0 ⎜ e δ + e δ ⎟ (26)
⎜ ⎟
⎝ ⎠

où l est la largeur du creuset.

ƒ brassage à double vortex, ascendant au milieu produit par deux brasseurs latéraux à
polarité inversée. La force moyenne résultante sera dirigée vers le bas et aura la forme
suivante :

⎛ − 2 x − 2( l − x ) ⎞
Fmz = − F0 ⎜ e δ + e δ ⎟ (27)
⎜ ⎟
⎝ ⎠

34
 Pour les deux configurations ci-dessus présentées on prendra F0 = 1000 N / m3 et
δ = 0.01m en analogie avec les cas traités en [6]. Cette force reste suffisamment faible
pour que l’écoulement soit laminaire mais suffisamment grande pour que l’influence du
brassage électromagnétique soit supérieure à celle induite par la convection naturelle dans
le bain liquide. Pour confirmer cela il suffit d’estimer la vitesse maximale dans le bain
liquide produite par la force électromagnétique En admettant que l’équilibre entre les
forces inertielles et les forces magnétiques s’établisse et en faisant une simple analyse
d’échelles de l’équation de mouvement dans le bain liquide on peut écrire :

∂V
ρV = Fmz
∂y
≈ (28)
V2
ρ ≈ Fmz
H

On peut facilement obtenir que:

Fmz H ⎛ −2 m ⎞
V≈ ⎜ = 2.7 ×10 ⎟ (29)
ρ ⎝ s⎠

Dans ces conditions la vitesse maximale dans le bain liquide est d’un ordre de
grandeur plus grand que la vitesse engendrée par la convection naturelle seule
( v 10−3 m / s , voir sous paragraphe précédent). L’effet prépondérant du brassage
électromagnétique permettra comme on va le voir le contrôle de l’écoulement dans le bain
liquide et par conséquent dans la zone pâteuse.

Résultats numériques
La force électromagnétique de l’équation (27), engendre un écoulement forcé à double
vortex ascendant au milieu (voir Figure 12c). On observe que l’amplitude de la vitesse se
situe autour du 1.4 cm/s, donc du même ordre de grandeur que la valeur estimée en (29).

Dans la Figure 12 les cartes de la fraction solide et de la concentration liquide sont

présentées à t = 200 s après le début de la solidification. On distingue déjà le début d’un
canal central ségrégé et le panache de liquide enrichi qui lui est associé. Ce canal est la
conséquence directe de l’écoulement dans le bain liquide. En effet, l’écoulement à double
vortex engendre une distribution particulière de pression au niveau du macrofront : des
hautes pressions là où le liquide impacte le macrofront (sur les parois latérales) et des
basses pressions au centre où le liquide remonte (Figure 13). Comme l’écoulement dans la
zone pâteuse est déterminé principalement par les gradients de pression locaux, la
distribution de pression au niveau du macrofront détermine un écoulement en zone pâteuse
qui se fera de l’extérieur vers l’axe du lingot (voir Figure 13). Celui-ci apportera du liquide
enrichi dans l’axe du lingot et y retardera la solidification. Ainsi une poche liquide
commence à se former qui à terme se transformera en un canal ségrége (voir Figure 12b).

35
 a) b)

c) d)
Figure 12: La a) concentration de soluté dans le liquide; b) fraction solide; c) vitesse; d)
température, à t = 200 s du début de la solidification.

Il est donc intéressant de remarquer que l’écoulement dans le bain liquide influence
l’écoulement dans la zone pâteuse par l’intermédiaire de la distribution de pression au
niveau du macrofront. Toutefois il faut remarquer que l’écoulement dans le bain liquide
détermine l’apparition d’un canal ségrégé seulement si l’écoulement pâteux généré reste
suffisamment fort par rapport aux conditions particulières de solidification. Une analyse
relativement simple de l’écoulement nous permet de quantifier la capacité de celui-ci de
créer un canal ségrégé ou non. Considérons par exemple une tranche élémentaire de zone
pâteuse ayant une fraction solide ε s et la température T . Supposons qu’un flux de liquide
le traverse (voir Figure 14), ce flux ayant une vitesse moyenne v f .

36
Figure 13: L’écoulement pâteux suit la distribution de pression au niveau du macrofront

Figure 14: Une tranche élémentaire de zone pâteuse et l’écoulement inter-dendritique

correspondant

Admettons aussi que le liquide est dans un état de mélange parfait ayant donc la
concentration à l’équilibre Cl* . Ceci est généralement valable pour la zone pâteuse à
l’exception d’une zone très mince située proche des pointes colonnaires. L’équation
caractérisant le transport de soluté dans cette tranche de zone pâteuse n’est plus que
l’équation moyenne de conservation de soluté dans le liquide :

∂ ( ε f ρ f Cl* )
+ div ( ε f ρ f Cl* v f ) = div ( ε f ρ f D f ∇Cl* ) − Cs*Γ s (30)
∂t

En négligeant la diffusion macroscopique de soluté (le premier terme de la partie

droite) et en tenant compte de la conservation de masse du liquide l’équation ci-dessus
devient :

37
 ∂Cl*
ε f ρf + ε f ρ f v f ∇Cl* = ( Cl* − Cs* ) Γ s (31)
∂t

Cette équation nous permet d’exprimer le transfert de masse Γ s et l’on observe aisément
que s’il n’y a pas de convection dans la zone pâteuse ( v f = 0 ) l’équation (31) ne
représente plus que la loi de Scheil. Par contre si l’écoulement pâteux est non nul celui-ci
va significativement influencer le processus de solidification par l’intermédiaire du terme
ε f ρ f v f ∇Cl* . Il est intéressant d’analyser le signe de ce terme. Pour ceci il est utile de
l’écrire sous la forme :

∇T ∇T
εf ρf vf nG = ε f ρ f v fG (32)
m m

où nG désigne un vecteur unitaire porté par le vecteur gradient thermique ∇T et v fG la

composante de la vitesse v f le long du vecteur gradient thermique. Ainsi on observe
facilement que si la vitesse v fG est positive le terme convectif ε f ρ f v f ∇Cl* devient négatif
et l’inverse. C'est-à-dire, si le flux de masse est localement orienté dans le même sens que
le gradient local de température la solidification diminue en intensité par rapport à un cas
purement diffusif et si le flux de masse est orienté dans le sens contraire avec le gradient
local de température la solidification est intensifiée. Pour le cas 2D analysé dans la Figure
12 il est maintenant facile de voir qu’aux endroits où le fluide remonte la zone pâteuse
(dans l’axe) la solidification diminue en intensité et aux endroits où le flux de masse
descend (sur les latérales) la solidification est accélérée. Il est intéressant maintenant
d’analyser les conditions critiques pour lesquelles la solidification est complètement
annulée. On obtient que la vitesse critique correspondante est telle que :

∂Cl* ∇T
εf ρf + ε f ρ f v fG =0 (33)
∂t m
où:
−1
⎛ ∂T ⎞
v crit
fG = − ∇T ⎜ ⎟
⎝ ∂t ⎠ (34)
= Visoth

c'est-à-dire la vitesse critique est égale à la vitesse locale des isothermes. Dans le cas où
dans la zone pâteuse la vitesse locale v f est dirigée dans le même sens que le gradient
local de température et que sa composante le long du même gradient est égale à la vitesse
locale des isothermes alors le taux local de solidification s’annule. Si

v fG > Visoth (35)

alors Γ s < 0 et des phénomènes locaux de refusion ont lieu.

38
 a) b)
Figure 15: La concentration moyenne (a) et la fraction solide (b) à t = 1200 s

En revenant sur les résultats numériques on observe que l’écoulement pâteux pour
le cas analysé est assez fort pour que la zone centrale reste non-solidifiée. Ceci indique que
l’écoulement dans le canal est assez fort pour que la solidification n’ait pas lieu dans le
canal. On remarque aussi le fort enrichissement du bain liquide dans le temps (voir Figure
15a, Cbain 15% ). Ceci est évidemment l’effet direct du liquide enrichi en soluté qui sort
du canal central et débouche dans le bain liquide. En regardant la carte de la concentration
moyenne à t = 1200 s (Figure 15a) on remarque clairement le fort enrichissement de la
zone centrale ainsi que la déplétion en soluté de la zone pâteuse entourant le canal,
conséquence de la présence du canal central (Figure 15b). On remarque aussi une
importante ségrégation normale, conséquence de l’enrichissement continu du bain liquide.

En analysant maintenant la deuxième configuration de la force électromagnétique,

équation (26), on observe que celle-ci produit dans le bain liquide un écoulement à double
vortex ascendant sur les côtés (Figure 16c). Si on regarde la carte de la fraction liquide et
de la concentration liquide à t = 1200 s (Figure 16a, b) on observe la présence des deux
canaux latéraux et des panaches de liquide enrichi correspondants. En effet, l’écoulement
dans le bain liquide influence de telle manière celui dans la zone pâteuse que cette fois
deux canaux latéraux apparaissent. En fait les zones où le liquide vient impacter le
macrofront (zone centrale) sont caractérisées par des pressions plus grandes que les
pressions caractérisant les zones où le liquide remonte (zone latérales). Ainsi un
mouvement dans la zone pâteuse apparaît se produisant de l’axe du lingot vers les zones
latérales. Celui-ci apporte du liquide enrichi en soluté vers les zones latérales ce qui y
retarde la solidification. Ainsi des poches de liquide se forment initialement proche des
parois latérales qui à terme évoluent vers des canaux ségrégés (Figure 16b). En regardant
la carte de concentration moyenne (Figure 16d) on observe le fort enrichissement des
canaux et la déplétion en soluté de la zone pâteuse centrale. On remarque aussi un fort
enrichissement du bain liquide ( Cbain 15% à t = 1200 s ) ceci étant la conséquence de la
petite dimension du domaine liquide analysé. En comparant la carte de la fraction solide à
t = 1200 s (Figure 16b) avec la carte correspondant au cas de la convection naturelle pure
(Figure 10c) on observe que la zone centrale du lingot pour le cas de la convection forcée
est caractérisée par une plus grande fraction solide par rapport au cas de la convection

39
naturelle. Ceci peut être expliqué par le fait que pour le cas de la convection forcée le
brassage de la zone pâteuse est plus intense que pour le cas de la convection naturelle. Ceci
est principalement du au fort brassage du bain liquide qui à son tour va influencer la
distribution de pression au niveau du macrofront de telle manière que le gradient de
pression au niveau du macrofront sera plus fort par rapport au cas de la convection
naturelle. Un brassage plus fort de la zone pâteuse signifie des vitesses plus grandes.
Comme nous l’avons déjà montré, un écoulement plus intense dirigé dans le sens contraire
au gradient local de température déterminera une solidification plus intense. Ainsi pour le
cas de la convection forcée la zone centrale subira une solidification plus intense par
rapport au cas de la convection naturelle.

a) b)

c) d)
Figure 16: La concentration dans le liquide (a), la fraction solide (b), la vitesse (c) et la
concentration moyenne (d) à t = 1200 s depuis le début de la solidification.

40
4.5.4.3 L’influence de la convection forcée sur la solidification – cas 3D

Le cas analysé
On a pu analyser et comparer avec les résultats de référence [6] plusieurs cas 2D de
solidification dirigée tout en mettant en évidence l’influence de la convection
naturelle/forcée sur le processus de solidification et les macroségrégations. Toutefois on
note que notamment dans le cas d’une solidification dirigée sous l’influence de la
convection naturelle, les effets tridimensionnels ne peuvent pas être négligés. En effet les
canaux ségrégés ont une structure essentiellement tridimensionnelle. De plus leur position
dans le lingot n’est pas toujours déterministe. Aussi on se propose par la suite d’analyser
une configuration de solidification tridimensionnelle : une solidification dirigée dans un
barreau cylindrique de 12 mm de diamètre et 30 mm de hauteur (voir Figure 17).

Figure 17: Géométrie du cas 3D et conditions limites correspondantes

En sachant que les conditions expérimentales des solidifications dirigées de type Bridgman
supposent des échantillons très allongés on a choisi un facteur de forme H / Φ 3 en
essayant de minimiser l’influence de la paroi supérieure imperméable et de maintenir en
même temps un temps de calcul acceptable. Les conditions limites sont semblables à celles
déjà utilisées pour les cas 2D, c'est-à-dire :
ƒ paroi inférieure imperméable ayant une température imposée et décroissante en temps,
équation (10) ;
ƒ paroi supérieure imperméable sur laquelle un flux de chaleur entrant est imposé,
équation (11) ;

Les conditions initiales sont elles aussi similaires avec les cas 2D :
ƒ un gradient vertical initial constant, G ;
ƒ un champ de concentration initial proche de C0 mais légèrement et aléatoirement
perturbé (équation (12)) . Comme nous l’avons expliqué antérieurement cette
perturbation initiale est introduite dans le système pour initier la déstabilisation de
l’écoulement proche du macrofront et donc pour initier la formation des canaux
ségrégés.

41
Tableau 4: Paramètres numériques pour le cas 3D
Alliage : Pb-10wt%Sn
Diamètre : 12 mm
Hauteur : 30 mm
No. de cellules : 20000
Schéma temporel : 1er ordre Euler
Schéma spatial : 2ème ordre UPWIND
Couplage pression-vitesse : SIMPLEC
Pas de temps : 0.02s
Temps de simulation : 1000s

Les paramètres numériques utilisés sont réunis dans le Tableau 4. Il faut noter que
les conditions de solidification (gradient thermique, vitesse de solidification) ont été
choisies similaires avec le cas 3D déjà analysé en [6]. Même si la géométrie est légèrement
différente (diamètre et hauteur du barreau plus grandes pour le cas présent) la densité du
maillage a été gardé constante. Ceci permettra une comparaison qualitative avec les
résultats de Ghislain Quillet [6].

Figure 18: La configuration de l’inducteur rotatif

En essayant de mettre en évidence l’effet de la convection naturelle et du brassage

électromagnétique à l’aide d’une seule expérimentation numérique on a choisi la
configuration temporelle suivante :
ƒ on commence l’expérience sans champ magnétique pour une période de 800 s ;
ƒ à partir de t = 800 s on applique un champ magnétique rotatif produisant une force
électromagnétique située dans le plan transversal au barreau cylindrique (xOy) suivant
la direction θ et ayant la variation radiale suivante :

r
Fθ = F0 (36)
R

42
 où R est le rayon du cylindre. En effet, en considérant l’inducteur autour du lingot
suffisamment long (voir Figure 18), on peut approcher la force motrice agissant sur le
liquide avec la distribution décrite dans l’équation ci-dessus. Comme intensité
maximale pour Fθ on a choisi F0 = 50 N / m3 , valeur similaire au cas analysé en [6].

Résultats numériques en convection naturelle

On observe très tôt après le début de la solidification l’apparition des instabilités de
l’écoulement au niveau du macrofront (voir Figure 19, 20). Celles-ci, par un mécanisme
similaire à celui déjà expliqué auparavant va engendrer l’apparition des poches liquides au
niveau du macrofront et à terme le développement des canaux ségrégés, les « freckles »
(Figure 21). On note que ces structures ont été fréquemment observées expérimentalement
[1] mais aussi numériquement [6, 8]. En effet on peut observer dans la Figure 2 la présence
de plusieurs canaux sur une métallographie obtenue après la solidification d’un alliage Pb -
Sn dans une géométrie cylindrique.

Figure 19: La concentration dans le liquide pour une coupe verticale à t = 120 s du début
de la solidification

Figure 20: Le module de la vitesse pour une coupe verticale à t = 120 s du début de la
solidification

43
Figure 21: La fraction solide dans une coupe transversale ( h = 1.5mm de la base du
creuset) à t = 200 s du début de la solidification

On remarque la présence des panaches de liquide enrichi en soluté qui sortent de

chaque canal et débouchent dans le bain liquide en l’enrichissant (Figure 19). Les vitesses
maximales au niveau de ces panaches sont de l’ordre de 1.5 mm/s (Figure 20). Dans la
Figure 21 on peut voir la présence de plusieurs canaux ségrégés dans la zone pâteuse. Il
faut remarquer la prédominance des canaux près de la paroi latérale ceci corroborant les
résultats obtenus en [6]. Toutefois, par opposition au cas 2D des canaux centraux
subsistent aussi. Ceci peut être expliqué par le fait que le domaine choisi pour ce cas 3D
( Φ = 12 mm ) est plus grand par rapport au cas 2D ( L = 5 mm ). En fait pour le cas 2D,
même si au début de la solidification des poches liquides se développent à l’intérieur du
lingot (Figure 6a), seuls les deux canaux latéraux subsistent. En effet il pourrait exister un
espacement minimal entre deux canaux latéraux. Ceci a déjà été observé numériquement
par Ghislain Quillet [6]. Il semble qu’en dessous d’un certain seuil (si les canaux sont trop
proches) les interactions solutales et hydrodynamiques trop fortes entre deux canaux
adjacents vont engendrer la disparition de l’un d’entre eux au profit de l’autre. Ainsi, la
survie de deux canaux latéraux pour le cas 2D peut s’expliquer par la dimension trop petite
du lingot par rapport à l’espacement minimal entre deux canaux adjacents. En choisissant
un domaine 3D plus grand des canaux centraux subsistent aussi à côté des canaux latéraux.

44
 a)

b) c)

d) e)
Figure 22: a) Fraction solide dans une section verticale ; b), c) Fraction solide et d), e)
concentration moyenne dans deux sections transversales (situées à h = 1.5mm
respectivement h = 7 mm de la base du creuset), à t = 800 s du début de la solidification.

Comme pour les observations de Quillet [6] on remarque nous aussi une évolution
de la configuration des canaux dans le temps. Ainsi au temps t=800 s on ne retrouve que

45
quatre canaux (Figure 22b, c) sur les 7 canaux ségrégés observés au début de la
solidification dans la zone pâteuse (Figure 21). Il est intéressant d’observer que les canaux
qui ont disparu ont laissé place à une configuration plus homogène dans le sens où les
quatre canaux finaux sont plus uniformément distribués dans l’espace. Ce phénomène peut
s’expliquer par le fait que la configuration initiale contenait des canaux qui étaient trop
proches l’un de l’autre par rapport à l’espacement minimal ci-dessus défini. Ceci a
engendré la disparition de certains d’entre eux, laissant place à une configuration
comprenant des canaux plus distants l’un de l’autre et aussi répartis dans l’espace de
manière plus homogène. Il serait intéressant dans la suite de cette thèse de se pencher
davantage sur cet aspect d’espace minimal en essayant de le quantifier. Une approche
intéressante serait celle de la théorie constructale [16] que nous avons déjà appliquée pour
calculer l’espacement primaire (référence I, Annexe 7.2). Toutefois on s’attend ici à une
analyse plus complexe car il semble que l’espace optimal entre deux canaux ségrégés soit
dépendant des interactions à la fois hydrodynamiques et solutales entre les canaux
adjacents.
On remarque aussi la forte ségrégation radiale produite par la présence des canaux
(Figure 22d, e) : le fort enrichissement des canaux et l’appauvrissement en soluté de la
zone pâteuse environnant les freckles.

Résultats numériques en brassage rotatif

On a vu qu’à partir de t = 800 s un brassage rotatif est pris en compte. Ce brassage
engendre évidemment un mouvement principal de rotation du liquide autour de l’axe du
cylindre (Figure 23). Toutefois à cause de la présence des parois supérieures et du
macrofront, l’écoulement ne peut pas rester uniquement dans le plan xOy car la
conservation de quantité de mouvement ne serait pas satisfaite. En effet un mouvement
secondaire prendra place dans le lingot cylindrique. Ceci correspond à un mouvement
toroïdal qui va brasser le liquide axialement. La cause de cet écoulement secondaire n’est
que la conséquence de la distribution de pression créée par le mouvement rotatif principal.
En effet la force centrifuge créée par le mouvement rotatif principal engendre une pression
plus élevée proche des parois latérales et plus basse dans l’axe du cylindre (voir Figure
24a). Ainsi un écoulement secondaire se forme en déplaçant du fluide des zones de haute
pression (latérales du cylindre) vers les zones de basse pression (l’axe du cylindre). Si
l’influence des gradients de soluté dans le bain liquide n’est pas prise en compte, deux
écoulements toroïdaux de sens contraire : un supérieur et un inférieur prendront naissance
dans le bain liquide (voir Figure 24a). En regardant maintenant la distribution réelle de
vitesse (voir la distribution de la composante verticale de vitesse, Figure 25) on observe
que les écoulements secondaires ont une configuration un peu plus compliquée dans le
sens où deux écoulements toroïdaux et non pas un existent dans la partie supérieure du
lingot. Ceci peut toutefois s’expliquer lorsque l’on tient compte de la distribution de soluté
dans le bain liquide. En effet, le mouvement toroïdal inférieur (Figure 24a) va apporter du
liquide pauvre en soluté près des parois latérales, ce qui va finalement accélérer la
solidification dans ces zones. Ainsi les canaux ségrégés latéraux vont être finalement
bouchés. En même temps la distribution particulière de pression au niveau du macrofront,
hautes pressions sur les latérales et basses pressions dans l’axe du lingot, va générer un
mouvement correspondant dans la zone pâteuse. En effet, comme l’écoulement dans la
zone pâteuse est principalement créé par la distribution locale de pression, un écoulement
prendra naissance dans la zone pâteuse de la paroi latérale vers le centre. Ainsi du liquide
enrichi est apporté dans l’axe du lingot et par un mécanisme identique à celui déjà
rencontré au cas 2D cet écoulement va engendrer la formation d’un canal ségrégé central.
On peut remarquer en effet dans la Figure 26 la configuration de la zone pâteuse engendrée

46
par le brassage électromagnétique glissant. On observe très bien la disparition des canaux
latéraux et l’apparition d’un gros canal central. Maintenant le canal central va déterminer à
son tour l’apparition d’un panache central de liquide enrichi. L’étain étant plus léger que le
plomb, le panache aura la tendance naturelle à monter. L’effet de poussée produit par ce
panache de liquide enrichi aura comme conséquence la modification sensible des
écoulements secondaires dans la partie supérieure du bain liquide de telle manière que le
liquide monte aussi dans la partie centrale du creuset (voir Figure 25).
On note aussi que la présence du canal central engendre de fortes ségrégations
radiales (Figure 27), phénomène auquel nous nous attendions puisque le canal central se
comporte comme une pompe de soluté drainant la zone pâteuse environnante.

Figure 23: Vecteurs vitesses dans la section transversale situé à h = 15 mm de la base du

creuset, à t = 1000 s du début de la solidification

a) b)
Figure 24: a) La configuration des écoulements secondaires si le mouvement d’origine
solutale n’est pas prise en compte ; b) La configuration réelle des écoulements secondaires
quand l’influence du panache de liquide enrichi est prise en compte.

47
 Canal
central

Figure 25: La magnitude de la composante verticale de la vitesse vz dans une section

verticale à t = 1000 s du début de la solidification

a) b)
Figure 26: a) La fraction solide à t = 840 s (40s après le début du brassage
électromagnétique) ; b) La fraction solide dans une section verticale à t = 1000 s .

48
Figure 27: La concentration moyenne dans une section verticale à t = 1000 s .

4.5.4.4 Conclusions partielles

En analysant plusieurs cas de solidification dirigée sous l’influence de la convection
naturelle et/ou forcée on a observé que la structure de l’écoulement à proximité du
macrofront a une très grande influence sur le développement des éventuels canaux
ségrégés. Le passage du cas sans brassage électromagnétique au cas où l’écoulement dans
le bain liquide est contrôlé est très bien reflété par la structure de la zone pâteuse. Dans le
cas de la convection naturelle plusieurs canaux ségrégés isolés sont présents dans la zone
pâteuse. Par contre l’écoulement forcé dans le bain liquide imposé par le brassage
électromagnétique modifie complètement la configuration des canaux. En effet, en
contrôlant l’écoulement dans le bain liquide, on peut aussi contrôler la distribution de
pression au niveau du macrofront. Principalement contrôlé par la distribution de pression
locale, l’écoulement dans la zone pâteuse va à son tour contrôler la distribution des canaux
ségrégés dans la zone pâteuse. Effectivement, un brassage important prend naissance dans
la zone pâteuse celui-ci étant dirigé des zones de basse pression vers les zones de haute
pression. Ce brassage amène du liquide enrichi en soluté vers les zones de basse pression
en y empêchant la solidification. Ainsi des canaux ségrégés sont donc susceptibles de se
développer aux endroits où l’écoulement pâteux débouche dans le bain liquide. Toutefois
les canaux ne peuvent exister seulement si l’écoulement dans la zone pâteuse est
suffisamment fort par rapport aux conditions locales de solidification. En effet on a montré
que les zones pâteuses où l’écoulement est dirigé dans le même sens que la vitesse locale
des isothermes Vis subiront un processus de solidification intensifié. Ces zones
correspondent aux zones de haute pression où l’écoulement pénètre dans la zone pâteuse.
Au contraire, aux endroits où l’écoulement pâteux se fait dans le sens contraire de la
vitesse locale des isothermes, le processus de solidification est diminué en intensité par
rapport à un cas purement diffusif. Si l’écoulement dépasse en intensité la vitesse des
isothermes, la solidification est annulée et un processus de refusion prend naissance
localement. Ainsi, un canal ségrégé peut naître et subsister dans le temps si seulement
l’écoulement pâteux local est suffisamment fort pour que la condition (35) soit respectée.
La présence d’un canal détermine aussi de fortes ségrégations dans la zone
pâteuse : le canal agit comme une pompe de soluté drainant la zone pâteuse. On a montré
aussi que le fort appauvrissement de la zone pâteuse autour d’un canal est du au fait que la

49
concentration moyenne caractérisant le flux de masse entrant dans la zone pâteuse est plus
petit par rapport à la concentration moyenne caractérisant le flux de liquide sortant de la
zone pâteuse et débouchant dans le canal.
On conclut ainsi que l’utilisation d’un brassage électromagnétique glissant ou
rotatif ne permet pas d’annuler les macroségrégations dues aux canaux ségrégés. Par contre
l’avantage de l’utilisation du brassage électromagnétique est le contrôle de la position des
canaux et donc la localisation précise des macroségrégations positives correspondantes en
opposition avec le cas de convection naturelle où la distribution des canaux était aléatoire.

4.5.5 Comparaison avec l’expérience ACCESS

Nous allons à présent exposer un essai de validation du modèle colonnaire à l’aide de
l’expérience de solidification sous l’influence d’un champ magnétique rotatif effectué au
laboratoire ACCESS, Aachen, Allemagne. Cette expérience a porté sur la solidification
dirigée d’un alliage binaire Al-7wt% Si.

4.5.5.1 Dispositif expérimental

L’expérience ACCESS est une expérience de solidification dirigée de type Bridgman où un
gradient thermique vertical G = 10 K / mm est imposé dans le four. L’échantillon est tiré
dans ce gradient thermique avec une vitesse de tirage constante Vt = 4.2 µ m / s . Le schéma
de l’expérience est présenté dans la Figure 28.

Figure 28: Le schéma de l’expérience de solidification dirigée ACCESS

Pour analyser l’influence du brassage électromagnétique sur la solidification, un aimant

permanent (voir Figure 28) est utilisé pour générer un champ magnétique rotatif dans la
section transversale du lingot. Les caractéristiques principales de cette expérience sont
résumées dans le Tableau 5. Une particularité intéressante de cette expérience est la
possibilité de décantation soudaine du liquide. Ce processus est réalisé par un basculement
rapide de l’échantillon qui en vertu de la force centrifuge importante expulse le liquide de
l’échantillon en exposant l’interface solide-liquide à un certain moment. Ceci permet de
quantifier l’interface solide-liquide au niveau du macrofront : rayon des pointes
colonnaires, espacement primaire, présence des éventuels canaux ségrégés dans la zone
pâteuse.

50
Tableau 5: Les caractéristiques principales de l’expérience ACCESS
Alliage utilisé : Al-7wt%Si-0.6wt%Mg
Gradient thermique vertical : 10000 K/m
Vitesse de tirage : 4.2 µ m / s
Le type d’aimant : permanent
Intensité maximale du champ magnétique : B0 = 2.5 × 10−2 T
Fréquence de rotation de l’aimant : 50 Hz
Diamètre de l’échantillon : 8 mm

4.5.5.2 Paramètres numériques

Une première approximation est faite au niveau de l’alliage. Nous serons en effet
contraints par le code est développé uniquement pour des alliages binaires de considérer
dans la suite la solidification d’un alliage binaire Al-7wt%Si même si l’alliage réel est Al-
7wt%Si-0.6wt%Mg. Toutefois en raison du faible pourcentage de Mg (0.6%) on s’attend à
ce que cette approximation n’introduise pas d’erreurs significatives.
Une deuxième approximation porte sur la distribution radiale de la force
électromagnétique. On considère que cette force agit uniquement dans le plan transversal à
l’échantillon, même si la force magnétique réelle n’est pas uniquement dans un plan
horizontal. Un champ uniforme d’intensité B0 = 25 mT a été mesuré dans l’axe de
l’aimant. La force électromagnétique qui agit dans un plan horizontal et suivant la direction
θ a l’expression générale suivante:

r
Fθ ( r , z ) = F0 ( z ) (37)
R

Toutefois il faut remarquer qu’en raison de la hauteur finie de l’aimant (h = 8mm) le

champ magnétique B0 subira une variation décroissante en fonction de la distance par
rapport à l’axe de l’aimant. En utilisant les mesures expérimentales mises à notre
disposition par ACCESS on a calculé la force électromagnétique résultante qui par la suite
a été approchée par la loi empirique suivante :

⎡ z − za
2 ⎤
Fθ ( r , z ) = F0 exp −
⎢ ⎥r (38)
⎢ (1.18 × 10−2 ) ⎥ R
2

⎣ ⎦

où F0 = 500 N / m3 et za est la coordonnée verticale de l’axe de l’aimant. On note aussi

que la position de l’aimant est légèrement décalée par rapport à l’isotherme Tliq
correspondante à la concentration C0 . En effet, le plan axial de l’aimant est positionné en
arrière de 16 mm par rapport à la position de Tliq . Ceci implique que l’intensité du champ
magnétique au niveau du macrofront n’est pas de 25 mT mais plus faible, d’environ 10 mT.
En tenant compte des conditions initiales choisies on considère qu’à t = 0 la température
sur la paroi inférieure est Tliq . Ainsi, la variation de la position de l’aimant peut être
aisément calculée comme étant :

51
 za (t ) = Vt t − ∆z0 (39)

où ∆z0 est le décalage de l’aimant par rapport à Tliq , ∆z0 = 16mm .

La hauteur de l’échantillon considérée dans les calculs est de 7cm plus petite que
celle réelle ( ≈ 12cm ). Toutefois comme on va pouvoir le constater, en raison du domaine
limité d’action de la force magnétique, le choix d’une hauteur de l’échantillon plus grande
que 7 cm ne change en rien les résultats finaux.
Comme la force électromagnétique est axisymétrique le problème réel 3D peut être
analysé avec succès avec un modèle axisymétrique de type swirl. En effet, même si il y a
un mouvement dans la direction azimutale θ , le problème peut être analysé dans un
domaine 2D car les dérivées ∂ / ∂θ de tous les champs physiques sont nulles. Toutefois
pour quantifier le transport convectif le long de la direction θ une équation de
conservation de QDM suivant θ doit être considérée. Au niveau de l’implémentation du
modèle en Fluent rien ne change par rapport à ce qui a été déjà dit auparavant (paragraphe
4.5.2) parce que le passage d’un modèle 3D à un modèle 2D axisymétrique se fait
automatiquement en Fluent. Il faut juste adapter le maillage au problème axisymétrique.
Les paramètres numériques utilisés dans la résolution numérique du cas de solidification
sont réunis dans le Tableau 6.

Tableau 6: Paramètres numériques du cas ACCESS

Alliage : Pb-10wt%Sn
Diamètre : 8 mm
Hauteur : 70 mm
Maillage : 20 X 350
Schéma temporel : 1er ordre Euler
Schéma spatial : 2ème ordre UPWIND
Couplage pression-vitesse : SIMPLEC
Pas de temps : 0.02s
Temps de simulation : 8000s

4.5.5.3 Résultats numériques et comparaison

A cause du brassage électromagnétique rotatif un écoulement de rotation s’établit
principalement dans le plan xOy (voir Figure 29a). Les vitesses maximales sont de l’ordre
de 1.1cm / s . Il faut aussi remarquer que l’écoulement engendré par le brassage
électromagnétique reste non-négligeable sur une hauteur finie au-dessus du macrofront.
Ceci est logique car l’aimant d’épaisseur fini crée un champ magnétique qui diminue en
intensité dès que l’on s’éloigne de l’axe de l’aimant, équation (38). Il faut remarquer
toutefois que l’écoulement ne reste pas uniquement dans le plan xOy. En effet un
mouvement secondaire toroïdal naît au dessus du macrofront comme le montre la Figure
29b. Celui-ci est du principalement à la distribution particulière de pression créée par
l’écoulement rotatif principal. Comme nous l’avons déjà expliqué auparavant, les forces
centrifuges qui agissent sur le liquide vont engendrer des pressions plus grandes proches de
la paroi latérale et plus basses dans l’axe du lingot (voir Figure 24a). Cependant, à cause de
la couche limite hydrodynamique au niveau du macrofront le maximum de pression se
trouvera non pas au niveau du macrofront mais à une certaine distance au-dessus de celui-
ci (voir Figure 30). En même temps l’écoulement diminue en intensité en s’éloignant du

52
macrofront. Alors les forces centrifuges et les pressions induites par celles-ci vont
diminuer aussi. Ainsi, l’existence de pressions plus basses caractérisera la zone éloignée du
macrofront (voir Figure 30). Cette distribution particulière de pression va engendrer deux
mouvements secondaires toroïdaux qui vont se diriger des zones de hautes pressions vers
les zones de basses pressions (voir Figure 29b et Figure 30).

a) b)
Figure 29: a) La magnitude de la vitesse et b) les lignes de courant dans le creuset à
t = 400 s du début de la solidification

Figure 30: Mécanisme d’apparition des écoulements secondaires

53
 De plus la distribution de pression au niveau du macrofront, les hautes pressions
proche de la paroi latérale et les basses pressions dans l’axe du barreau générent un
brassage de la zone pâteuse (Figure 30) qui apporte du liquide riche en soluté dans l’axe du
lingot. Ceci ajouté au mouvement toroïdal de même sens dans le bain liquide juste au-
dessus du macrofront donne naissance à un gros canal ségrégé central (Figure 31a)
similaire à celui déjà obtenu dans le cas 3D précédemment analysé. L’écoulement dans la
zone pâteuse aura comme effet l’enrichissement du canal et l’appauvrissement de la zone
pâteuse environnante par le mécanisme de pompage de soluté déjà identifié au paragraphe
4.5.4.1 (Figure 31b). Il est important de noter qu’expérimentalement un gros canal central
similaire à celui obtenu numériquement a été observé. On peut observer dans la Figure 32
l’interface solide-liquide au niveau du macrofront découverte par la technique de
décantation du liquide. On peut donc facilement observer un gros canal central. Sa
dimension est tout à fait comparable à celle du canal obtenu dans la simulation numérique.
Il est aussi important de noter qu’expérimentalement il a été observé que la zone centrale
du lingot était remplie par des grains équiaxes en opposition avec la zone extérieure du
lingot qui restait colonnaire. On observe très bien dans la Figure 33 la structure des grains
équiaxes qui caractérise la zone centrale du lingot. La corrélation entre le canal central et la
zone équiaxe centrale met en évidence le fait que des germes naissent dans le liquide
enrichi du canal probablement par nucléation homogène et/ou fragmentation des grains
colonnaires. En effet des phénomènes de refusion ont lieu dans le canal dus à l’écoulement
du liquide dans le sens contraire à la vitesse locale des isothermes. Ceci peut amplifier les
phénomènes de fragmentation des colonnes. Les fragments ainsi obtenus vont ensuite se
sédimenter au fond du canal donnant naissance à cette structure centrale équiaxe.

Tliq

a) b)
Figure 31: a) La fraction solide et b) la concentration moyenne dans le creuset à t = 2000 s
du début de la solidification

Les cartes de la microstructure suivant une section transversale et longitudinale

sont présentées dans la Figure 34. On observe une fois de plus la structure colonnaire
extérieure et une structure centrale équiaxe enrichie en Si. En effet on remarque l’existence
d’une fraction eutectique (zone en gris foncé) beaucoup plus importante dans la zone
centrale par rapport aux zones périphériques. Ceci confirme l’effet prononcé

54
d’enrichissement du canal central et l’appauvrissement de la zone colonnaire environnant
le canal. Un autre phénomène intéressant est l’augmentation du diamètre de la zone
centrale le long du lingot. Pour le moment on ne peut pas avancer une explication
raisonnable de ce phénomène, mais comme on va le découvrir dans la suite, la
modélisation numérique semble fournir une réponse pertinente à cette question.

Figure 32: Vue macroscopique des deux interfaces solide-liquide obtenues par la technique
de décantation du liquide. On remarque la présence d’un canal central (Weiss et al.)

a) b)
Figure 33: La structure des grains dans une section transversale (a) et une section
longitudinale (b). On observe la présence des grains équiaxes dans la partie centrale du
lingot (Weiss et al.).

En revenant maintenant aux résultats numériques on observe des phénomènes très

semblables. Premièrement on note le fort enrichissement du canal ségrégé central (Figure
31b, Figure 35b,d). Celui-ci reste liquide jusqu’à pratiquement la température eutectique au
fond du canal. En dessus de cette température la réaction eutectique se déclenche. Ainsi au
fur et à mesure que le front de solidification avance, une phase eutectique importante
occupera la partie centrale en concordance avec les résultats expérimentaux.
Deuxièmement à cause de l’existence du canal central le bain liquide s’enrichit en soluté

55
au fur et à mesure que le front de solidification avance. Ceci est évidemment la
conséquence du flux de liquide enrichi sortant du canal et débouchant dans le bain liquide.

a)

b)
Figure 34: La microstructure dans une section transversale (a) et une section
longitudinale (b). Les deux sections polies montrent la microstructure dendritique avec
l’enrichissement de la phase eutectique dans la partie centrale. Expériences réalisées par
ACCESS (Weiss et al.).

Toutefois il est intéressant d’observer qu’à cause de la limitation spatiale de l’écoulement

forcé dans le bain liquide, l’enrichissement du bain liquide n’est que partiel. En effet, seule
la zone où l’effet de mélange produit par la convection forcée est non-négligeable
s’enrichit en soluté (voir Figure 31b, Figure 35b,d). Ce phénomène a une conséquence très
importante sur l’évolution de la solidification. En effet, en raison de la présence de la zone
de mélange limité, le liquide au-dessus du macrofront s’enrichit rapidement par rapport à
un cas de brassage glissant où l’effet de mélange serait étendu à la totalité du bain liquide.
Ceci a plusieurs conséquences importantes :
ƒ d’une part cet enrichissement est tellement rapide que le macrofront se voit vite ralenti
par rapport à sa vitesse initiale de 4.2 µ m / s . Ceci entraîne le rapprochement de l’axe
de l’aimant par rapport au macrofront. On peut aisément remarquer ce phénomène dans
la Figure 31a et la Figure 35a,c où la position de l’isotherme Tliq est montrée. On
observe que le macrofront est de plus en plus en arrière par rapport à l’isotherme Tliq ,
ce qui signifie que l’axe de l’aimant s’approche de plus en plus du macrofront.
ƒ D’autre part à cause du rapprochement de l’aimant du macrofront, le liquide au dessus
du macrofront est plus fortement brassé. On peut remarquer qu’à t=8000 s la vitesse est
augmentée jusqu’à 2.8 cm/s (Figure 36c) par rapport à une vitesse proche de 1 cm/s à
t=400 s (Figure 29b). Ceci a une influence importante sur la distribution de pressions
proche du macrofront. En effet, la configuration de la couche limite hydrodynamique
au-dessus du macrofront dépend fortement du nombre de Reynolds local. La théorie de
Prandtl prédit une modification de l’épaisseur de la couche limite en fonction du Re
local comme suit : δ / R ∼ 1/ Re . Ainsi pour un brassage plus fort ( Re ) l’épaisseur
de la couche limite hydrodynamique diminue. Ceci déplacera le maximum de pression

56
 plus proche du macrofront et aura comme conséquence une modification sensible de la
structure de l’écoulement secondaire au dessus du macrofront. On observe clairement
cette modification dans la Figure 29b et Figure 36b, d. En regardant l’évolution de la
convection au dessus du macrofront on observe qu’avec le temps les deux boucles de
convection au dessus du macrofront deviennent de plus en plus aplaties et inclinées
vers les parois latérales. Ceci aura une influence significative sur la configuration du
canal central. En effet on observe (Figure 35b, d) une augmentation de la largeur du
canal dans le temps due à la modification de l’écoulement au-dessus du macrofront. Il
est important de noter que ce phénomène correspond qualitativement aux résultats
expérimentaux. Ainsi les résultats numériques semblent éclairer ce point qui était
jusqu’à maintenant non résolu.
ƒ L’enrichissement rapide du bain liquide a une conséquence directe sur la hauteur de la
zone pâteuse. En effet cette dernière diminue dans le temps en raison de la diminution
de l’intervalle de solidification ∆T0 = Tliq (Cbain ) − Teut . En effet de t=0s à t=4000s la
concentration au-dessus du macrofront augmente de 7% à 10% (voir Figure 35b).
L’intervalle de solidification diminue par conséquent et, pour un gradient thermique
donné, l’épaisseur de la zone pâteuse diminue aussi. Ceci est facilement observable en
comparant la Figure 31a avec les Figure 35a, c.
ƒ Finalement l’augmentation de la concentration du liquide au dessus du bain liquide
déterminera l’apparition, en dehors des ségrégations radiales créées par le canal central,
d’une macroségrégation longitudinale (normale) importante (voir Figure 35b, d).

Tableau 7: Résultats expérimentaux sur les macro ségrégations

Distance à partir du Diamètre de la Concentration Concentration
bas de l’échantillon région centrale moyenne du Si dans moyenne du Si dans
la région centrale la région périphérique
9 mm 2,0 mm 10.8 wt% 3.8 wt%
17 mm 3,2 mm 10.2 wt% 4.1 wt%
27 mm 3,5 mm 11,0 wt% 5.9 wt%

A ce stade il est intéressant de faire une comparaison directe entre les résultats numériques
et expérimentaux obtenus au niveau des macroségrégations. Les auteurs de l’expérience
nous fournissent la variation radiale de la concentration moyenne pour trois sections
transversales différentes. Ces résultats sont résumés dans le Tableau 7. Il faut noter que ces
profils sont obtenus après une procédure de prise de moyenne sur la partie centrale et
périphérique du lingot et ne représentent pas la variation radiale exacte de la concentration
moyenne locale. En Figure 37 les résultats numériques et expérimentaux sont superposés.
On observe que le modèle reproduit bien le niveau de la concentration dans la partie
centrale qui est proche de la concentration eutectique. On observe un bon accord avec
l’expérience en ce qui concerne le niveau de concentration dans les zones périphériques.
On remarque notamment la ségrégation normale positive importante : l’augmentation de la
concentration périphérique avec la hauteur des sections dans le lingot. Malgré l’accord
général satisfaisant avec l’expérience on remarque un désaccord quantitatif au niveau de la
concentration dans la zone centrale et ceci pour les trois sections analysées. Une
explication pour ce désaccord pourrait provenir du fait que la solidification équiaxe n’est
pas prise en compte par le modèle actuel. En effet, on a observé expérimentalement que la
zone centrale était toujours occupée par des grains équiaxes qui très probablement se
sédimentent au fond du canal central. La présence du solide dans le canal ainsi que leur
effet de sédimentation aura comme effet une diminution de la concentration moyenne du

57
canal par rapport à un cas où le canal resterait liquide jusqu’à un régime très proche de
l’eutectique. En fait c’est précisément ce qu’on obtient : la concentration de la zone
centrale est surestimée par rapport aux résultats expérimentaux.

axe

axe
Tliq

a) b)

axe

Tliq

axe

c) d)
Figure 35: Résultats du modèle. La fraction solide (a) et la concentration moyenne (b) à
t = 4000 s et la fraction solide (c) et la concentration moyenne (d) à t = 7400 s .

58
 axe

axe

a) b)

axe

axe

c) d)
Figure 36: La magnitude de la vitesse (a) et les lignes de courant (b) à t = 4000 s et la
vitesse (c) et les lignes de courant (d) à t = 8000 s .

59
Figure 37: Comparaison entre les résultats du modèle et les résultats expérimentaux de la
variation radiale de la concentration moyenne pour trois sections transversales différentes.
Expériences réalisées par ACCESS.

4.5.5.4 Conclusions partielles

On a essayé de modéliser l’expérience de solidification dirigée sous l’influence du
brassage électromagnétique rotatif, expérience réalisée au laboratoire ACCESS. L’accord
des résultats numériques avec ceux expérimentaux est aussi bon sous l’aspect qualitatif que
quantitatif. Plus précisément :
ƒ Le modèle reproduit le gros canal central produit par la convection forcée dans le bain
liquide. Les dimensions du canal central sont très proches des dimensions obtenues
expérimentalement.
ƒ Les ségrégations radiales et longitudinales sont en bon accord avec les résultats
expérimentaux à l’exception de la zone centrale qui est légèrement surestimée. La
cause de ce désaccord semble être la non-prise en compte de la zone équiaxe.
ƒ L’élargissement du canal dans le temps est aussi qualitativement reproduit par le
modèle. La cause de cet élargissement réside dans la modification des recirculations
secondaires au niveau du macrofront. Ces dernières sont à leur tour dues au
ralentissement du macrofront et au rapprochement du plan axial de l’aimant du
macrofront.
ƒ A cause de la présence du canal central, le bain liquide au dessus du macrofront
s’enrichit en soluté. A cause du brassage limité en extension dans le bain liquide, cet
enrichissement se fait très rapidement. En même temps l’enrichissement du liquide au
dessus du macrofront a comme conséquence directe la diminution sensible de
l’épaisseur de la zone pâteuse.

60
5 Conclusions finales
Dans une première partie de la thèse (référence I, Annexe 7.2) on se concentre sur
l’application de la théorie constructale dans la prédiction de la structure dendritique solide.
Premièrement on analyse le critère de stabilité marginale vu dans la perspective de la
théorie constructale. En ayant comme principe de base la loi constructale on conclu que la
forme de la dendrite s’adapte de telle manière que la vitesse de croissance de la pointe
dendritique soit maximale. Dans ce contexte on a montré que, parmi le domaine des rayons
possibles de la pointe dendritique prédit par l’analyse de stabilité, la pointe dendritique va
choisir le plus petit rayon ça veut dire un rayon égal avec la plus petite longueur d’onde
des perturbations donnant naissance à des instabilités. On identifie aussi l’existence d’une
compétition entre la croissance contrôlée par la diffusion moléculaire et la croissance
dendritique. En analysant cette compétition on a été capable de déterminer le temps
critique et le rayon critique pour lesquels la déstabilisation de l’interface solide-liquide se
déclanche. Dans un deuxième temps on développe un modèle pour l’espacement
dendritique secondaire. En analysant la croissance d’une seule pointe dendritique on
conclu que la croissance latérale de la pointe dendritique est proche d’une croissance
contrôlée par la diffusion moléculaire et par conséquent est caractérisé par une vitesse de
croissance qui diminue dans le temps. Dans ce contexte la théorie constructale prédit une
optimisation de la croissance latérale de la dendrite. En faite on identifie encore une fois
une compétition entre la croissance latérale contrôlée par la diffusion moléculaire et la
croissance dendritique. En analysant cette compétition on a été capable de caractérisé le
mécanisme de la croissance latérale des branches dendritiques secondaires et de calculer
finalement l’espacement dendritique secondaire. Le résultat est en bon accord avec le
modèle de J.S. Langer and H. Muler-Krumbhaar, Acta Metallurgica 26, 1681 (1978) mais
aussi avec des différents résultats expérimentaux. Finalement, l’espacement dendritique
primaire est analysé dans la perspective de la loi constructale. Quand on l’applique à un
front colonnaire la loi constructale prédit que le seul moyen dont le pointes dendritique
dispose pour l’optimisation du processus de solidification est la minimisation de l’espace
entre deux pointes adjacentes, c'est-à-dire la minimisation du λ1 . Le mécanisme
responsable pour la diminution du λ1 a été identifié dans le mécanisme de division
dendritique. Toutefois, à cause des interactions solutales entre les pointes adjacentes il a
été trouvé qu’il existe un espacement critique λcr en dessous duquel la division des pointes
colonnaire ne peut plus se faire. Dans ce sens la loi constructale prédit que l’espacement
primaire caractérisant l’état quasi-stationnaire du front colonnaire sera proche de
l’espacement stable minimum λcr . En quantifiant la division dendritique et les interactions
solutales entre les pointes colonnaire on a finalement été capable d’obtenir un modèle pour
l’espacement dendritique primaire. En effet, on obtient que λ1 est de même ordre de
grandeur avec λcr et plus précisément que λ1 ≈ 1.5λcr . Le résultat est par la suite validé
avec plusieurs donnés expérimentales et on obtient que le nouvel modèle est en très bon
accord avec les expériences.
Dans une deuxième partie de la thèse (référence II, Annexe 7.2) on a développé un
nouveau modèle multiphasique eulérien valide pour la solidification colonnaire et équiaxe
et pour un cas purement diffusif (sans convection). Les équations moyennes ont été
obtenues à l’aide d’un moyennage statistique de phase et la formulation finale du modèle
est valable aussi bien pour une structure équiaxe que colonnaire. Premièrement le cas
d’une structure colonnaire pure ou équiaxe pure est analysé. A l’aide de la densité de

61
probabilité des particules les plus proches, f
(1)
( t , x; z ) , on a pu calculer rigoureusement
1

les termes et les tenseurs inconnus existant dans les équations de conservation moyennes
non-fermées. En effet, en utilisant « l’approximation de la cellule » [5] pour cette fonction
de densité de probabilité, les champs moyens peuvent être facilement calculés à l’aide de
simples intégrations volumiques dans une cellule sphérique caractéristique de rayon R1 . Le
modèle final valide soit pour une structure colonnaire soit pour une structure équiaxe est
similaire à celui obtenu par Wang et Beckermann [10]. On note que l’équivalence entre les
deux modèles est une conséquence directe de l’utilisation de l’hypothèse de l’équilibre
thermique local. En effet, si l’hypothèse d’homogénéité thermique locale est considérée
alors la moyenne statistique d’ensemble reproduit l’approche de la prise de moyenne
volumique. Toutefois le moyennage statistique est fondamentalement différent du
moyennage volumique. En effet, si l’hypothèse d’homogénéité thermique locale n’est plus
valide, l’utilisation de la prise de moyenne statistique donne naissance à des expressions de
fermetures différentes de celles obtenues par l’application du moyennage volumique. En
plus, les résultats obtenus avec le moyennage volumique ne sont pas physiquement
corrects. Dans ce contexte il est montré en référence II (Annexe 7.2) que l’utilisation de la
prise de moyenne statistique couplée avec le modèle de la cellule permet avec succès
l’intégration des effets des inhomogénéités locales (gradients locaux des champs
physiques) sur les expressions de fermeture. Deuxièmement, le cas de la coexistence entre
les structures colonnaires et équiaxes (c/e) est analysé du point de vue du moyennage
statistique. Plus précisément, le cas de la pénétration du front colonnaire dans une zone
équiaxe développée est analysé. Ce type de couplage peut être fréquemment rencontré
pendant les phénomènes de solidification et spécialement dans le cas des phénomènes de
transition colonnaire-équiaxe (CET). Le principal avantage de la prise de moyenne
statistique réside dans le fait qu’à l’aide des distributions de probabilité des particules les
( )
plus proches, f c / e t , x; z1 , on peut quantifier d’une manière rigoureuse les influences des
particules colonnaires/équiaxes (c/e) en x. En utilisant l’approximation de la cellule pour
f c / e deux cellules sphériques centrées en x et de rayon R1c* et R1e* ont été identifiées. En
utilisant ces deux cellules il est possible de séparer l’influence des structures c/e dans le
point x en définissant des moyennes d’ensemble correspondantes à chaque structure. Ces
moyennes vont quantifier les effets moyens en x produit par chaque structure. Ainsi,
suivant le cas où la particule la plus proche de x est colonnaire ou équiaxe, le modèle nous
conduit à différentier deux types principaux de valeurs moyennes : respectivement
colonnaires et équiaxes. Dans ce contexte deux fractions importantes sont identifiées : la
fraction colonnaire et celle équiaxe, ε c et ε e . Ces deux fractions quantifient la manière
dont les deux structures partagent l’espace en x. Une autre caractéristique importante de la
moyenne statistique est le fait qu’elle nous permet de quantifier le blocage mécanique
produit par la structure équiaxe sur les pointes colonnaires. On a démontré que les
structures équiaxes et colonnaires interagissent suivant un double mécanisme de blocage.
Le premier effet de blocage mécanique déterminera une raréfaction de la zone colonnaire
au moment où cette dernière pénètre dans la zone équiaxe. En effet, à cause de la
dimension finie des grains équiaxes les pointes colonnaires disposent seulement d’un
espace limité (1 − ε ge ) dans lequel elles peuvent germer. Ainsi seulement une fraction,
1 − ε ge , du nombre initial des colonnes, nc0 , va être capable de pénétrer à travers la zone
équiaxe. Le deuxième effet de blocage mécanique reflète le fait que pour une zone mixte
(colonnaire+équiaxe), les deux structures sont forcées de partager localement un même
espace. Par conséquent, l’espace disponible pour chaque structure sera différent et plus

62
petit par rapport à l’espace correspondant disponible juste avant l’état de coexistence.
Ainsi l’évolution de la zone colonnaire+équiaxe sera très différente de celle caractérisant
une zone purement colonnaire ou purement équiaxe. Finalement, en utilisant
l’approximation de la cellule pour f c / e et les hypothèses physiques valides à l’échelle du
grain nous avons pu fermer rigoureusement le système des équations moyennes. Un
modèle complet décrivant la coexistence des grains colonnaires et équiaxes a été
finalement obtenu.
Dans une troisième partie de la thèse (référence III, Annexe 7.2) nous nous sommes
concentré sur l’utilisation du nouveau modèle mathématique pour la simulation de la
transition colonnaire-équiaxe. Premièrement nous avons effectué la simulation numérique
des solidifications quasi-stationnaires dans le but d’obtenir les cartes de la CET. Nous
avons trouvé qu’en fonction du choix de la vitesse de tirage par rapport à deux vitesses
critiques caractérisant le zone équiaxe devant les pointes colonnaires on pouvait identifier
sur la carte CET trois zones : une zone purement colonnaire, une zone purement équiaxe et
finalement une zone mixte colonnaire+équiaxe. En quantifiant le mécanisme de division
des dendrites colonnaires on a pu déterminer si la zone mixte colonnaire+équiaxe
correspondait à un état quasi-stationnaire aux pointes colonnaires. Le mécanisme de
division des colonnes (mécanisme de redensification de la zone colonnaire) a été quantifié
à l’aide d’un nouveau modèle pour l’espacement dendritique primaire. Ce modèle a à sa
base une analyse de la stabilité des pointes colonnaires par rapport aux conditions locales
de solidification (gradient thermique, vitesse des isothermes). Il a été aussi validé par
plusieurs données expérimentales. Finalement, la zone mixte colonnaire+équiaxe a été
quantifiée à l’aide de la fraction colonnaire ε c , qui à son tour quantifie rigoureusement
l’état de coexistence des deux structures. Il est important de noter que les effets de blocage
solutal et mécanique produits par la zone équiaxe sur les pointes colonnaires sont
intrinsèquement inclus dans le nouveau modèle. Ainsi ce modèle réussit à unifier
l’approche semi-empirique de Hunt [3] (le blocage purement mécanique) et celui de
Martorano et al. [4] (le blocage purement solutal). Dans une deuxième partie de ce travail
on s’est concentré sur l’utilisation du nouveau modèle pour la simulation des expériences
unidirectionnelles de solidification en [11, 12]. On a tout d’abord mis en évidence le
couplage complexe entre la zone colonnaire et la zone équiaxe devant les pointes
colonnaires. Ensuite la stationnarité du front colonnaire par rapport aux conditions locales
de refroidissement a été analysée. Nous avons ainsi pu mettre en évidence que le front
colonnaire évoluait dans un état quasi-stationnaire par rapport aux conditions locales de
refroidissement jusqu’à des moments très proche de la transition colonnaire-équiaxe. Nous
avons également trouvé que la surfusion de nucléation des grains équiaxes était très proche
de la surfusion maximale des pointes colonnaires et qu’ainsi la CET était virtuellement
indépendante de la zone équiaxe devant le front colonnaire. Ceci est en concordance avec
les résultats de Martorano et al. [4] et vient appuyer l’idée que la CET est principalement
causée par la fragmentation des dendrites colonnaires. Si la zone équiaxe n’est pas prise en
compte par le modèle on observe que la vitesse du front colonnaire subit une croissance
brusque au début de la solidification suivi par un quasi-plateau correspondant à un état
quasi-stationnaire aux pointes colonnaires et finalement, après un temps critique, suivi par
une évolution oscillatoire. Il est très intéressant de noter que le début de l’évolution
oscillatoire du front colonnaire a pu être corrélé avec la position de la CET mesurée dans
les expériences. On a montré aussi que l’évolution oscillatoire du front colonnaire
correspondait en fait à un état non stationnaire des pointes colonnaires par rapport aux
conditions locales de refroidissement. On a trouvé aussi que cette évolution oscillatoire du
front colonnaire était très favorable à la fragmentation des dendrites colonnaires et par

63
conséquent très favorable à la CET. Finalement, on a conclu que le régime non stationnaire
du front colonnaire par rapport aux conditions locales de refroidissement représentait la
cause principale de la CET, au moins pour les alliages non raffinés.
Dans une quatrième partie de la thèse (référence IV, Annexe 7.2) le modèle à prise
de moyenne statistique développé dans la deuxième publication (référence II, Annexe 7.2)
est étendu pour prendre en compte les effets de la convection du fluide. En similitude avec
le modèle initial, le nouveau modèle approche la moyenne d’ensemble à l’aide des densités
( )
de probabilité des particules les plus proches, f c / e t , x; z1 . De cette manière les
interactions solutales et mécaniques entre les deux structures colonnaire et équiaxe (c/e)
mises en évidence par le modèle initial sont aussi prises en compte par le nouveau modèle.
Toutefois, à cause de la prise en compte de l’écoulement fluide les hypothèses physiques
valides à l’échelle du grain sont réévaluées par rapport à celles initialement décrites en
référence II. Premièrement, l’écoulement fluide autour des grains perméables colonnaires
et équiaxes est analysé. Deuxièmement, la croissance du grain en présence de la
convection est décrite. Finalement, on fait une analyse du transfert de soluté autour des
grains c/e en tenant compte de la convection. On trouve ainsi que le transfert de masse à
l’interface du grain c/e en présence de la convection est plus intense par rapport au cas sans
convection. On propose des corrélations pour le calcul du coefficient de transfert de masse
global dans la limite de très petites et de grandes fractions granulaires. Ces hypothèses
physiques valides à l’échelle du grain couplées avec le modèle de la cellule nous
permettent de fermer le système complexe formé par les équations de conservation.
Premièrement les équations de conservation des grains sont dérivées. La nucléation
équiaxe est modélisée comme un événement instantané tandis que la nucléation colonnaire
(l’apparition soudaine des pointes colonnaires dans un point x) est modélisée à l’aide d’un
algorithme de suivi du front. On propose ainsi une technique simple de suivi du front ayant
à la base le modèle du volume fluide (« volume of fluid model »). De plus, les dimensions
des cellules R1c* , R1e* et le calcul des fractions colonnaires et équiaxes ( ε c et ε e ) sont ici
déterminés à l’aide d’un nouveau modèle de coexistence qui prend en compte la
convection des grains équiaxes. Les équations de conservation de masse, de la quantité de
mouvement et de l’énergie sont dérivées par la suite. Il faut remarquer que pour des raisons
de simplicité l’équation de conservation de quantité de mouvement pour le liquide est
écrite pour le fluide total et non pas séparément pour le fluide colonnaire et équiaxe. Pour
l’équation de l’énergie une seule équation de mélange est écrite. Finalement, les équations
de conservation de soluté sont dérivées. Ainsi deux équations pour le solide colonnaire et
équiaxe ainsi qu’une équation pour la conservation de soluté dans le fluide total sont
considérées. De plus, les deux équations de conservation de soluté dans le liquide inter-
dendritique sont utilisées pour le calcul du transfert de masse liquide-solide. On propose
aussi un algorithme pour le traitement implicite du calcul de la concentration dans le
liquide extra-dendritique. Ceci est fait par le couplage à chaque pas de temps de l’équation
de conservation de soluté dans le liquide extra-dendritique avec l’équation de conservation
de soluté dans le fluide total. Enfin un modèle complet à prise de moyenne statistique
d’ensemble est proposé, celui-ci étant capable de modéliser le couplage complexe entre les
structures colonnaire et équiaxe en présence de la convection.
Finalement, dans un but de validation, le modèle mathématique développé dans la
référence IV (Annexe 7.2) a été implémenté dans le logiciel FLUENT. Toutefois, il est
important de noter que pour l’instant seulement le modèle colonnaire pur a été
complètement implémenté et testé en Fluent. Dans un premier temps, à l’aide du modèle
numérique plusieurs cas de solidification dirigée sous l’influence de la convection naturelle
et/ou forcée ont été analysés. On a observé que la structure de l’écoulement à proximité du

64
macrofront a une très grande influence sur le développement des éventuels canaux
ségrégés. Le passage du cas sans brassage électromagnétique au un cas où l’écoulement
dans le bain liquide est contrôlé est très bien reflété par la structure de la zone pâteuse.
Dans le cas de la convection naturelle plusieurs canaux ségrégés isolés sont présents dans
la zone pâteuse. Par contre l’écoulement forcé dans le bain liquide imposé par le brassage
électromagnétique modifie complètement la configuration des canaux. En effet, en
contrôlant l’écoulement dans le bain liquide la distribution de pression au niveau du
macrofront est également contrôlée. Principalement contrôlé par la distribution de pression
locale, l’écoulement dans la zone pâteuse va à son tour contrôler la distribution des canaux
ségrégés dans la zone pâteuse. Effectivement, un brassage important prend naissance dans
la zone pâteuse celui-ci étant dirigé des zones de basse pression vers les zones de haute
pression. Ce brassage amène du liquide enrichi en soluté vers les zones de basse pression
en y empêchant la solidification. Ainsi des canaux ségrégés sont donc susceptibles de se
développer aux endroits où l’écoulement pâteux débouche dans le bain liquide. Toutefois
les canaux peuvent exister seulement si l’écoulement dans la zone pâteuse est
suffisamment fort par rapport aux conditions locales de solidification. Nous avons en effet
montré que les zones pâteuses subissait un processus de solidification intensifié, là où
l’écoulement est dirigé dans le même sens avec la vitesse locale des isothermes Vis . Ces
zones correspondent aux zones de haute pression où l’écoulement pénètre dans la zone
pâteuse. Au contraire aux endroits où l’écoulement pâteux se fait dans le sens contraire à la
vitesse locale des isothermes, le processus de solidification est diminué en intensité par
rapport à un cas purement diffusif. Si l’écoulement dépasse en intensité la vitesse des
isothermes, la solidification est stoppée et un processus de refusion prend naissance
localement. Ainsi, un canal ségrégé peut naître et subsister dans le temps seulement si
l’écoulement pâteux local est suffisamment fort pour que la vitesse locale du liquide soit
plus grande ou égale à la vitesse locale de l’isotherme. La présence d’un canal détermine
aussi de fortes ségrégations dans la zone pâteuse : le canal agit comme une pompe de
soluté drainant la zone pâteuse. On a montré aussi que le fort appauvrissement de la zone
pâteuse autour d’un canal provenait du fait que la concentration moyenne caractérisant le
flux de masse entrant dans la zone pâteuse était plus petite par rapport à la concentration
moyenne caractérisant le flux de liquide sortant de la zone pâteuse et débouchant dans le
canal. On conclut ainsi que l’utilisation d’un brassage électromagnétique glissant ou rotatif
ne permet pas d’annuler les macroségrégations dues aux canaux ségrégés. Par contre
l’avantage de l’utilisation du brassage électromagnétique réside dans le contrôle de la
position des canaux et donc la localisation précise des macroségrégations positives
correspondantes, en opposition donc avec le cas sans brassage où la distribution des canaux
était aléatoire (les freckles).

Ensuite, dans un but de validation, le modèle numérique a été utilisé dans la

modélisation de l’expérience de solidification dirigée sous l’influence du brassage
électromagnétique rotatif, expérience réalisée au laboratoire ACCESS. L’accord entre les
résultats numériques et les résultats expérimentaux est aussi bon tant sur le plan qualitatif
que quantitatif. D’une part le modèle reproduit le large canal central induit par la
convection forcée dans le bain liquide : les dimensions du canal central sont très proches
des dimensions obtenues expérimentalement. D’autre part les ségrégations radiales et
longitudinales sont en bon accord avec les résultats expérimentaux à l’exception de la zone
centrale qui est légèrement surestimée. La cause de ce désaccord semble être la non-
modélisation de la zone équiaxe. Finalement l’élargissement du canal dans le temps est
aussi qualitativement reproduit par le modèle. La cause de cet élargissement réside dans la
modification des recirculations secondaires au niveau du macrofront. Ces dernières sont à

65
leur tour dues au ralentissement du macrofront et au rapprochement du plan axial de
l’aimant du macrofront.

En conclusion, les objectifs principaux de cette thèse, à savoir :

ƒ Le développement d’un modèle mathématique eulérien à prise de moyenne statistique
capable de modéliser le couplage complexe entre les structures colonnaire et équiaxe,
ƒ La validation de ce modèle avec des résultats expérimentaux et notamment avec des
expériences de solidification dirigée faisant intervenir le phénomène de transition
colonnaire-équiaxe,
ƒ L’implémentation du modèle dans le logiciel FLUENT et la validation de celui-ci avec
des expériences,
ont été atteints. Toutefois, il est important de noter que pour l’instant seul le modèle
colonnaire pur a été complètement implémenté et testé dans Fluent. Une version du modèle
équiaxe pur est en phase de développement mais les résultats actuels n’ont pas permis une
validation approfondie de celui-ci. Comme perspectives de continuation de ce travail on
note :
ƒ L’implémentation dans Fluent du modèle mathématique permettant la modélisation de
la structure équiaxe pure et la validation de celui-ci,
ƒ L’implémentation dans Fluent et la validation du modèle mathématique complet
développé en référence IV (Annexe 7.2) permettant la simulation du couplage
équiaxe/colonnaire en présence de la convection.

66
6 Références
1. Sarazin, J.R., and A. Hellawell: "Channel Formation in Pb-Sn, Pb-Sb and Pd-Sn-Pb
Alloy Ingots and Comparison With the System.", Metallurgical Transactions A, 1988,
vol. 19A, pp. 1861-1871.

2. Jackson, K.A., et al.: "On the Origin of the Equiaxed Zone in Castings.", Transactions
of the metallurgical society of AIME, 1966, vol. 236, pp. 149-158.

3. Hunt, J.D.: "Steady State Columnar and Equiaxed Growth of Dendrites and Eutectic.",
Materials Science and Engineering, 1984, vol. 65, pp. 75-83.

4. Martorano, M. A., C. Beckermann, and Ch.-A. Gandin: "A Solutal Interaction

Mechanism for the Columnar-To-Equiaxed Transition in Alloy Solidification.",
Metallurgical and Materials Transactions A, 2003, vol. 34A, pp. 1657-1674.

5. Drew, D.A., and S.L. Passman. Theory of Multicomponent Fluids. . New York:
Springer-Verlag, 1999.

6. Quillet, G. "Influence de la Convection, Naturelle ou Forcée, sur l'apparition des

Mésoségrégations lors de la Solidification des Alliages Métalliques.", Thèse de
doctorat, INPG-ENSHMG, Grenoble, 2003.

7. Ciobanas, A.I. "Modélisation numérique de la solidification interdendritique. Influence

du brassage électromagnétique sur les macroségrégations. ", DEA de l’INPG, 2002

8. Felicelli, S.D., J.C. Heinrich, and D.R. Poirier: "Simulation of Freckels During Vertical
Solidification of Binary Alloys.", Metallurgical Transactions B, 1991, vol. 22B, pp.
847-859.

9. Rappaz, M., and Ph. Thévoz: "Solute Diffusion Model for the Equiaxed Dendritic
Growth.", Acta Metallurgica, 1987, vol. 35.7, pp. 1487-1497.

10. Wang, C.Y, and C. Beckermann: "A Multiphase Solute Diffusion Model for Dendritic
Alloy Solidification.", Metallurgical and Materials Transactions A, 1993, vol. 24A, pp.
2787-2802.

11. Mahapatra, R.B., and F. Weinberg: "The Columnar to Equiaxed Transition in Tin-Lead
Alloys.", Metallurgical Transactions B, 1987, vol. 18B, pp. 425-432.

12. Ziv, I., and F. Weinberg: "The Columnar-To-Equiaxed Transition in Al 3 Pct Cu.",
Metallurgical Transactions B, 1989, vol. 20B, pp. 731-734.

13. Carman, P.C. Flow of Gases Through Porous Media. . London: Butterworth Scientific,
1956.

14. Medina M. "Ségrégation en solidification en présence de convection forcée ", Thèse de

doctorat, INPG – ENSHMG, Grenoble, 2000.

67
15. Worster, M.G.: "Convection in Mushy Layers.", Ann. Rev. Fluid Mech., 1997, vol. 29,
pp. 91-122.

16. Bejan, A. Shape and Structure, From Engineering to Nature. . New York: Cambridge
University Press, 2000.

68
7 Annexes
7.1 Les propriétés physiques des alliages utilisés
Tableau 8: Propriétés de l’alliage Pb – 10wt% Sn
Coefficient de conduction thermique du solide, λs W .m −1.K −1 24,65
Coefficient de conduction thermique du fluide, λs W .m −1.K −1 24,65
Chaleur latente de solidification, L J .kg −1 40000
Chaleur spécifique du solide, cs J .kg −1.K −1 145
Chaleur spécifique du fluide, c f J .kg −1.K −1 145
Coefficient d’expansion thermique, βT K −1 −1, 2 ×10−4
Coefficient d’expansion solutale, β C ( wt % )
−1
−5,15 × 10−3
Coefficient de diffusion solutale dans le solide, Ds m2 .s −1 0
Coefficient de diffusion solutale dans le liquide, m 2 .s −1 3 ×10−9
Df
Viscosité dynamique, µ kg.m −1.s −1 1,8 ×10−3
Mass volumique du solide, ρ s kg.m −3 10000
Mass volumique du liquide, ρl kg.m −3 10000
Coefficient de partage, k -- 0,31
Pente du liquidus, m K . ( wt % )
−1 −2,32
Température de fusion du Pb pur, T f K 600, 65
Température du liquidus à C0 , Tliq K 577, 4
Concentration eutectique, CE wt % 61,9
Température eutectique, TE K 456, 42

Tableau 9: Propriétés de l’alliage Al –7wt% Si

Coefficient de conduction thermique du solide, λs W .m −1.K −1 100
Coefficient de conduction thermique du fluide, λs W .m −1.K −1 100
Chaleur latente de solidification, L J .kg −1 400000
Chaleur spécifique du solide, cs J .kg −1.K −1 1140
Chaleur spécifique du fluide, c f J .kg −1.K −1 1140
Coefficient d’expansion thermique, βT K −1 1, 2 × 10−3
Coefficient d’expansion solutale, β C ( wt % )
−1
−2,5 ×10−4
Coefficient de diffusion solutale dans le solide, Ds m2 .s −1 ≅0
Coefficient de diffusion solutale dans le liquide, m .s2 −1
3 ×10−9
Df
Viscosité dynamique, µ kg.m −1.s −1 2,52 ×10−3

69
Mass volumique du solide, ρ s kg.m −3 2452
Mass volumique du liquide, ρl kg.m −3 2452
Coefficient de partage, k -- 0,13
Pente du liquidus, m K . ( wt % )
−1 −6, 62
Température de fusion de l’Al pur, T f K 933, 45
Température du liquidus à C0 , Tliq K 887,11
Concentration eutectique, CE wt % 12, 6
Température eutectique, TE K 850

70
7.2 Les références en anglais

Référence I

71
INSTITUTE OF PHYSICS PUBLISHING JOURNAL OF PHYSICS D: APPLIED PHYSICS
J. Phys. D: Appl. Phys. 39 (2006) 5252–5266 doi:10.1088/0022-3727/39/24/023

Dendritic solidification morphology

viewed from the perspective of constructal
theory
A I Ciobanas1 , A Bejan2 and Y Fautrelle1
1
EPM/CNRS Laboratory, ENSHMG, BP 95, 38402 Saint Martin d’Hères cedex, France
2
Duke University, Department of Mechanical Engineering and Materials Science, Durham,
NC 27708-0300, USA
E-mail: aciobanas@yahoo.com

Received 16 July 2006, in final form 23 October 2006

Published 1 December 2006
Online at stacks.iop.org/JPhysD/39/5252

Abstract
In this paper we focus on the application of the constructal theory in
predicting the dendritic solid structure. First we analyse the marginal
stability criterion from the perspective of the constructal principle. Having as
a guiding principle the constructal law we have shown that among the whole
range of possible dendrite tip radiuses predicted by the stability analysis the
dendrite tip will choose the smallest one, that is a radius equal with the
smallest perturbation wavelength leading to instabilities. We identify as well
the existence of a competition between the diffusion controlled growth and
the dendritic growth. Second, we develop a model for the secondary arm
spacing. We identify a competition between the lateral diffusion controlled
growth of a needle and the dendritic growth of lateral secondary arms. By
analysing this competition we are able to characterize the sidebranching
mechanism and to finally compute the secondary arm spacing. The result is
in good agreement with the experimental results. Finally, the primary arm
spacing is analysed from the perspective of the constructal law. The
constructal law predicts that the only way the columnar tips can optimize the
solidification process is to minimize the spacing between two adjacent tips,
namely λ1 . By quantifying the two mechanisms responsible for the selection
of λ1 , the dendrite division and the dendrite overgrown mechanisms, we
were finally able to obtain a model for the primary arm spacing. This model
is also validated against various experimental data.

Nomenclature Tts Isotherm corresponding to the (K)

quasi-steady columnar tips
Symbol Meaning Unit xst Position of the quasi-steady colum- (m)
C Solute concentration field (wt%) nar tips
Cs∗ Interfacial concentration at the (wt%) G Thermal gradient (K m−1 )
solid interface Gc Solutal gradient at the solid–liquid (wt% m−1 )
Cl∗ Interfacial concentration at the (wt%) interface
liquid interface λ1 Primary arm spacing (m)
T Temperature field (K) λ2 Secondary arm spacing (m)
Tliq Liquidus temperature correspond- (K) λmin Minimum perturbation wavelength (m)
ing to the initial concentration C0 leading to instabilities
Tf Fusion temperature of the pure (K) α Thermal diffusivity (m2 s−1 )
solvent Ds Chemical diffusivity of the solute (m2 s−1 )
in the solid
0022-3727/06/245252+15$30.00 © 2006 IOP Publishing Ltd Printed in the UK 5252
 Dendritic solidification morphology

Dl Chemical diffusivity of the solute (m2 s−1 ) solid structure that forms for the most rapid solidification given
in the liquid certain physical constraints (i.e. given undercooling, isotherm
m Slope of the liquidus line (K wt%−1 ) velocity or thermal gradient). As we will point out, when seen
k Partition coefficient from the perspective of the constructal law, the sidebranching
trs Return time to the steady state of a (s) mechanism appears as a needed mechanism in order to counter
tip perturbed from its initial quasi- the slower lateral growth (diffusion controlled growth) of the
steady position primary arm. By quantifying the competition between the
tdw Time scale characterizing the (s) diffusion controlled growth and the dendritic growth we are
strength of the solutal interactions able to predict the secondary arm spacing.
between two adjacent columnar Concerning the primary arm spacing, various models exist
tips in the literature [6–12]. We will depart once more from the
Rt Dendrite tip radius (m) classical approach to the prediction of primary arm spacing
Vt Dendrite tip velocity (m s−1 ) by forcing ourselves to look at the columnar structure from
Vis Isotherm velocity (m s−1 ) the perspective of the constructal law. We conclude that
Pe Tip Peclet number (= Rt Vt /2Dl ) the natural tendency of the columnar tips is to minimize the
 Dimensionless undercooling average spacing between them. In fact, by minimizing the
primary arm spacing (λ1 ) the columnar structure is optimizing
1. Introduction the solidification process behind the columnar tips. By
identifying and quantifying the two mechanisms responsible
It is well known that a columnar dendritic structure can be for the selection of a unique primary arm spacing (the
characterized with a precise length scale named the primary dendrite division mechanism and the overgrown mechanism)
arm spacing. In fact the primary arm spacing is a measure of we are finally able to obtain a model for λ1 . This model
the average spacing between two adjacent columnar tips. This is subsequently validated against various experimental data
length scale defines a quasi-steady state columnar structure [2, 13–15].
and hence was directly linked to the thermal gradient and The paper comprises four sections. In section 2, the
the isotherm velocity defining the local cooling conditions. marginal stability criterion is analysed from the perspective
Moreover, both the columnar and the equiaxed dendritic of the constructal theory. As well, we introduce the concept
structures can be characterized with a smaller length scale, of ‘competition’ between the diffusion controlled growth and
the secondary arm spacing. The secondary arm spacing is a the dendritic growth. Using this concept we are able to predict
measure of the length scale between two adjacent secondary the moment corresponding to the solid–liquid destabilization.
dendrite arms and it is usually an order of magnitude smaller In section 3 the secondary arm spacing is also analysed from
than the primary arm spacing. The presence of primary the perspective of the constructal theory. The proposed model
and secondary dendrite arms within an equiaxed or columnar is in very good agreement with the model of L–M [1] and
zone thus contributes to the complexity of the dendritic solid with different experimental data [2, 16]. Finally throughout
structure. Hence, the existence of a model which could section 4 a model for the primary arm spacing is developed.
predict these two important lengths scales would be of great Again the concept of the selection of the primary arm spacing
interest. In fact various models were developed in the past is analysed from the perspective of the constructal theory. The
decades aiming to predict both the primary and secondary model is also tested against different experimental data.
arm spacing. Concerning the secondary arm spacing, Lager
and Muller-Krumbhaar (L–M) [1] developed a theoretical
2. The stability criterion
model to predict the stability of dendrite tips and the inherent
instability that gives rise to sidebranches (secondary arms).
One should note that the theoretical findings in the two
Their model predicts that the secondary arm spacing (λ2 ) is
following sections are entirely based on the findings of Bejan
almost equal to the double of the primary arm dendrite tip
et al [4,5]. However we introduce here two original approaches
radius, Rt . One should note that this result was subsequently
regarding the stability criterion and the computation of the
validated experimentally [2] and it was observed that the L–M
secondary arm spacing.
model fits very well with the experimental data. One should
In [5], in order to quantify the evolution of a solid region in
however note the high complexity of the stability analysis
space and time Bejan relied entirely on the constructal law [4].
which stands behind this relatively simple result. In this
Basically the constructal law states that the geometric form of
respect we will present in the following study a more original
a freely changing flow system develops in time in such a way
(and simpler) approach to the secondary arm spacing. In
that the currents find the path of maximal access. Applied to
fact we will propose a model for the secondary arm spacing
a solutal controlled solidification the constructal law means
developed from the perspective of the constructal law [3–5].
that [5]:
Basically the constructal law states that the geometric form
of a freely changing flow system3 develops in such a way 1. the heat currents generated on the solidification fronts will
that the currents find the path of maximal access. In this shape their fronts such that they flow with least resistance;
context one should see the dendritic structure as an optimal 2. the solid region grows as fast as possible;
3 Flow does not refer to fluid flow only but has a more general meaning in 3. equilibrium is reached in minimal time because of the
the sense of energy flows (heat, mass, etc). generation of geometric structure.

5253
A I Ciobanas et al

to the phase diagram (the liquidus line) and can be computed

with respect to the local temperature. Usually one has
T − Tf
Cl∗ = , (4)
m
where m is the liquidus line slope and Tf the fusion temperature
of the pure solvent. The main consequence of this particular
solid behaviour is the continuous enriching of the liquid phase
due to the inherent solute rejection at the solid–liquid interface.
The rejected solute will diffuse away from the interface as
the solid–liquid interface is advancing (see figure 1(b)). In
fact for binary alloys the solid growth is mainly controlled
by this solute diffusion process at the solid–liquid interface.
Figure 1. (a) The isothermal solidification of a solid nuclei; (b) the Moreover, the solute conservation written at the solid–liquid
concentration distribution around the solid grain.
interface enables one to link the solid growth rate and the solute
fluxes at the solid–liquid interface. One has
Even if the study in [5] focuses on the solidification of a
pure substance we will in the following consider the case of a ∂Cl ∂Cs
(Cl∗ − Cs∗ ) V = −Dl + Ds , (5)
binary alloy. Let us consider a volume V (see figure 1(a)) ∂nl ∂ns
filled with a liquid binary alloy, having a uniform solute
concentration C0 and a uniform temperature T0 . We will where V is the normal velocity of the solid–liquid interface.
consider as well that the liquid is undercooled, meaning that Since usually Ds  Dl the solute flux in the solid can be
its temperature is lower than the liquidus temperature, Tliq , neglected.
corresponding to C0 . One has Due to the liquid undercooling, subsequent to the solid
nucleation, the concentration at the solid–liquid interface will
T0 = Tliq − T0 . (1) jump to Cl∗ (> C0 ). Hence the solid nucleation will trigger a
spherical solute diffusion wave around the nucleation site. The
We will also consider that the volume V is maintained at a distance travelled by this diffusion wave can be expressed as a
constant temperature T0 . Let us suppose now that at t = 0 s function of time as follows:
a solid germ nucleates within the liquid phase. Due to the
undercooled state of the liquid phase the solid germ will start x = (Dl t)1/2 . (6)
to grow for t > 0. Moreover, due to the relative high thermal
A solutal boundary layer will develop around the solid grain
diffusivity of the liquid metals the latent heat will be rapidly
triggering a non-zero solute flux at the solid–liquid interface.
diffused away from the solid–liquid interface. In this respect,
In fact, it is precisely this solute flux which also drives the
one can consider that a thermal equilibrium exists at the scale
grain growth. This particular free boundary diffusion problem
of the solid grain and that the solidification process is very
is no more than the Stefan free boundary problem. Written for
close to an isothermal one (T = T0 = ct). One should also
plane solid interface (see figure 2) the Stefan problem writes
note that for liquid metals the Lewis number,
as follows:
α
Le = , (2) ∂C ∂ 2C
Dl =D 2,
∂t ∂x
is very large. Hence, the thermal diffusion process at the solid– C (x = ∞, t) = C0 ,
(7)
liquid interface will be much faster than the solutal diffusion C (x = xsf , t) = Cl∗ ,
process. Therefore unlike for pure substances where the solid

∗ dxsf ∂C
growth is controlled by the thermal diffusion process of the Cl (1 − k) = −D .
dt ∂x xsf
latent heat away from the moving solid–liquid interface, for
binary alloys the solid growth is mainly driven by the solute The solution to the Stefan problem can be easily found in
diffusion process at the solid–liquid interface. the literature [17]. The main result concerns the time evolution
One interesting aspect of the binary alloy solidification of the planar solid front. Indeed, for the case of small St
is that usually the solubility of the solute in the solid is lower numbers, one obtains that the position of the solid front as
than the solubility in the liquid phase. As a consequence, as the a function of time is
solidification process evolves, the growing solid phase accepts
less solute than the liquid is actually able to offer. Basically xsf ∼
= (2StDl t)1/2 , (8)
the concentration at the solid interface is smaller than the one
at the liquid interface (see figure 1(b)). One has where St is the dimensionless Stefan number. For the
particular case of a binary alloy St number is no more than
Cs∗ = kCl∗ , (3) the dimensionless undercooling:

where k < 1 is called the partition coefficient and Cl∗ is the Cl∗ − C0
St (= ) = . (9)
solute concentration at the liquid interface. The latter is linked Cl∗ − Cs∗

5254
 Dendritic solidification morphology

Figure 3. The solid grain radius Rsf and the solid growth rate Vsf as
a function of time for an Al–3wt%Cu alloy ( = 0.3).
Figure 2. The uni-dimensional Stefan problem.
decreasing growth rate, see equation (11), in respect with the
Even if the problem analysed above is valid for a planar solid constructal law, the solid will be forced to select a different
front, one can assume that the central result in (8) applies as geometry, a more effective one. As pointed out by Bejan
well for the case of growing sphere but only as an order of et al [5] the geometric alternative for the solid is to invade
magnitude estimate; that is, the liquid in the form of needles.
The needles which are at the base of the dendritic form of
Rsf ∼ (2  Dl t, )1/2 , (10) the solid are a direct consequence of the solid–liquid interface
instabilities. Indeed, given some particular solidification
where Rsf is the solid grain radius. An interesting aspect of conditions (thermal gradient G and solutal gradient Gc at
the diffusion controlled growth in (10) is that the solid growth the interface) the solid–liquid interface becomes instable to
rate small perturbation of its shape. The Mullins–Sekerka [18]
dRsf 1
Vsf ∼ ∼ (2  Dl )1/2 t −1/2 (11) linear stability analysis provided, for the case of a planar front,
dt 2 the conditions under which the planar solid–liquid interface
is continuously decreasing in time. In fact at the very beginning becomes instable. Indeed, using the Mullins–Sekerka results,
of the solidification, due to the infinite propagation speed of the Kurtz and Fisher [7] computed the perturbation wavelengths
diffusion wave, the solid growth rate is also infinite, equation range leading to instabilities. Their central result states that
(11). However, soon after the nucleation, the solid growth rate for a perturbation wavelength λ,
decreases rapidly (Vsf ∝ t −1/2 ). This is mainly due to the rapid
1/2
increase in the solute boundary layer (δ) around the spherical 
λ > 2π , (14)
solid grain. This boundary layer can be easily estimated from mGc − G
the solute conservation at the solid–liquid interface:
the solid–liquid interface becomes instable. In (14) Gc is the


∗ ∂C C ∗ − C0 solutal gradient and G is the thermal gradient at the solid–
Cl (1 − k) Vsf = −D =D l . (12) liquid interface. As Kurtz and Fisher pointed out, one expects
∂r Rsf δ
the dendrite point radius (the needle radius) to be equal to the
Using the result in (11) one can easily obtain that perturbation wavelength λ, namely that

δ ∼ (2  Dl )1/2 t 1/2 ∼ Rsf . (13) Rt = λ. (15)

This means that the solutal boundary layer around the grain This result coupled with equation (14), would mean that
is of the same order of magnitude as the radius of the solid for a given directional solidification experiment one should
grain. In figure 3 the representation of the grain radius and observe various dendrite tip radii corresponding to different
of the grain growth rate as a function of time is presented perturbations wavelengths within the range determined in (14).
for an Al–3wt%Cu alloy and an undercooling of  = 0.3. However in reality, one observes a unique radius for the
The Al–3wt%Cu alloy physical properties can be found in dendrites tip. The solution to this problem was given by Langer
table 1. We can thus state that the diffusion controlled growth and Muler-Krumbhaar [19]. Indeed, they proposed a selection
is very effective at the very beginning of the solidification and criterion (the marginal stability criterion) for the dendrite tip,
rapidly decreases in efficiency after the nucleation. From the namely that the radius of the dendrite tip is equal to the
point of view of the constructal law the natural tendency in the minimum Mullins–Sekerka wavelength leading to instabilities.
evolution of the solid structure will be the maximization of the That is,
1/2
radius Rsf through the generation of the solid geometry. Since 
Rt = λmin = 2π . (16)
the maintenance of a spherical geometry means a continuous mGc − G

5255
A I Ciobanas et al

We note that this result was extensively validated with the

experiments. In this context, an interesting question would
be why the solid selects a precise geometry among its infinite
possibilities, equation (14). Even if L–M answered this
question by means of a complex stability analysis we will
hereafter propose a simple explanation by citing the constructal
principle. Indeed, the geometric form of the solid will develop
in such a way that the solidification process will be the
most effective one. As already pointed out, applied to the
solidification of a binary alloy, the constructal law predicts
that the shape of the solid shape will adapt in order to grow
as fast as possible. Let us consider the case of an isothermal
solidification (G = 0) of a binary alloy. Putting G = 0 in (14)
and using equation (15) we can state that the range of possible
radiuses of the dendrite tip (needle) is

1/2

Rt > 2π . (17)
mGc Figure 4. Diffusion controlled growth versus dendritic growth
(Al–3wt%Cu alloy,  = 0.3).
Using now the solute balance at the solid–liquid interface,
Cl∗ (1 − k)V = −Dl Gc , the above inequality can be rewritten instability. Moreover Bejan [4] used this particular observation
as follows: to predict the laminar–turbulent transition and the obtained

 results are very convincing. In this context it appears that the
Rt V 2 V
> π2 , (18) turbulent structure and dendrites in solidification are analogous
2Dl m(k − 1)Cl∗ Dl
manifestations of the constructal phenomenon of generation
where V is the growth velocity of the solid needle. We of flow configuration. In fact, in an isolated system the
recognize in the LHS of the above equation the tip Peclet generation of configuration brings the nonequilibrium system
number P e = Rt V /2Dl . If we approach the needle tip with to equilibrium the fastest.
a hemisphere we can further link the Pe number to the tip Returning now to the evolution of the solid nuclei we are
dimensionless undercooling  [7]: able to state that due to the continuous decrease in the solid
growth rate (figure 3) the solid is forced to select a different
P e = . (19) more effective geometry. We concluded that a valid alternative
for the solid shape is the needle (the dendrite tip). Indeed, the
Now one can easily obtain most interesting feature of the needle is that due to its slender
m(k − 1)Cl∗ Dl 2 shape the dendrite tip is able to advance in the undercooled
V <  (20) liquid with a constant velocity:
π 2
meaning that the growth velocity of the solid needle will be m(k − 1)Cl∗ Dl 2
Vt =  . (22)
always smaller than a maximum velocity. One can easily π 2
recognize that the RHS of (20) is no more than the dendrite tip
Note again that the above equation is valid if we assume a
velocity obtained if the marginal stability criterion of L–M is
spherical shaped dendrite tip. Given a needle shape for the
used [19]; that is,

 solid one can easily see that the radius of the solid region
 1/2 would increase linearly with time:
Rt (= λmin ) = 2π
mGc
 1/2 Rsf = Vt t. (23)
Dl
= 2π . (21)
m (k − 1) Cl∗ V If the above equation is plotted against the diffusion controlled
The constructal principle enunciated above says that the needle growth in equation (10) one can easily see (figure 4) that
shape will adapt in such a way as to grow with a maximum at the beginning of the solidification the diffusion controlled
velocity. Looking at (20) it is now easy to see that a growth is more effective than the dendritic solidification (the
maximum dendrite tip velocity can be achieved only if among needle geometry). However soon after nucleation, the needle
the whole range of possible radii, equation (17), the dendrite geometry wins over the diffusion controlled growth. With
tip chooses the smallest one, equation (21). We can thus respect to the constructal principle this would also mean that
see how the constructal law manages to explain in a simple at about tcr (see figure 4) from the start of the solidification
way the difficult concept of the marginal stability criterion. the solid interface will select a more effective shape, that is the
It is interesting to note that the choice of the minimum needle geometry. These findings are very interesting since they
stable tip radius is somewhat similar to the phenomenon of are evidence of a competition between the diffusion controlled
laminar–turbulent transition. Bejan [20] showed that the growth and the dendritic solidification. It is worthwhile to
choice between the viscous diffusion and the ‘needles’ of note that the competition between these two solidification
eddy flow is also made at the smallest (neutral) wavelength of regimes has been observed numerically by means of phase

5256
 Dendritic solidification morphology

field simulations of solid nuclei growth [21]. Note as well

that figure 4 is the same as the figure for eddy formation in
turbulence presented in [4]. This emphasizes once more the
similarity between the turbulent structure and the formation of
dendrites in solidification. It would be interesting to compute
the critical time tcr . This can be easily achieved by equalling
(10) and (23). One can easily obtain
2Dl
tcr = . (24)
Vt2

The critical radius at which the shifting of the geometry occurs

therefore becomes
2Dl
Rcr = . (25)
Vt
Using now the result in equation (19) one can easily see that

Rcr = Rt . (26)
Figure 5. The grain radius (Rsf ) and the minimum perturbation
The above result is remarkable since it states that the critical wavelength leading to instabilities (λmin ) as a function of time
radius of the solid nuclei which marks the shift between the (Al–3wt%Cu alloy,  = 0.3).
two growth regimes, the diffusion controlled and the dendritic
regime, is no more than the stable dendrite tip radius. As Rsf , it is easy to see that the grain interface becomes unstable
Bejan et al [5] observed, tcr and Rcr represent no more than the for t > t ∗ . In this context t ∗ will be a measure of the time
smallest time and length scales of the dendritic structure, the at which the destabilization of the interface will take place.
‘heartbeat and capillarity of the much more complex dendritic Furthermore, t ∗ will quantify the time at which the grain growth
structure that develops as time progresses’. commutes from a diffusion controlled regime to a dendritic
One may say that the concept of competition between regime. It would be interesting to compute this critical time.
diffusion controlled growth and dendritic growth does not Putting λmin = Rsf one can easily obtain
have a rigorous physical explanation. However as we will
 2
show in the following, one can easily explain this particular π 2 8Dl
phenomenon using a more rigorous stability analysis. Indeed t∗ = 8Dl = . (28)
m(k − 1)Cl∗ Dl 2 Vt2
we have shown that just after the solid nucleation a solute
diffusion wave starts to propagate around the solid grain. This It is easy to see now that,
will determine the creation of a solutal boundary layer around
the grain having a thickness of the same order of magnitude t ∗ = 4tcr , (29)
as the grain radius, equation (13). By knowing the solutal
obtaining thus that the destabilization time t ∗ computed
gradient at the solid–liquid interface one can easily compute
rigorously from a stability analysis is of the same order of
the minimum perturbation wavelength leading to instabilities
magnitude as the critical time tcr predicted based on the
using the equation (14). One can easily obtain
 1/2 constructal principle. Moreover one obtains that the critical
 radius at which the interface destabilization takes place is
λmin = 2π  
−m Cl∗ − C0 /δ 4Dl
 1/2
 R∗ = = 2Rcr . (30)
 2Dl 1/4 1/4 Vt
= 2π t . (27)
mCl∗ (k − 1)  We thus conclude that the results obtained by applying the
It would be interesting to compare now the minimum constructal principle are of the same order of magnitude as
perturbation wavelength λmin with the radius of the grain the ones obtained by means of a more complex stability
Rsf , equation (10) (see figure 5). One can see that at the analysis. Hence, the idea of a competition between the
very beginning of the solidification the minimum wavelength diffusion controlled growth and the dendritic growth and
leading to instabilities is larger than the actual radius of the consequently between two solid geometries (the sphere and
grain. One should also note that a real interface perturbation the dendrite tip) has a secure physical basis. In this context
cannot have a wavelength larger than the actual radius of the the constructal principles do not appear as a collection of ad
grain. In fact one expects the perturbation wavelengths to be hoc established statements but as a physical reality having a
of the same order of magnitude as the grain radius. Hence, rigorous explanation.
during this first stage of the grain solidification (λmin > Rsf )
the solid grain will be stable to perturbations of its solid– 3. The secondary arm spacing
liquid interface. However soon after the solid nucleation (at
t ∗ , see figure 5) the minimum perturbation wavelength leading We have therefore shown that in view of the constructal
to instabilities (λmin ) equals the grain radius Rsf . Since the law the solid grain will choose the most effective geometry
interface perturbations are of the same order of magnitude as (either a sphere or a dendrite tip) with respect to the local

5257
A I Ciobanas et al

Figure 7. The evolution of the initial solid nuclei into a dendritic

Figure 6. The growth of the dendrite tips from the initial spherical equiaxed grain. One can observe the increasing undercooled region
solid grain. that forms between the primary dendrite arms.

conditions at the solid–liquid interface (i.e. the local solute

gradient). However one should note that when the solid–
liquid destabilization triggers, the dendrite tip will not develop
randomly around the initial spherical grain. In fact, and as
Bejan [4,5] pointed out, the needles cannot form at every point
on the surface of the solid sphere because each needle needs
space to breathe. In order to survive each needle needs a certain
amount of undercooled liquid region around it. Hence needles
can only grow from a few discrete points on the surface of the
initial solid sphere. Indeed, the angles between the needles are
controlled by the inner molecular structure of the substance.
For instance, for metallic alloys which form a body-centred
cubic crystal the dendrite tips will grow at an angle of 90◦ . Figure 8. The dendrite arm viewed as a string of discs growing
Hence a total of six dendrite tips will develop from the initial following a diffusion controlled regime.
solid sphere (see figure 6). Note that the dendrite tips will
grow provided that they advance into a reservoir of undercooled where x − x0 is the distance measured from the dendrite tip
liquid. and Vt is the tip velocity (see figure 8). It is obtained therefore
At this point it would be interesting to focus on the that the dendrite tip has a parabolic shape. Or it is well known
third identified constructal principle. This one states that the that in reality the tip does not have a hemispheric shape but a
geometric structure of the solid adapts in such a way that parabolic one. For a parabolic shaped tip the tip Peclet number
the solutal equilibrium is reached in minimal time. If one would no longer be equal to the dimensionless undercooling
carefully analyses the solid structure of the four dendrite points  as for the spherical tip. In reality one has
in figure 7 one can easily observe that as the tips grow and
advance into the undercooled melt an increasing amount of P e = I v −1 (), (33)
undercooled liquid is left between the tips. The increasing
undercooled space between the dendrite tips therefore calls where I v −1 represents the inverse Ivantsov function [22] (a
for an optimization of the solid structure. complex integral function). However even though the dendrite
To be able to predict the optimal geometric structure one tip has a parabolic shape the tip Peclet number P e = Rt Vt /2Dl
should first analyse in more detail a single dendrite tip. As will still be of the same order of magnitude as the local
Bejan et al [5] pointed out the needle could be seen as a straight dimensionless undercooling. Putting thus P e ∼  in (32)
line string of discs that are growing following a diffusion one finally obtains
controlled regime, thus obeying equation (10). Hence the tip
R ∼ [Rt (x − x0 )]1/2 . (34)
arriving at point x0 at time t0 (figure 8) will trigger the growth
of a disc of radius R: We have thus obtained that the dendrite tip radius is of the
R ∼ [2  Dl (t − t0 )] 1/2
. (31) same order of magnitude as the minimum Mullins–Sekerka
wavelength leading to instabilities, equation (21). In fact,
Points where solidification has been triggered earlier are from equation (34) one has that the dendrite tip radius is Rt /2.
situated near the ‘root’ of the needle and more recent points Or the marginal stability criterion states that the dendrite tip
are positioned near the tip (see figure 8). In sum, at a fixed radius is precisely equal to the minimum Mullins–Sekerka
time t the transversal radius of the dendrite trunk is wavelength leading to instabilities. Hence the real dendrite
 1/2 shape is described by the following law:
2  Dl
R∼ (x − x0 ) , (32)
Vt R = [2Rt (x − x0 )]1/2 , (35)

5258
 Dendritic solidification morphology

where R is the transversal radius of the dendrite at a coordinate

x0 behind the dendrite tip and Rt is the dendrite tip radius
verifying equation (21). However, using equation (21) solely
does not permit the computation of the tip radius and the tip
velocity. Therefore one needs one more equation linking Rt
and Vt . This one is obtained from the solution of the stationary
solute transport at the parabolic tip. A solution to this problem
has been found by Ivantsov [22]. Ivantsov was able to link
the tip Peclet number P e = Rt Vt /2Dl to the dimensionless
undercooling  by means of equation (33). By coupling now
equation (33) with the marginal stability criterion, equation
(21), the tip velocity and the tip radius can be computed
analytically. An interesting particularity of the parabolic shape
of the dendrite is that behind the dendrite tip the solid spreads
laterally as
R = (2Rt Vt )1/2 (t − t0 )1/2 , (36)
meaning that the lateral solid growth rate
dR 1
= (2Rt Vt )1/2 (t − t0 )−1/2 (37)
dt 2 Figure 9. The lateral growth of the dendrite at a fixed position x0 :
the full line represents the diffusion controlled growth and the
is continuously decreasing in time. This property is dashed line the dendritic growth of lateral secondary arms.
responsible for the slenderness of the dendrite and, at the same
time, for the increasing amount of undercooled liquid between
meaning that
the primary dendrite tips (see figure 7). In this context the
xcr = 2Rt . (40)
third constructal principle calls for an optimization of the solid
structure in order to minimize the time during which solutal The distance xcr has a very important meaning. In fact, as the
equilibrium is reached between the primary dendrite arms. primary dendrite trunk is advancing into the undercooled melt,
Since the maximum solid velocity is the dendrite tip velocity, secondary arms will always develop at a distance xcr behind
the fastest way to invade the undercooled liquid between the the primary dendrite tip (see figure 10). In this respect, as
primary dendrite arms would be if secondary dendrites arms the primary arm is advancing, an array of secondary dendrite
would grow in a direction perpendicular to the primary arms arms will develop laterally behind the primary dendrite tip
(see figure 10). Indeed, since the liquid concentration between (figure 10). One can easily observe that at 2tcr , 3tcr , etc
the primary arms remain close to C0 , the local undercooling from the moment corresponding to the destabilization of the
could finally trigger the growth of secondary dendrite arms initial spherical nucleus new secondary dendrite arms will start
perpendicular to the primary ones. It would be interesting to grow laterally. Hence a network of orthogonal secondary
to analyse from the perspective of the constructal law the dendrite arms develops between the primary dendrite arms.
conditions for the growth of lateral dendrites arms. In fact, The distance between two adjacent lateral arms will be no more
somewhat similar to the preceding case, one should analyse the than xcr . Hence xcr will also be a measure of the secondary
competition between the lateral diffusion controlled growth of arm spacing λ2 . One has thus
the primary arm, equation (36), and the dendritic growth of the
lateral secondary arms. Given a dendritic growth regime, the λ2 = xcr = 2Rt . (41)
solid region would spread laterally as
It is important to note the equivalence between the result
R = Vt (t − t0 ) , (38) in (41) and the result of the stability analysis made by
where Vt is the tip velocity given the local undercooling . L–M [1]. Moreover, the last result can be easily validated
Plotting now the above equation versus the diffusion controlled with experiments. In [2] and [16] for instance the authors
growth, equation (36), one can easily observe (figure 9) the measured the secondary arm spacing and their conclusion was
following: at the beginning of solidification, that is just unanimous: λ2 is almost equal to the double of the primary
behind the dendrite tip, the diffusion controlled growth is more arm tip radius
effective than the dendritic growth. However, soon after the λ2 ∼= 2Rt . (42)
start of the solidification (that is at a small distance behind The agreement between the present theory, the model
the dendrite tip) the dendritic growth regime becomes more in [1] and the experiments is remarkable. This good
effective than the diffusion controlled growth. The constructal agreement emphasizes once more the physical correctness of
principle calls for the choice of the optimal solid shape. Hence the constructal principle when applied to predict the solid
at tcr = tcr − t0 from the start of the solidification at x0 (that geometry of a dendritic structure.
is at a distance xcr = Vt (tcr − t0 ) behind the dendrite tip) a
dendrite arm will start to grow laterally (see figure 10(a)). The
critical time can be easily estimated as 4. The primary arm spacing
2Rt Many studies dealing with the modelling of the primary arm
tcr = tcr − t0 = , (39)
Vt spacing have been conducted in the past 20 years [6–12].

5259
A I Ciobanas et al

Second, in the upper cell/dendrite spacing range the

mechanism of dendrite division is identified. Indeed, for a large
average spacing between dendrites, the solutal interactions
between adjacent dendrites are weak and the liquid behind the
dendrite points will remain close to the initial concentration
C0 due to the thin solutal boundary layers around the columnar
tips (see figure 12). Therefore this solutally untouched space
leaves an undercooled liquid melt between the columnar
grains. Secondary or even tertiary arms developed behind
the columnar tips may evolve into vertical primary arms thus
reducing the primary arm spacing. In this respect one expects
to find at the intersection of these two unstable regimes a
stable dendrite array configuration defined by the primary arm
spacing λ1 (see figure 11). Lu [8] managed to quantify the
stable regime of the dendrite array by using their full numerical
model describing the growth of a dendrite tip. Despite the
accuracy of the Lu model the primary arm spacing can only be
obtained by solving the complex numerical dendrite model.
Therefore, an important limitation of the model is that no
general analytical expression for λ1 can be obtained.
In this context the existence of a simple and reliable
solution for the primary arm spacing would be of great
Figure 10. The evolution of the dendrite structure at four different interest. Let us first analyse the columnar dendrite array from
moments: (a) at 2tcr ; (b) at 3tcr ; (c) at 4tcr and (d) at 5tcr the perspective of the constructal theory. We have already
from the moment corresponding to the destabilization of the initial
spherical solid nuclei. The lateral arrows point to the positions on
identified in section 2 three constructal principles which define
the primary arm where new (lateral) secondary arms start to grow. an optimal solidification process. When applied to a columnar
front forced to advance into an undercooled melt with a
The majority of primary arm models [6, 7, 9, 10] manage to constant velocity (Vis ) and within a given constant gradient
obtain an analytical expression for λ1 by fixing in an ad hoc G, the three constructal principles predict that the only way
manner the solid (or the grain fraction [9, 10]) at the eutectic the columnar tips can optimize the solidification process is
temperature behind the columnar dendrite points. Even if the to minimize the spacing between two adjacent tips, namely
results obtained for λ1 reproduce well enough the experimental λ1 . Indeed, since the tip velocity is fixed at the level of
data the approach chosen in these models hides partially the the isotherm velocity, Vis , the only way the columnar tips
fact that λ1 is mainly controlled by the columnar tip kinetics, can optimize the solidification process is by decreasing the
hence by what happens at the tip and not at a certain distance dendrite array spacing, λ1 . By reducing λ1 , the time during
behind the tip. A completely different approach is used in which the solutal equilibrium is reached behind the columnar
[11, 12]. Here the primary arm spacing is regarded from the tips will also be reduced since the lateral diffusion process
perspective of the stability of a dendrite array advancing within around the dendrite tips is characterized by a diffusion time
an imposed thermal gradient and isotherm velocity. Besides scale depending directly on λ1 (τdw ∼ = λ21 /4/Dl ). At the
the complex mathematical derivation, these models fail to fit same time, a smaller λ1 will also mean that the solid is
accurately the experimental data. However the major trends more effectively invading the undercooled liquid. As already
in the variation of λ1 with respect to the thermal gradient pointed out, the mechanism responsible for the decrease in λ1
and isotherm velocity are well reproduced. Finally Lu [8] was identified as the ‘dendrite division mechanism’. However,
manages by using a complex numerical model of the dendrite the division of columnar grains cannot reduce λ1 indefinitely
tips to closely predict the variation of the primary arm spacing due to stability reasons. Indeed, one expects that due to strong
as well as the cellular spacing. An important feature of the solutal interactions between adjacent dendrite tips, there is
Lu [8] model is that the cell/primary arm spacing is obtained a critical λcr below which the division of the columnar tips
from a stability criterion. Indeed, the authors emphasized cannot take place. As we will see in the following, below λcr
the idea based upon which the observed cell/primary arm the mechanism of overgrowth becomes non-negligible and the
spacing defines a stable cell/dendrite configuration. In dendrite array becomes unstable to small perturbation of the
[8] two mechanisms determining the existence of a stable dendrite tips. Therefore, with respect to the constructal law the
configuration are identified. First in the lower cell/dendrite primary arm spacing characterizing the quasi-steady columnar
spacing range there is the mechanism of overgrowth which tips will be close to the minimum stable spacing, λcr .
limits from below the stable cell dendrite spacing. Indeed if In the following we will quantify analytically the two
the spacing between the dendrites array is low enough, due to mechanisms described above (the dendrite division and the
strong solutal interactions between adjacent dendrite tips, some dendrite overgrowth mechanism), to obtain a simple analytical
of the dendrites are overgrown by the surrounding members of expression for the primary arm spacing λ1 defining the stable
the array. Hence the natural tendency of a highly dense dendrite dendrite array configuration.
array is to increase its average dendrite spacing by means of Let us consider a columnar directional solidification
the overgrowth mechanism. experiment involving a constant vertical thermal gradient G

5260
 Dendritic solidification morphology

Figure 11. The two mechanisms responsible for the selection of the
primary arm spacing.

Figure 13. The perturbed columnar dendrite array.

where a = 0.4567 and b = 1.195. One can see now that Tts
Figure 12. A quasi-steady state columnar dendrite array. can be computed from (43).
To analyse whether or not the configuration in figure 12
and a constant isotherm velocity Vis (figure 12). The columnar is stable one should first consider the existence of a small
dendrite tips will be forced to advance within an undercooled perturbation at the dendrite tips. Let us consider that one
melt of initial composition C0 with a growth velocity equal dendrite is slightly perturbed from its initial quasi-steady
to the isotherm velocity. After an initial transition time the position defined by the isotherm Tts . In figure 13 it was
columnar dendrites, growing opposite to the heat flux direction, considered that the dendrite tip is perturbed such that the tip
will finally reach a quasi-steady state when their tip velocity found itself behind the quasi-steady position. As we will see
Vt will equal Vis . The quasi-steady state is defined with in the following, considering the reverse would not change
respect to the dendrite tips only. Since the dendrites always anything in the conclusions that we draw. It is now important
grow opposite to the isotherm velocity (see figure 12) the tip to notice that the natural tendency of the perturbed tip would
temperature (and consequently the local tip undercooling) will be to reach back its quasi-steady position. Indeed, since the
remain constant in time when the condition Vt = Vis is verified. new dendrite tip temperature Tt is lower than Tts , the local
Hence, if no perturbation of the tip is considered, the dendrite undercooling  will be higher than s . Hence one expects the
tip velocity will remain constant in time as well, defining thus perturbed tip to have a higher velocity than Vis . Since Vt > Vis
the quasi-steady state at the tip. Note that we are interested in one expects that in time the perturbed tip will get closer to the
the primary arm spacing defining a quasi-steady state and not a quasi-steady position defined by the isotherm Tts . The same
transition stage in the evolution of dendrite tips. Therefore we can be stated for the case where the perturbed tip found itself
will focus in the following on the stability of the dendrite array, ahead of the isotherm Tts . In this case the tip temperature
given the hypothesis of a quasi-steady state at the columnar will be greater than Tts and the tip velocity will be lower
tips. than the isotherm velocity. Consequently the tip temperature
Let us consider a directional solidification experiment will decrease in time, the perturbed tip getting closer and
(fixed G and Vis ) in which the columnar dendrite array is closer to the quasi-steady state position. In this context one
advancing with the velocity Vt = Vis , that is in a quasi- important aspect of the dendrite tip behaviour is revealed:
steady state. Let us denote with λ the average spacing between subject to perturbations from their steady state position the
two adjacent dendrites of the dendrite array (figure 12). Let dendrite tips will tend to reach back their initial quasi-steady
us denote as well with Tts the columnar tip temperature position. However this process will not be instantaneous but
corresponding to this quasi-stationary state. This temperature will follow a certain dynamics. We propose in the following to
can be easily computed since the columnar tip velocity determine the characteristic time within which the perturbed
Vt (= Vis ) can be directly linked to the local tip undercooling dendrite tips reach back the quasi-steady position. Let us
C ∗ −C0
s = C ∗l(1−k) by means of a tip kinetics law. Indeed, by denote this time with trs (see figure 13). Let us consider too that
l
assuming a parabolic tip shape one can write the perturbed dendrite tip is deviated at a distance x0 from its
initial quasi-steady position (figure 13). To determine trs one
Dl m (k − 1) Cl∗ 2
Vt = I v −1 (s ) , (43) should therefore resolve the kinetic equation of the dendrite tip
π 2
where Cl∗ = (Tts − Tf )/m is the equilibrium concentration dx
= Vt , (45)
corresponding to the tip temperature Tts and I v −1 represents dt
the inverse Ivantsov function. Since this function has no
where x is the position of the dendrite tip. One question
simple analytical solution, we will use in the following the
remains, however: what is the relevant tip velocity Vt ? We can
fit computed by Wang and Beckermann [23]:
assume that the tip velocity can be linked at any moment to
I v −1 () = a [/ (1 − )]b , (44) the local tip undercooling using a standard tip kinetic theory,

5261
A I Ciobanas et al

equation (43). The latter hypothesis is not straightforward The differential equation to be solved becomes
since equation (43) is valid for a steady state dendrite tip. It is
  
worthwhile to note that experimental evidence exists to support dx ∂Vt Tts − T0
= Vis + x− − Vis t , (49)
the steady state approximation to the tip velocity Vt . Indeed, dt ∂x x=xst G
Liu [16] showed that the measured instantaneous tip velocity
of a decelerating columnar dendrite array (linear decrease where for t = 0, x(0) = xst − x0 . The ordinary differential
of the isotherm velocity) could be almost exactly correlated equation can be solved, and one obtains
with the steady state velocity computed with the help of local 
 
undercooling conditions at the tip. ∂Vt
x − xst = −x0 exp t . (50)
In fact, the use of a stationary solution for the tip velocity ∂x x=xst
Vt in a case where the local tip undercooling is unstationary
were to be valid only if the velocity response time of the tip to Notice that (∂Vt /∂x)x=xst < 0. Hence, it is easy to observe that
changes in local conditions at the tip were to be much smaller the initial tip deviation (x0 ) from its quasi-steady position
than the time scale characterizing the same changes in the is constantly reducing in time. Looking at the particular
local conditions. For rapid solidification cases involving fast exponential variation in (50) one can identify a characteristic
advancing columnar dendrites, the dendrite tip Rt is very small time scale, −(∂Vt /∂x)−1 xst , within which the perturbed tip
(∼
=10−6 m). Since the solutal boundary layer at the tip is of the almost reaches back the quasi-steady position xst . It is
order of Rt the time scale characterizing the solute transport at important to note that this time scale is independent of the initial
the tip would be of the order of τt ∼ = Rt 2 /4Dl , thus small too perturbation x0 . This is a very important remark because
∼ −9 2 −1
(Dl = 10 m s ). Because the tip velocity is intrinsically the determined time scale will characterize in a unique way
linked to the solute transport at the tip the velocity response the dynamics of the perturbed tip. In fact −(∂Vt /∂x)−1 xst is
time of the tip to changes in local undercooling conditions will only dependent on parameters external to the tip perturbation
be of the order of τt as well. Now, the time scale characterizing such as the thermal gradient, isotherm velocity, initial alloy
the variations in local undercooling conditions will be no more composition and tip kinetics law. Since the initial deviation is
than trs , which is the tip return time to the quasi-steady position. almost entirely reduced during a time equal to the double of
Hence, the tip velocity Vt can be approximated with the steady −(∂Vt /∂x)−1 xst one can state that the return time to the quasi-
state solution only if τt  trs . For the moment we consider this stationary state, trs , is of the order of
as a valid hypothesis. The inequality τt  trs will be verified

a posteriori once the time scale trs is determined. ∂Vt −1
trs ∼
= −2 . (51)
As pointed out, to compute the return time trs one should ∂x xst
resolve equation (45) where Vt is expressed with the help of
(43). However, the use of (43) for the tip velocity would Hence a unique time scale characterizing the dynamics of the
render very difficult, if not impossible, the analytical solving dendrite array has been determined. It is somewhat evident
of the differential equation (45). In this respect one should that this time would be a key parameter in determining the
look for a simple approximation of the kinetic law. Remember stable regime of the dendrite array. In addition to quantifying
that we need to analyse the dynamics of the tip subject to the return time to the stationary state of a perturbed tip, trs
small perturbation from the steady state. Hence equation (45) will also quantify the time during which a dendrite tip issued
must be solved for very small deviations x0 from the steady from a dendrite division mechanism (like the tertiary arm
position. In this context a reasonable approximation for the evolving into a primary arm) will reach back the isotherm Tts
Vt variation about the quasi-steady position would be its linear corresponding to the quasi-stationary state. Hence, trs will
approximation (the first order Taylor expansion): also be a measure of the dendrite division mechanism. In

 order to determine whether or not the division mechanism can
∂Vt
Vt ∼ = Vis + (x − xst ), (46) take place one should try to quantify as well the mechanism
∂x x=xst
opposing the natural tendency of dendrite tips to reach back
where xst is the position of the isotherm Tts corresponding the steady state position. This mechanism can be found in
to the quasi-steady state of the tip. Notice that xst is not a the existing solutal interactions between two adjacent tips. A
static coordinate since the isotherm Tts is moving with the dendrite tip that falls behind the quasi-steady position will be
imposed isotherm velocity Vis . Indeed the temperature field solutally influenced by the adjacent tips. A question remains
in the vertical ingot is to be answered however: how can one quantify these solutal
T (x) = T0 + Gx − GVis t, (47) interactions? To answer this question one should take into
account the existence of the solutal boundary layers around
where T0 is the temperature at the bottom of the ingot at the tips. An advancing dendrite tip will always leave behind
t = 0. Since considering an initial overheating at the bottom it a solutal diffusion wave which spreads laterally at the speed
of the ingot would not change anything in the problem we wdw (x) ∼ 2D 1/2 (t ∗ )−1/2 , where t ∗ = t − tp is the time
will hereafter consider that T0 is the liquidus temperature computed from the time tp at which the dendrite tip reached
corresponding to the initial concentration C0 , that is T0 = a certain coordinate x. Looking in the frame of reference
Tf +mC0 . In this respect one can identify the quasi-steady state linked to the isotherm Tts one can see the perturbed tip trying
position of the tip, xst , knowing the corresponding temperature to reach back the steady state position but always positioned
at the tip Tts . Indeed, from (47) one obtains behind the two adjacent tips (figure 13). Since the two adjacent
Tts − T0 tips are located ahead of the perturbed one they can solutally
xst = + Vis t. (48)
G influence it by means of lateral diffusion waves. The travel

5262
 Dendritic solidification morphology

time of a diffusion wave from one tip to another can be easily

estimated [4] as
2
∼ λ .
tdw = (52)
4Dl
In conclusion, with the help of tdw we have managed to quantify
the strength of the solutal interaction between adjacent tips.
In this context it would be interesting to compare the two
mechanisms: the dendrite division mechanism (trs ) and the
solutal interactions between adjacent tips (tdw ). If trs  tdw ,
then the solutal diffusion wave has the time to capture the
perturbed tip before the latter reaches back the quasi-steady
position. Therefore it is likely that the adjacent tips solutally
overgrow the perturbed one. Hence, the dendrite array is
not stable to small perturbation of the dendrite tips. On the
contrary if trs  tdw then the perturbed tip has the time to
reach back the quasi-steady position before any lateral solutal Figure 14. Comparison of the present theory and the Hunt [6] and
wave captures it. Hence, the dendrite array is stable to small Kurtz [7] models with the experimental results in [2]:
perturbations of the dendrite tips. However, one expects for succinonitrile–5.5 mol% acetone system, G = 67 K cm−1 .
this configuration the dendrite division mechanism to be active.
Indeed, perturbed tertiary arms can easily transform into a At this moment it is interesting to note that the analysis
primary arm since they have the ability (the time) to reach of the dendrite array from the perspective of constructal
back the isotherm Tts (the steady state position) before any theory helped us to conclude that the primary arm spacing
lateral diffusion wave stops them. This configuration will not characterizing a columnar dendrite array should be close to the
be stable with respect to the dendrite division mechanism. critical spacing, λcr . And we have indeed identified the real
In the lower range of the dendrite array spacing (λ  primary arm spacing as being of the order of 1.5λcr , equation
(4Dl trs )1/2 ) the dendrite array is unstable to small perturbations (54). Basically the dendrite array is always searching for an
of the dendrite tips whereas in the upper dendrite spacing
optimum configuration given some external constraints (G,
range (λ  (4Dl trs )1/2 ) the dendrite array is stable to small
Vis ). The constructal theory helps us to define these optimal
perturbation of the dendrite tips but unstable with respect to
conditions (in our case a minimum dendrite spacing) and to
the dendrite division mechanism (figure 11). One should look
identify the mechanism pointing to the optimal configuration
therefore for the stable dendrite spacing somewhere at the
(in our case the dendrite division and the dendrite overgrown
intersection of these two domains, that is somewhere around
mechanisms).
the critical spacing defined by trs = tdw :
At this moment it would be interesting to compare the
λcr = (4Dl trs )1/2 . (53) main result in equation (54) with different experimental data.
First the present model is compared with the experimental
However, since λcr represents the upper limit of the unstable data provided by Somboonsuk et al [2]. The directional
regime to small perturbation of the dendrite tips one expects experiments in [2] have been carried out on a succinonitrile–
the real primary arm spacing to be somewhat larger than λcr . 5.5 mol% acetone system at a constant thermal gradient of
Similarly to the Lu [8] hypothesis, we suppose that the dendrite G = 67 K cm−1 . The properties used in the present calculation
division mechanism becomes active for a local spacing greater are summarized in table 1. In figure 14 the comparison of the
than two times the critical spacing λcr . In this respect one experimental data with the present model as well as with the
expects the real primary spacing to be somewhat between λcr well-known Hunt [6] and Kurtz [7] models is presented. As
and 2λcr . Hence, we will consider that the primary arm spacing one can see, despite the relative simplicity of the model, the
will be of the order of agreement of the present model with the experimental data is
very good. Equation (54) does much better than the Hunt [6]

 
∂Vt −1 and Kurtz [7] models. Moreover, it better predicts the slope
λ1 ∼
= 1.5λcr ∼
= 18Dl − . (54) of the real primary arm variation with respect to the pulling
∂x xst
velocity.
Next a comparison with the experimental data of
We have thus obtained a simple analytical expression for the
McCartney and Hunt [13] obtained for an Al–0.34 wt % Si–
primary arm spacing. At this moment it would be interesting to
0.14 wt% Mg system is presented in figure 15. We note again
validate the hypothesis we made regarding the tip response time
a very satisfactory agreement of the present theory with the
to changes in local conditions, τt  trs . Since it was proven
experimental data. Notice that in table 1 we have approximated
that τt ∼
= (Rt )2 /4/Dl and trs ∼
= (λ1 /1.5)2 /4/Dl it is now easy
the real ternary alloy with an effective binary Al–0.34 wt % Si
to see that the τtip  trs inequality would be equivalent to
alloy.
Rt  λ1 /1.5. (55) In figure 16 a comparison with the experimental data
provided by Young and Kirkwood [14] on three different Al–
The above equality is obviously true since it was proven [7] Cu alloys is presented. One can observe the good agreement
that Rt  λ1 . of the present theory with the experimental data. Notice again

5263
A I Ciobanas et al

Table 1. The physical properties used for the comparison of the present theory with experiments.
Alloy system Properties and units
C0 (wt %) Tf (K) m (K wt%−1 ) k Dl (m2 s−1 )  (mK)
−9
Scn – 0.35 wt % acetone 0.35 331.08 −2.8 0.1 1.3 × 10 0.64 × 10−7
Scn–5.5 mol% acetone 5.5 (mol %) 331.08 −2.2 (K mol%−1 ) 0.1 1.27 × 10−9 0.64 × 10−7
Al–0.34 wt % Si–0.14 wt % Mg 0.34 933.4 −5.5 0.18 3.5 × 10−9 2.4 × 10−7
Al–2.4 (4.4; 10.1) wt % Cu 2.4 (4.4; 10.1) 934.2 −2.6 0.14 3 × 10−9 1 × 10−7
Pivalic acid–0.076 wt % ethanol 0.076 308.9 −2.94 0.16 0.6 × 10−9 4.34 × 10−8

Figure 15. Comparison of the present theory and the Hunt [6] and Figure 17. Comparison of the present theory and the Hunt [6] and
Kurtz [7] models with the experimental results in [13]:Al–0.34 wt % Kurtz [7] models with the experimental results in [15]:
Si–0.14 wt% Mg system, G = 30 K cm−1 . succinonitrile–0.35 wt% acetone system, G = 37.6 K cm−1 .

cellular and dendritic structure coexist within a finite band of

pulling velocities. In fact, when the experiment was repeated
several times, the authors observed cells in some experiments
and dendrite in others. In this respect it would be interesting
to modify the present model in order to compare it with the
cellular spacing within this finite band of velocities. Until
now, in order to obtain relevant data for the full dendritic
regime we have used for the computation of the tip velocity
gradient, −(∂V /∂x)−1 xst , the Ivantsov solution for the dendrite
tip kinetics. However, since a parabolic shape approximation
to the dendrite shape can no more be used for the cellular arrays,
one should look for a more appropriate approximation. Since
the tip of the cells obtained within this finite velocity band is
close to a spherical shape [15] an interesting approximation for
the dendrite tip shape would be the spherical approximation.
Figure 16. Comparison of the present theory and the Kurtz [7] In this respect, the tip Peclet number, P e, becomes equal to
model with the experiments in [14]: full lines, dashed lines and the dimensionless tip undercooling, . Hence, the tip velocity
dashed-dot lines are theory and squares, circles and triangles are would become
experiment for Al–2.4 (G = 27), Al–4.4 (G = 26) and Al–10.1
wt% Cu (G = 53 K cm−1 ), respectively. Dl m(k − 1) Cl∗ 2
Vt =  . (56)
π 2
that equation (54) better predicts the trend in the real primary By using the above law for the computation of the velocity
arm spacing variation with respect to the pulling velocity when gradient, −(∂V /∂x)−1 xst and time scale trs one can approach now
compared with the Kurtz [7] model. the cellular spacing. The comparison between the quadratic
Next a comparison with the Eshelman [15] experiments approach to λ1 and the experimental data is presented in
made on succinonitrile–0.35 wt% acetone system at G = figure 17. One can notice the excellent agreement with the
37.6 K cm−1 is assessed in figure 17. As one can see, the real cellular spacing. The agreement of the present model with
agreement with the model in the velocity range corresponding both the dendrite and cellular spacing emphasizes once more
to the dendrite structures is excellent. In fact, in [15] the the fact that the primary arm spacing λ1 is mainly controlled
authors provided spacing data for the dendrite domain as well by the tip kinetics, hence by what happens at the dendrite tips
as for the cell domain. It is interesting to observe that both and not at a certain distance behind the tips.

5264
 Dendritic solidification morphology

place. Moreover, these two critical time and length scales

are computed by means of a more rigorous stability analysis.
We conclude thus that the results obtained by applying the
constructal principle are of the same order of magnitude as
the ones obtained by means of the more complex stability
analysis. Hence, the idea of a competition between the
diffusion controlled growth and the dendritic growth and
consequently between two solid geometries (the sphere and
the dendrite tip) has a secure physical basis. The constructal
principles do not appear as ad hoc established statements but
as a physical reality having a rigorous explanation.
Second, we developed a model for the secondary arm
spacing having as the basis the same concept of the constructal
theory. We analysed first the growth of a single dendrite tip
and concluded that the lateral growth of a needle is close to
Figure 18. Comparison of the present theory and the Hunt [6] and a diffusion controlled growth, therefore being characterized
Kurtz [7] models with the experimental results in [15]: pivalic by a continuously decreasing growth rate. In this respect
acid–0.076 wt% ethanol system, G = 29.8 K cm−1 . the constructal theory calls for an optimization of the lateral
growth. This can be achieved if secondary dendrite arms
Finally a comparison with the experiments in [15] on start to grow laterally. In fact, we identify once more a
pivalic acid–0.076 wt% ethanol is presented in figure 18. Here competition between the lateral diffusion controlled growth
again the agreement is very satisfactory. and the dendritic growth. By analysing this competition we
We finally conclude that, in spite of its relative simplicity, are finally able to compute the secondary arm spacing. It
the present theory correctly predicts the primary arm spacing. appears that the secondary arm spacing is equal to the double
Even more, by means of the two time scales trs and tdw , we of the primary arm dendrite tip radius. This result is in good
have managed to quantify in a simple way the two mechanisms agreement with the model of L–M [1] as well as with different
responsible for the selection of the primary arm spacing: experimental results [2, 16].
the dendrite division mechanism and the solutal interactions Finally, the primary arm spacing is analysed from the
between two adjacent tips, the latter also being responsible for perspective of the constructal law. When applied to a columnar
the dendrite overgrown phenomena. front forced to advance into an undercooled melt with a
constant velocity (Vis ) and within a given constant gradient
G, the constructal law predicts that the only way that the
5. Conclusions columnar tips can optimize the solidification process is by
minimizing the spacing between two adjacent tips, namely
In this paper we have focused on the application of the λ1 . By minimizing the primary arm spacing the columnar
constructal theory to predicting the dendritic solid structure. structure is also optimizing the solidification process behind
First we have analysed the marginal stability criterion from the columnar tips. The mechanism responsible for the decrease
the perspective of the constructal principle. Having as guiding in λ1 was identified as the dendrite division mechanism.
principle the constructal law we have concluded that the However, the division of columnar grains cannot reduce λ1
dendrite shape will adapt in such a way as to grow with a indefinitely due to stability reasons. Due to strong solutal
maximum velocity. In this context we have shown that among interactions between adjacent dendrite tips, there is a critical
the whole range of possible dendrite tip radii predicted by the λcr below which the division of the columnar tips cannot
stability analysis the dendrite tip will choose the smallest one, take place. We have shown that below λcr the mechanism of
that is a radius equal to the smallest perturbation wavelength dendrite overgrowth becomes non-negligible and the dendrite
leading to instabilities. This is somewhat similar to the array becomes unstable to small perturbation of the dendrite
phenomenon of laminar–turbulent transition where the choice tips from their quasi-steady position. Therefore, with respect
between the viscous diffusion and the ‘needles’ of eddy flow is to the constructal law the primary arm spacing characterizing
also made at the smallest (neutral) wavelength of instability. In the quasi-steady columnar tips will be close to the minimum
this context it appears that the turbulent structure and dendrites stable spacing, λcr . By quantifying the two mechanisms, the
in solidification are analogous manifestations of the constructal dendrite division and the dendrite overgrowth mechanisms,
phenomenon of generation of flow configuration. we were finally able to obtain a model for the primary arm
We have also shown that the initial diffusion controlled spacing. We found that λ1 is of the same order of magnitude
growth regime of the solid grain is characterized by a as λcr , and more precisely that λ1 ∼ = 1.5λcr . This result is
continuously decreasing growth rate. By applying again the validated subsequently against various experimental data. We
constructal principle we obtain that solid is forced to select showed that this new theory is in very good agreement with
a different more effective geometry: the dendrite tip. In this the experiments.
respect we identify the existence of a competition between We conclude that the constructal theory can be
the diffusion controlled growth and the dendritic growth. By successfully used in predicting the dendritic solid structure.
analysing this competition we were able to determine the A complete review of the progress on the constructal theory
critical moment and the critical radius of the solid grain at of generation of configuration in nature can also be found
which the destabilization of the solid–liquid interface takes in [24, 25].

5265
A I Ciobanas et al

References [14] Young K P and Kirkwood D H 1975 The dendrite arm

spacings of aluminium–copper alloys solidified under
[1] Langer J S and Muler-Krumbhaar H 1978 Theory of dendritic steady state conditions Metall. Trans.
growth:I. Elements of a stability analysis Acta Metall. A 6 197–205
26 1681–7 [15] Eshelman M A, Seetharman V and Trivedi R 1988 Cellular
[2] Somboonsuk K, Mason J T and Trivedi R 1984 Interdendritic spacing: I. steady state growth Acta Metall.
Spacing: I. Experimental Studies Metall. Trans. A 15 36 1165–74
967–75 [16] Liu S, Lu S-Z and Hellawell A 2002 Dendritic array growth in
[3] Bejan A 1997 Advanced Engineering Thermodynamics 2nd the systems NH4 Cl–H2 O and [CH2 CN]2–H2 O: the
edn (New York: Wiley) pp 794–804 detachement of dendrite side arms induced by deceleration
[4] Bejan A 2000 Shape and Structure, From Engineering to J. Cryst. Growth 234 740–50
Nature (Cambridge: Cambridge University Press) pp 175–8 [17] Bianchi A M, Fautrelle Y and Etay J 2004 Transferts
[5] Bejan A, Fautrelle Y and da Silva A K 2002 Dendritic Thermiques (Lausanne: Presses Polytechniques et
solidification structures deduced from constructal theory Universitaires Romandes) pp 339–44
(unpublished work) [18] Mullins W W and Sekerka R F 1964 Stability of a planar
[6] Hunt J D 1979 Cellular and primary dendrite spacings interface during solidification of a dilute binary alloy J.
Solidification and Casting of Metals (London: Metals Appl. Phys. 35 444–51
Society) pp 3–9 [19] Langer J S and Muler–Krumbhaar H 1977 Stability effects in
[7] Kurtz W and Fisher D J 1981 Dendrite growth at the limit of dendritic crystal growth J. Cryst. Growth 42 11–14
stability: tip radius and spacing Acta Metall. 29 11–20 [20] Bejan A 1984 Convection Heat Transfer (New York: Wiley)
[8] Lu S-Z and Hunt J D 1992 A numerical analysis of dendritic pp 215–19
and cellular array growth: the spacing adjustement [21] Boettinger W J, Warren J A, Beckermann C and Karma A
mechanism J. Cryst. Growth 123 17–34 2002 Phase-field simulation of solidification Annu. Rev.
[9] Ma D and Sahm P R 1998 Primary spacing in directional Mater. Res. 32 163–94
solidification Metall. Mater. Trans. A 29 1113–9 [22] Ivantsov G P 1947 Temperature field around spherical,
[10] Ma D 2002 Modeling of primary spacing in dendrite arrays cylindrical and needle-shaped crystals which grow in
during directional solidification Metall. Mater. Trans. B 33 supercooled melt Dokl. Akad. Nauk SSSR
223–33 58 567–9
[11] Warren J A and Langer J S 1990 Stability of dendritic arrays [23] Wang C Y and Beckermann C 1996 Equiaxed dendritic
Phys. Rev. A 42 3518–25 solidification with convection: I Multiscale/multiphase
[12] Warren J A and Langer J S 1993 Prediction of dendritic Modelling Metall. Mater. Trans. A 27 2754–64
spacing in a directional-solidification experiment Phys. Rev. [24] Bejan A 2006 Advanced Engineering Thermodynamics 3rd
E 47 2702–12 edn (Hoboken, NJ: Wiley) chapter 13
[13] McCartney D G and Hunt J D 1981 Measurements of cell and [25] Bejan A and Lorente S 2006 Constructal theory of generation
primary dendrite arm spacings in directionally solidified of configuration in nature and engineering J. Appl. Phys.
aluminium alloys Acta Metall. 29 1851–63 100 041301

5266
Référence II
Submitted to Journal of Physics D: Applied Physics

Ensemble averaged multi-phase eulerian model for

columnar/equiaxed solidification of a binary alloy
Part I. The mathematical model
A I Ciobanas1 and Y Fautrelle1
1
EPM/CNRS Laboratory, ENSHMG, BP 95, 38402 Saint Martin d’Hères cedex, France

E-mail: aciobanas@yahoo.com

Abstract
A new multi-phase Eulerian model for the columnar and equiaxed dendritic solidification has been
developed. The mean conservation equations are derived by means of a statistical phase averaging
technique, and the mathematical formulation of the model can be used for both columnar and
equiaxed solidification. The model uses three different phases respectively the columnar and
equiaxed solid and the liquid. The new set of equations enables us to simulate the CET during the
directional solidification of a binary alloy. Owing to the statistical nature of the model, we are able
to treat rigorously the coexistence of equiaxed and columnar structures and consequently the CET
phenomena. The averaged equations are closed by means of the cell model approximation. This
technique can be successfully used to model the various interactions between the liquid and the
solid. It may also incorporate the effects of the inhomogeneities of the various scalar fields, e.g.,
the solute and temperature gradients. An envelope model is used to parameterize the small scales,
i.e., the dendrite scale. This leads us to distinguish two types of liquid, namely the extra-dendritic
and inter-dendritic liquid. In part I the equations are rigorously derived in the purely diffusive case,
whilst in part II we present one-dimensional simulations of Sn-Pb and Al-Cu directional
solidification experiments involving CET phenomena. Quasi-steady state CET maps are as well
computed.

PACS: 81.30.Fb Solidification, 82.20.Wt Computational modeling; simulation

1. Introduction
Modeling of the solidification at the scale of the process remains a challenging issue. The reason is
obviously linked to the complexity of the phenomena occurring in real practical situations. The first
difficulty rests in the existence of many length scales, from the dimension of the ingot (a few
10cm) to the chemical diffusion microscale (a few microns). Another major problem lies in the
nature of the solidification process. The growth of the solid may be columnar or equiaxed, and a
competition between the two types often occurs during the solidification process: the so-called
columnar-to-equiaxed transition (CET).
Concerning the prediction of solidification, numerical models have been developed in the
past twenty years [1-7], in which equations for heat transfer, fluid flow, solidification and solute
transfer are solved in a coupled way. The objective of such models is to simulate how operation
parameters (alloy composition, thermal and hydrodynamical conditions) can influence some
characteristics of the defects (composition difference, position in the product). So far most of the
models are based on continuum models [1] or space-averaged equations [2-7]. In the latter case, the
local equations are integrated in an elementary representative volume (REV) whose size is greater
than the dendrite scale [2]. Such an averaging procedure has proved its efficiency. However, the
method itself may be questionable. Firstly, the spatial averaging procedure is based on a scale
separation between the micro and macro-scales. This is not true in practice, and the existence of
mesoscale phenomena; e.g., the freckles, contradicts the separation hypothesis. Secondly, the
extension of the models based on spatial averaging method to situations where the flow is
turbulent, especially when an electromagnetic stirring is applied, is generally rigorously impossible.
The statistical approach avoids those objections. Furthermore, it is consistent with the random
nature of the germination process as well as the behavior of the equiaxed grains.

II-1
Submitted to Journal of Physics D: Applied Physics

The aim of the present work is to transpose the statistical phase averaging procedure, used
in two-phase flows [8, 9], to the modeling of the solidification. Such an approach has been
suggested first by Furmanski [10]. A first preliminary set of equations has been proposed by
Fautrelle et al. [11] and Ciobanas et al. [12] using the statistical phase averaging approach. In the
present paper, we extend the idea proposed by Fautrelle et al. [11] to propose a complete set of
equations which describe both the columnar and the equiaxed solidification. Moreover, combined
with the so-called cell model [9] the averaging procedure can take into account the effects of
inhomogeneous nature of the solute and temperature fields.

2. Ensemble averaging
To approach numerically a solidification process one has to solve a set of balance equations,
characterizing the balance of mass, energy, momentum and solute in the ingot. At the smallest scale
one can easily write down the local form of these equations but the resolution of the system at the
smallest length scale present in the ingot will be a virtual impossible task for real solidification
problems usually involving ingots larger than 10−1 m. On the other hand one is not always
interested on variations of parameters at the smallest scale but instead on average parameters like
velocity, solute concentration characterizing large vortices or mesosegregation. Moreover the latter
phenomena are also more easily captured experimentally. In this respect we are interested in a set
of average equations which could reproduce the evolutions in time of different averaged
parameters, like velocity, temperature, solute concentration, etc. These average equations can be
obtained by applying an average operator to each of the local balance equation (Kataoka [13]).

2.1. Ensemble operator

To be able to apply a statistical averaging operator to a physical process, the latter must be a
repeatable one. To be more precise the process must not be very sensitive to small changes of
initial and boundary conditions. If it is not, from one experiment to another, one will obtain
incompatible results which cannot be validated by different experimentalists. The first hypotheses
of our analytical study will refer thus to property of repeatability of the process.
Under this condition we can define the ensemble average [9] as the average which allows
the interpretation of physical phenomena in terms of the repeatability of the multiphase
phenomenon. However, it is worthwhile to notice that since the initial and boundary conditions will
be slightly different from one experience to another, each new experiment (realization) of a
multiphase process will be unique. Following Drew [9] by assuming the flow as deterministic, the
randomness of the multiphase flow arises from the uncertainty in the initial conditions. Thus, each
experiment will give birth to a new element of the ensemble that contains all the possible
realizations of the experiment. In the case of multiphase flows this ensemble contains a virtual
infinite number of realizations. This ensemble is of great importance since the averaging operator
will act on it. Note that, the physical parameter to be averaged is completely at the choice of the
researcher: it can be the pressure, velocity, temperature or solute concentration field or the
positions of the interface between the solid and the liquid.
Thus, if we denote with µ the realization and with ψ ( t , x ) the physical field we are
interested in, then ψ ( t , x; µ ) will denote the dependence of the field ψ with respect to the
realization ensemble µ . Let now θ be the ensemble made of all possible realizations of the
multiphase system. Note that each subset of θ is an ensemble too Drew [9]. However this
ensemble has to posseses a particular feature in order to be useful in the statistical description of
the system. One should be able to assign probabilities for any subset of θ , or in mathematical
language, that θ be a Borel set. Given this proprety, one could define a measure (the probability)
on the set of all possible realizations θ . In this context the ensemble average of ψ can be defined
as follows [9]:

ψ ( t , x ) = ∫ψ ( t , x; µ ) dm ( µ ) (1)
θ

II-2
Submitted to Journal of Physics D: Applied Physics

with dm ( µ ) the probability density on the Borel set θ .

2.2. Ensemble averaged equations

The first step in obtaining of the macroscopic equations is the writing of the local conservation
equations of a generic field ψ ( x,t ) [9, 13]:

∂ ( ρψ )
∂t
( )
+ div ( ρ vψ ) = div jψ + ρ bψ , (2)

along with the jump conditions (the conservation of ψ ) at the solid-liquid interface:

⎡ ρψ ( v − v i ) + jψ ⎤ ⋅ n = mψ , (3)
⎣ ⎦

where ρ represents the density, v the velocity, vi the solid-liquid interface velocity, jψ the
molecular diffusion flux of ψ , bψ the body source of ψ , n the unit normal at the solid-liquid
interface and mψ is the interfacial source of ψ . In (2) ψ is a generic physical field and can be
replaced with 1, the velocity, the specific enthalpy or the solute concentration, obtaining
respectively the local mass, momentum, energy and solute conservation equations. These local
balance equations (2) and (3) are valid in each point of the two-phase system, but, as already
pointed out, their numerical resolution at the smallest length scale would be a virtual impossible
task. Moreover, these equations states for the conservation of ψ inside the whole multiphase
system. As one would desire to track the average evolution for each phase separately, the effect of
each phase on the total balance of ψ has to be identified first. This is done by multiplying the local
conservation equations with the characteristic phase function:

⎧⎪1 if ( t , x ) is within the k phase

X k (t, x ) = ⎨ (4)
⎪⎩0 if ( t , x ) is not within the k phase
Then, in order to have an insight on average quantities, one has to further apply the ensemble
average operator defined in (1). Using then the topological relations characterizing X k [9], one can
easily obtain the average conservation equations for ψ in each phase k as well as the
corresponding jump equations:

(
∂ X k ρψ ) + div
∂t
( X k ρ vψ ) = div ( X k jψ ) + X k ρbψ + ρψ ( v − vi ) ⋅ ∇X k − jψ ⋅ ∇X k (5)

∑ ⎡⎣⎢ ρψ ( v − vi ) ⋅ ∇X k − jψ ⋅ ∇X k ⎤⎦⎥ = m (6)

k

Equation (5) states for the average balance of ψ inside each phase k in the multiphase
system and equation (6) for the conservation of ψ across the phase interfaces. The term ∇X k has
to be understood in the sense of distributions (see Drew [8]) and therefore the two terms in (5) and
(6) containing ∇X k are intrinsically linked to the interface and have a meaning only on the solid-
liquid interface. Indeed, ∇X k can be also written as [9]:

∇ X k = −n k δ ( x − x i ( t ) ) (7)

II-3
Submitted to Journal of Physics D: Applied Physics

where n k is the unit normal to the interface of phase k , pointing out of the phase interface and
δ ( x − xi ( t ) ) represents the Dirac delta function at the interface.
To be useful in a numerical model, the formulation of (5) and (6) has to be structured by
defining adequate average variables. These average variables can be of different types following
the choice of the weighted function: average variables weighted with the characteristic function
X k , mass weighted averages (Favré weighted) with X k ρ and finally, interfacial averages
weighted with ∇X k , having a meaning at the interface only. Definition of all these variables is
synthesized in Table 1.

Table 1: Average variable definition

Definition Variable description
εk = Xk The "volume" fraction of phase k
S k = −n k ⋅ ∇ X k The interfacial area density of phase k
Xk ρ
ρk = The average density within phase k
εk
X k ρψ
ψk = The mass-weighted average of ψ within phase k
εk ρ k
X k jψ The average molecular flux of ψ within phase k
jψ k =
εk
X k v′kψ k′
j′ψ k = − The fluctuation (Reynolds) flux within phase k
εk
X k ρ bψ
bψ k = The mass-weighted average body souces of ψ within phase k
εk ρ k
Φψ k = − jψ ⋅ ∇X k The average molecular flux of ψ at the interface of phase k
The average source of ψ at the interface of phase k, due to the
Γψ k = ρψ ( v − vi ) ⋅ ∇X k
exchange between phases

With the help of definitions in Table 1 one is able now to write down the ensemble
averaged conservation equation for the general physical field ψ and the corresponding jump
conditions:

(
∂ ε k ρ kψ k ) + div ε ρ vkψ = div ⎡ε jψ + j′ψ ⎤ + ε ρ bψ k
∂t
( k k k ) ⎢⎣ k ( k k )⎥⎦ k k (8)
+ Φψ k + Γψ k ; k = {s, l}

∑
k
(Φ
ψ
k )
+ Γψ k = m ; k = {s, l} (9)

The final system of averaged equations describing the solidification process will consist of mass,
momentum, energy and solute balance written for each phase in the multiphase system. This is
done by particularizing the general field ψ , the corresponding molecular flux jψ and the
volumetric source bψ with the appropriated expressions for each balance equation. Table 2
synthesizes all of this parameters. Note that the molecular fluxes for the to momentum, energy and
solute balance correspond respectively to the classical stress tensor of a newtonian fluid, the

II-4
Submitted to Journal of Physics D: Applied Physics

Fourier heat flux, and the Fick solute flux as follows:

T = − pI + µ ⎡∇v + ( ∇v ) ⎤ + λv ∇ ⋅ vI
T
⎣ ⎦
q = − λ ∇T (10)
jc = − ρ D ∇C

where p is the local pressure, µ is the dynamical fluid viscosity, λv is the bulk viscosity, λ
represents the thermal conductivity and D is the solute diffusivity.

Table 2: Particularization of the general averaged equations

Balance equations: ψ Molecular flux jψ Body sources bψ

Mass 1 0 0
Momentum velocity stress tensor gravity or magnetic force
v ⎡ m.s −1 ⎤ T ⎡ kg .m −1.s −2 ⎤ g or Fm / ρ ⎡ m.s −2 ⎤
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
Energy enthalpie molecular heat flux: body heat source
h ⎡ J .kg −1 ⎤ −q ⎡ J .s −1.m −2 ⎤ Q ⎡ J .s −1.m −3 ⎤
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
Solute solute conc. molecular mass flux: (no chemical reactions)
C [ wt %] − jc ⎡ kg .s −1.m −2 ⎤
⎣ ⎦ 0

3. The pure columnar or equiaxed solidification

The concept of ensemble averaging as defined in equation (1) is rather abstract. Consequently the
unknown terms arising from the averaging procedure will be difficult to model. In order to better
understand the nature of this average technique one should approach the ensemble average through
the distributions functions attached to the macroscopic process to be studied.

Figure 1: A typical directional solidification distinguishing the full solid, the columnar, the
equiaxed and the bulk liquid zone.
Let us consider a solidification process taking place in an ingot filled initially with a liquid
metallic alloy. During the solidification process solid phase is forming and, since we are placing
ourselves in the case of quite rapid solidification, the solid growth will give birth to columnar
or/and equiaxed structures. We are dealing thus with a two-phase system having an extremely
complex liquid-solid interface and more, its evolution in time may imply sudden changes of the
solid structure (e.g. columnar-to-equiaxe transition). During the solidification one can identify in

II-5
Submitted to Journal of Physics D: Applied Physics

the ingot a very large number of columnar and equiaxe grains. Let us denote by N this number.
Note however that this number N is extremely variable in time due to various phenomena like
nucleation of equiaxed grains within an undercooled liquid, fusion of equiaxed grains convected
into overheated liquid and changes in column number due to changes in local cooling conditions
(cooling rate and thermal gradient). Moreover the grains in the ingot may have very different
dimensions since each columnar or equiaxed grain has a unique growth history.
In fact after nucleation, the solid nuclei will grow with respect to local undercooling
conditions. For a quite rapid solidification case the liquid-solid interface will destabilize soon after
nucleation giving birth to highly dendritic grains (Figure 1). For the equiaxed grains, due to
specific interfacial anisotropies, a precise number of primary arms are developing forming the 3D
structure. These primary arms are advancing in the undercooled melt with a velocity depending on
the local undercooling conditions at their tips. During their growth the primary arms develop
laterally a network of secondary arms and in some cases even tertiary ones. Thus, soon after
nucleation the solid-liquid interface becomes quickly highly complex. For the columnar grains the
situation is somehow different in the sense that each columnar grain evolves around a main primary
arm oriented along the heat flux direction, giving birth to the well known quasi-parallel columnar
structure (Figure 1). However during its growth the primary arm develops lateral secondary or
tertiary arms, similar to the equiaxed case. In this light, the main difference between the two
structures lies in the fact that columnar grains evolve along a preferred direction (opposite to the
heat flux lines) and have no growth constraints far from the primary arm tips. In contrast equiaxed
grains have no preferred growth direction, and their growth is constrained by the presence of other
equiaxe grains uniformly distributed around it. As a consequence of these constraints, in contrast
with the relative isotropic equiaxed grains, the columnar grains are highly anisotropic ones due to
their preffered growth direction (Figure 1). Hence one deals with a very complex two-phase system
since to describe it one should have information on various parameters describing each of the
columnar or equiaxed grains: their position, dimensions, volume grain density as well as their
orientation. In this context a simplified appoach in describing the grains will be addopted.
First, due to their quasi-isotropic arrangement in space, the equiaxed grain distribution will
be in the following approximated with a local homogenoeus distribution of spherical dendritic
grains. Their local grain density will be denoted with n ⎡ m −3 ⎤ and their radius with a [ m ] . Note as
⎣ ⎦
−1/ 3
well that one can easily identify the length scale R1 = ( 4 / 3π n ) defining locally the average
spacing between two neighbour particles. However, that the equiaxed grains can be highly
dendritic. In this context the radius a must be seen as a measure of the extension in space of the
whole grain solid matrix and not of the smallest solid scale (inter-dendritic spacing). This apprach
is similar with the envelope model introduced by Rappaz and Thévoz [14] who introduced a virtual
surface envelopping the solid matrix and separating thus the inter-dendritic liquid from the extra-
dendritc liquid. The same apprach has been succesfully used by Wang and Beckermann [6] within
their volume averaged model.
Secondly, due to their highly anisotropic shape, the columnar grains could be locally
approximated with a cylindrical shape oriented along their prefered growth direction. However,
besides from being more complex this approach would also mean that one would have to use two
different models for the equiaxed and the columnar zone respectively. In an attempt to obtain a
average model that uses the same mathematical formulation for both equiaxed and columnar grains
we will try to adopt in the following a simplified approach for the columnar grains. In fact we will
approximate locally the columnar grains as beeing quasi-isotropic. Given this hypothesis one can
approach locally the columnar grains with a spherical model.

II-6
Submitted to Journal of Physics D: Applied Physics

a) b)
Figure 2: a) The columnar structure and the corresponding grain fraction variation along their
preferred growth direction; b) The equivalent columnar structure formed by spherical grains.
Let us consider a columnar structure (Figure 2a) being characterized with a primay arm
spacing, λ1 . Using the cylindrical approximation to the columnar grain enveloppe, one can easily
compute the grain fraction variation along their preferred growth direction:

a2
ε gcyl = (11)
λ12

Approximating locally the columnar grains with an isotropic model would mean to replace the
original columnar zone with an equivalent columnar structure (Figure 2b) formed by spherical
grains. However, one should use for the spherical structure the same main parameters - the average
spacing between grains ( λ1 ) and the grain volume fraction ( ε gcyl ) – as those defining the original
columnar structure:

R1 = λ1 , ε gsph = ε gcyl (12)

Hence, to a first approximation, we have approached the columnar zone with an equivalent
spherical model for which the two main geometrical characteristics, the average spacing between
the particles and the grain fraction were inherited from the real columnar zone, ( λ1 and ε gcyl ).
Thus, the columnar zone is locally approximated with a homogeneous spherical grain distribution,
having a grain density determined from the primary arm spacing as follows:

−1
⎡4 3⎤
nc = ⎢ π ( λ1 / 2 ) ⎥ (13)
⎣3 ⎦

This is somewhat equivalent with transforming the columnar zone into an equivalent equiaxed one.
However this is a particular “equiaxed” zone since one must have in mind the fact that there is a
strong statistical correlation between these spherical particles: first no mobility of the grains is
allowed and secondly an axial gradient of the local grain fraction ε g would be equivalent with a
local variation of the columnar grain diameter along their preferred growth direction. It is
worthwhile to notice that this approximation was already used by Martorano et al. [15]. This
hypothesis was also supported by numerical results [15]. Indeed, the computation of the solute

II-7
Submitted to Journal of Physics D: Applied Physics

diffusion process in the extra-dendritic liquid for the cylindrical and spherical approach
respectively showed that a negligeble difference exists between the two approaches. Since the
molecular diffusion processes (e.g. of solute or momentum) around the grains depend strongly on
the length scale of the extra-dendritic liquid, ll = λ1 − a , it would be interesting to evaluate the error
introduced by the spherical approximation of the columnar zone when computing ll . One has:

llcyl 1 − ε g1/ 2
= (14)
llsph 1 − ε g1/ 3

As one can easily observe, in the range ε g ∈ [ 0,1] , the difference between llcyl and llsph does not
exceed 50%, and the two length scales remain thus of the same order of magnitude. Therefore the
corresponding diffusion processes in the extra-dendritic liquid would remain of the same order of
magnitude as well, supporting thus the numerical findings of Martorano et al. [15].
In this context, in the following we shall treat both the equiaxed grains and the column in
the same way, that is using a spherical approximation.

3.1. Probability distribution functions

A detailed derivation of the probability distribution for a case of a simple two-phase system can be
found in Drew [9]. However the two-phase environment considered in that study was a simple one:
the number of solid particles in the system was constant (N), the spherical particle were identical
and had a known dimension (a). This is completely different from the two-phase system one could
encounter during a solidification process (variable number of solid particles having variable
dimension as well).
In this light, the new probability distribution functions must be computed taking into
account the complexities of a solidification problem. The main difficulty in deriving these
distributions resides in the fact that the number of particles in the ingot is varying with time.
Indeed, considering the multiphase system as a repeatable one, Drew [9] supposed the existence of
a master function f ( ) ( t , z1 , v1 , a1 ,… , z N , v N , aN ) able to describe the randomness of the system
N

such as:

f ( ) ( t , z1 , v1 , a1 ,… , z N , v N , aN ) dz1dv1da1 … dz N dv N daN
N
(15)

is the probability of finding at time t a particle within dz1 of z1 , having the velocity within dv1 of
v1 , the diameter within da1 of a1 and finding a particle within dz 2 of z 2 , having the velocity
within dv 2 of v 2 and the diameter within da2 of a2 , etc.
For a case in which N is variable in time the existence of a single master function has no
more meaning. In turn, one can define a family of master functions each corresponding to a certain
grain number N in the ingot, that is:

f ( N , t , z1 , v1 , a1 ,… , z N , v N , aN ) , N ∈ +
(16)

The distribution function above has the meaning that:

f ( N , t , z1 , v1 , a1 ,… , z N , v N , aN ) dz1dv1da1 … dz N dv N daN (17)

is the probability that at the moment t we found N particles in the ingot (columnar or equiaxed
grains), each particle i ( = 1, N ) being centered within dz i of z i , having the velocity within dvi of
vi and the diameter within dai of ai . It is interesting to note that with the help of these
distributions one can easily compute the probability of finding in the ingot, at time t, N particles:

II-8
Submitted to Journal of Physics D: Applied Physics

PN = ∫∫ ∫ f ( N , t , z1 , v1 , a1 ,…, z N , v N , aN ) dz1dv1da1 … dz N dv N daN (18)

Obviously the probabilities PN verify:

∞
∑Pi = 1 (19)
i =1

To simplify the mathematical formulation from now on we will use the following master
distribution function which integrates only the position of particles:

f ( N , t , z1 ,… , z N ) = ∫∫ ∫ f ( N , t , z1 , v1 , a1 ,…, z N , v N , aN ) dv1da1 … dv N daN (20)

where f ( N , t , z1 ,… , z N ) dz1 … dz N is the probability of finding at time t in the ingot N particles,

centered within dz i of z i . It is important to notice that the use of the particle distribution f in (20)
does not imply a loss in generality since all information related to particle velocity or dimension are
included in f by means of integrations in (20).
With the help of (20) we can define an important density function which will be
extensively and this is the unconditional density function for any sphere center being at z1 :

∞
f ( ) ( t , z1 ) = f (1, t , z1 ) + ∑ ⎡⎣ ∫∫ ∫ f ( i, t , z1 ,…, zi ) dz 2 … dzi ⎤⎦
1
(21)
i =2

having the property that f ( ) ( t , z1 ) dz1 is the unconditional probability that at time t in the ingot
1

one can find a particle centered within dz1 of z1 . Note that f ( ) verifies as well:
1

(1)
∫ f ( t , z1 ) dz1 = 1 (22)

This distribution is a very important one since it can be linked to the bulk particle density
n, a key parameter defining the nucleation phenomena. Indeed, if we admit at z1 a local
( )
homogeneous grain density n t , z1 we have:

f ( ) ( t , z1 ) = n ( t , z1 )
1
(23)

Note however that all distributions defined until now refer only to the position of particles
at the time t. In turn we would like to compute the ensemble average of a certain field ψ at x and
at time t. Thus, to describe statistically a physical field ψ ( x,t ) one should define a distribution
function which includes the position of particles with respect to a certain point P ( x, t ) . Following
Drew [9] a simple way to do this would be to order the N particles in the ingot following their
distance from the reference point x . Let us consider the following rearrangement of the particles:

z i1 , z i2 ,…, z iN (24)

such as the particles i1 , i2 ,…, iN be ordered in order of closeness to the point x , that is:

x − z i1 ≤ x − z i2 ≤ … ≤ x − z iN (25)

II-9
Submitted to Journal of Physics D: Applied Physics

Since (24) is only a rearrangement of the original series of particles z1 ,…, z N it is obvious
that

( )
f N , t , z i1 , z i2 ,… , z iN = f ( N , t , z1 ,… , z N ) (26)

By denoting the positions of the rearranged particles as follows:

z1 = z i1 ; z 2 = z i2 ;…; z N = z iN (27)

one can define a new family of distribution function: the nearest-neighbor distribution functions
( )
f N , t , x; z1 , z 2 ,… , z N . As noted the new distribution function is no more that a transformation of
(20):

( )
f N , t , x; z1 , z 2 ,… , z N = f ( N , t , z1 ,… , z N ) (28)

Roughly the distribution f has the interpretation that:

( )
f N , t , x; z1 , z 2 ,…, z N d z1d z 2 … d z N (29)

is the probability that one find in the ingot, at time t, N particles and that the closest particle to the
point x is centered within dz1 of z1 , that the next closest particle is centered within dz 2 of z 2 ,
etc.
The nearest-neighbor distribution in (28) is of extreme importance since it involves a
reference point x in space. This eneables ones to easily analyse statistically any physical field ψ
at x and t.. Moreover the distribution f include valuable information on the position of particles in
order of closeness to point x . Indeed knowing that the closest particles influence to a greater
extend the field ψ at x than the farthest ones do, we can think to limit our interest to the closest
particles to the point x . Doing this we could simplify extensively the closure problem inherent to
any average approach. In this respect the computing of the unconditional distribution function that
the closest particle to the point x is centered within d z1 of z1 , would be of great interest:

∞
(1)
f ( ) (
t , x, z1 = f 1, t , x, z1 + ∑ ⎡ ∫∫
⎣ ) ∫ f ( i, t , z1 ,…, zi ) d z 2 … d zi ⎤⎦ (30)
i =2

(1) (1)
The interpretation of f is that f ( t , x, z1 ) d z1 represents the unconditional probability that the
closest particle to the point x is centered within d z1 of z1 .

3.2. The ensemble average

The derivation of probability distribution functions enables us now to approach annalytically the
abstract ensemble average formulation defined in (1). Let us denote by ψ ( t , x ) a generic physical
field. The value of ψ at t and at x will vary naturally with respect to each realization of the
process and consequently with the corresponding configuration of the two-phase system (number
of particles in the ingot, their position, velocity, dimensions, etc.)

ψ ( t , x ) ≡ ψ ( t , x; N , z1 , v1 , a1 ,… , z N , v N , aN ) (31)

II-10
Submitted to Journal of Physics D: Applied Physics

Since the ensemble average ψ ( t , x ) is the average of ψ ( t , x ) on the ensemble of all possible
realizations of the two-phase process, ψ ( t , x ) can be expressed as the integration of ψ ( t , x ) over
all possible configuration of the multiphase system, integration weighted with the corresponding
configuration probability, that is f ( i, t , z1 , v1 , a1 ,… , z N , v N , aN ) :

∞
ψ ( t , x ) = ∑ ⎡⎣ ∫∫ ∫ f ( i, t , z1 , v1 , a1 ,…, zi , vi , ai )ψ ( t , x ) dz1dv1da1 … dzi dvi dai ⎤⎦ (32)
i =1

To simplify the mathematical formalism on one hand and to concentrate only on the
position of the particles on the other hand we will introduce from now on conditional averages of
ψ . Indeed it is easy to observe that (32) can be rewritten as follows:

∞
ψ ( t , x ) = ∑ ⎡⎣ ∫∫ ∫ f ( i, t , z1 ,…, z i )ψ ( t , x i, z1 ,…, z i ) dz1 … dzi ⎤⎦ (33)
i =1
in which:
1
ψ ( t , x i, z1 ,… , z i ) = ∫ f ( i, t , z1 , v1 , a1 ,…, z i , vi , ai )
f ( i, t , z1 ,… , z i ) ∫∫ (34)
× ψ ( t , x ) dv1da1 … dvi dai

is the conditionally averaged field ψ , averaged on the condition that in the ingot are i particles,
each of them centered in z i ,… , z i . We have thus restricted ourselves to the position of the particles
in the ingot solely, the information regarding the particle velocity and their dimension being now
included in f ( i, t , z1 ,… , z i ) and ψ ( t , x i, z1 ,… , z i ) expressions respectively. However, the
computation of ψ ( x, t ) with the help of equation (33) still remains a challenging issue and this, of
course, to the large number of particles in the ingot. Therefore we will restrict even more our
interest to a single particle, the one centered in z1 . Using again the conditional averages we can
write:

ψ ( t , x ) = ∫ f (1) ( t , z1 )ψ (1) ( t , x z1 ) dz1 (35)

where:
1
ψ (1) ( t , x z1 ) =
f ( ) ( t , z1 )
1

(36)
⎧
⎪ ∞ ⎡
× ⎨ f (1, t , z1 )ψ ( t , x 1, z1 ) + ∑ ⎢
∫∫ ∫ f ( i, t , z1 ,…, zi ) ⎤ ⎫⎪
⎥⎬
⎪⎩ i =2 ⎢
⎣× ψ ( t , x i , z1 ,… , z i ) dz 2 … d z i⎥
⎦ ⎭⎪

is the conditionally averaged field ψ , averaged on the condition that there is a particle centered
within dz1 of z1 .The distribution f ( ) ( t , z1 ) is no more than the unconditional density function
1

for any sphere center being at z1 already defined in (21) and (23).
In a similar way, using this time the nearest-neighbor distribution functions, we can define
the ensemble average of ψ as follows:

ψ ( x, t ) = ∫ f
(1)
( t, x, z1 )ψ (1) ( t , x z1 ) d z1 (37)

II-11
Submitted to Journal of Physics D: Applied Physics

(1)
in which f ( t, x, z1 ) is the unconditional probability already defined in (30) and:
ψ
(1)
(t, x z ) =
1
1
f
(1)
( t , x, z1 ) (38)
⎡ ∞ ⎤
( )
× ⎢ f 1, t , x, z1 ψ ( t , x ) + ∑ ⎡ ∫∫
⎣ ( )
∫ f i, t , x, z1 ,…, z i ψ ( t , x ) d z 2 … d z i ⎤⎦ ⎥
⎣ i =2 ⎦

represents the conditionally averaged field ψ , averaged on the condition that the closest particle to
the point x is centered within d z1 of z1 . The ensemble average of ψ expressed as in equation
(37) is of extreme importance because, given certain assumptions, one is able to express
analytically the distribution f
(1)
( t, x, z1 ) and the conditional average ψ (1) ( t , x z1 ) .
Indeed, we have seen that if we consider locally at z1 and t an uniform grain density
n ( t , z1 ) then the unconditional distribution function f ( ) ( t , z1 ) for any sphere being at z1 equals
1

n , equation (23). Using this result, Drew [9] proved that the nearest-neighbor distribution
f
(1)
( t, x, z1 ) has the following analytic expression:
f ( t , x, z1 ) = n ⋅ exp ⎢ − π n ( x − z1 ) ⎥
(1) ⎡ 4 3⎤
(39)
⎣ 3 ⎦
As we will see in more detail later, the conditionally averaged field ψ
(1)
(t, x z )
1 can be
approached analytically if one assumes that that the field ψ ( t , x ) is mainly influenced by the
closest grain to x and that the influence of distant grains is negligible. This would be equivalent to
what Drew [9] pointed out, namely “to assume that each grain is isolated in the sense that it only
interacts with its neighbors through the average fields”.

3.3. The cell model

The use of the exact distribution function in (39) to compute the ensemble average in (37) will
increase excessively the complexity of the closure problem. Thus, a simpler approximation for
(1)
f ( t, x, z1 ) would be of great interest. Looking at the exact distribution (Figure 3) one can see

that the probability to find the closest grain at the distance r = x − z1 from the point x decreases
(1)
exponentially with the increase of r . Indeed f ( t, x, z1 ) remains of order of n for exponential

arguments 4 3π nr 3 of order of one and decreases dramatically with further increase of 4 3π nr 3 .

II-12
Submitted to Journal of Physics D: Applied Physics

Figure 3: The exact form (continuous line) and the cell model approximation (dashed line) of the
(1)
probability distribution f ( t, x, z1 ) for n = 109 m−3
(1)
This variation for f clearly evidence the existence of a characteristic distance R1 from
x above which the probability to find the closest particle to x becomes negligible. In this light, an
(1)
interesting approximation for f would be the following "top-hat" distribution:

⎧n for x − z1 ≤ R1
(1) ⎪
f ( )
t , x, z1 = ⎨ (40)
⎪⎩0 for x − z1 > R1

This approximation is called the "cell model" approximation to the one particle
distribution [9]. The characteristic length R1 can be analytically computed since we know that the
probability of finding the center of the closest particle to x between x and ∞ is 1, that is:

⎛ ⎞
(1)
lim ⎜ ∫ f
R →∞ ⎜
⎜ x − z1 ≤ R
(
t , x, z1 d z1 ⎟ = 1
⎟⎟ ) (41)
⎝ ⎠

Using (40) together with (41) one can easily obtain:

− 13
⎛4 ⎞
R1 = ⎜ π n ⎟ (42)
⎝3 ⎠

This is a very important result since it evidences the existence at each point x of a sphere of radius
R1 within which we can find one and only one particle. So, using the probability distribution
functions we were able to retrieve a local characteristic length which quantifies in a certain point of
the two-phase system the amount of space within which one can find one and only one particle. For
this reason a cell of radius R1 enveloping a particle will also enclose the space proper to that
particle, the “zone of influence” characterizing that particle (Figure 4). Outside this cell one will
found itself under the influence of a neighbor particle. In other words, the characteristic length R1
represents half of the average distance between two adjacent grains. Notice however that all the
above conclusions are based on the assumption of a local homogeneous grain distribution n at x.

II-13
Submitted to Journal of Physics D: Applied Physics

Figure 4: The average distance between two adjacent grains

Using the cell approximation to f

(1)
( t, x, z1 ) , the ensemble average ψ ( t , x ) defined in
(37) becomes:
ψ ( x, t ) = ∫ f
(1)
( t, x, z1 )ψ (1) ( t , x z1 ) d z1
(t, x z ) d z
(1) (43)
=n ∫ ψ 1 1
x − z1 ≤ R1

(1)
Notice that the use of the cell approximation for f transforms the improper integral in (37) into
a definite volume integral over the sphere of radius R1 and centered at x . Notice as well that the
integration in (43) is made with respect to z1 , meaning that we integrate over all position of the
grain in the cell. This result is of great importance because, one one hand it simplifies the ensemble
average formulation and on the other hand, and most importantly, it quantifies the ensemble θ of
all possible realisations of the multiphase flow, so abstactly used in (1). Indeed the ensemble θ
indentifies now with the ensemble of flows for which the nearest particle to x is centered
anywhere within the sphere ("cell") of radius R1 and centered at x (Figure 5a). Hence, the
integration in (43) is made over all positions of the grain in the cell that is over the ensemble of all
possible realisation of the multiphase system. The equation (43) represents no more than the "cell
approximation to the ensemble average”.

Figure 5: a) The cell approximation to the ensemble average; b) The cell approximation to the
volume average

II-14
Submitted to Journal of Physics D: Applied Physics

However, to be able to compute ψ ( x,t ) one has to approach analytically ψ

(1)
(t, x z ) . As
1

already anticipated one can assume that the field ψ ( x,t ) is mainly influenced by the closest grain
to x . In this respect, by negleting the influence at x of all grains except the closest one to x, the
(1)
conditionnaly averaged field ψ can be approximated with the exact solution of ψ ( x,t ) given a

single grain at z1 . Hence ψ

(1)
(t, x z ) ≅ ψ (t, x; z ) . Equation (43) becomes:
1 1

ψ (t, x ) = n ∫ ( )
ψ t , x; z1 d z1 (44)
x − z1 ≤ R1

One important consequence of (44) is that, in order to compute ψ , one has to evaluate the
(
exact field ψ t , x; z1 ) around a particle only to maximum distance R1 from its center z1 . As
already explained, above this limit the field would not be anymore “proper” to the particle centered
in z1 , but to its neighbours. Using equation (44) let us compute now some of the average quantites
defined in Table 1 for a simple two-phase particulate system (spherical particles of radius a ,
uniformely distributed in space and having a known particle density n p ).

Figure 6: Computation of a solid (a) and of an interfacial (b) ensemble average

The solid fraction becomes:

(1)
ε s (t, x ) = ∫ f ( t, x; z1 ) X s ( t, x ) d z1
x − z1 ≤ R1
(45)
= np ∫ X s ( t , x ) d z1
x − z1 ≤ R1

Taking into account the definition of the characteristic function X s the last integral transforms in
an integral over the positions of the particle within the cell such as the point x is within the solid
phase. This is equivalent with an integral of the particle position over the spherical domain
x − z1 ≤ a (Figure 6a):

II-15
Submitted to Journal of Physics D: Applied Physics

ε s (t, x ) = n p ∫ d z1
x − z1 ≤ a
(46)
4 ⎛ a3 ⎞
= n p π a3 ⎜⎜ = 3 ⎟⎟
3 ⎝ R1 ⎠

Using a similar reasoning the liquid fraction ε l ( t , x ) becomes:

εl (t, x) = n p ∫ d z1
a ≤ x − z1 ≤ R1
(47)
⎛ a3 ⎞
4
= n p π R13 − a3
3
( ) ⎜⎜ = 1 − 3 ⎟⎟
R1 ⎠
⎝

In the latter case the integration is made over the positions of the particle within the cell such as the
point x is within the liquid phase, that is over the sperical domain a ≤ x − z1 ≤ R1 (Figure 6a). It is
worthwhile to notice that the results for the solid/liquid fraction are identical with the ones obtained
by Wang and Beckermann [6] using the volume average technique and the cell approximation to
the volume average. However (46) and (47) result from a closure technique fundamentally different
from the one used in [6]. Indeed, the equations (45) and (47) engage integrations over the position
of the closest particle to x, z1 , such as the point x be always in the solid or in the liquid phase,
respectively. Therefore, during the average process the field to average is always evaluated in the
phase for which one computes the average. Note that this seemingly minor feature is one of the key
advantages of the cell approximation to the ensemble average.
The approach used by Wang and Beckermann [6], the cell approximation to the volume
average, considers an equivalent cell of radius R1 (Figure 5b). However, during the average
process the particle is fixed and centered at x . All fields are averaged in this cell and the average
results (including the liquid and interfacial averages) are artificially attached to the point x , a point
that it is always in the solid. As we will see later on this incoherence may produce wrong results,
especially at the evaluation of interfacial averaged fields.
For a generic field ψ ( t , x ) the solid and liquid ensemble average becomes respectively:

(1)
ψ s (t, x) =
1
εs ∫ f ( t, x, z1 ) X s ( t , x )ψ ( t , x; z1 ) d z1
x − z1 ≤ R1

⎛ ⎞ (48)
np
=
εs ∫ (
ψ t , x; z1 d z1 ) ⎜= 1
⎜⎜ 4 3π a3 ∫ (
ψ t , x; z)
1 d z1
⎟
⎟⎟
x − z1 ≤ a x − z1 ≤ a
⎝ ⎠
and:
np
ψ l (t, x ) =
εl ∫ (
ψ t , x; z1 d z1 )
a ≤ x − z1 ≤ R1
(49)
=
1
∫ ( )
ψ t , x; z1 d z1
(
4 3π R13 − a3 ) a ≤ x − z1 ≤ R1

Notice again that the integrations in (48) and (49) are made over the position of the particle such as
the point x is respectively in the solid or the liquid phase.
Let us compute now the interfacial average S s (the area density, see Table 1):

II-16
Submitted to Journal of Physics D: Applied Physics

Ss = n p ∫ n s ⋅ n sδ ( x − xi ( t ) ) d z1
x − z1 ≤ R1
(50)
= np ∫ dΩ ( = n p 4π a 2 )
Ω( a )

Because ∇X k has a meaning at the solid-liquid interface only, the volume integral in (50)
transforms into an integral over the positions of the particle such as the point x is on the interface,
namely into a surface integral over the spherical envelope of radius a , centered at x (Figure 6b).
Notice that, since the liquid and the solid share the same interface, S s = Sl .
Let us compute now the two averaged interfacial sources Φψ k and Γψ k which quantifies
the two types of interactions between the solid and the liquid (see Table 1). Since we are dealing
with solid rigid particles, the relative velocity w k = ( vi − v )k and the molecular flux jψk at the
phase k interface have only normal components:

w k = wk n k
(51)
jψ = jψk n k

Having this in mind Φψ k and Γψ k become:

Φψ k = n p ∫ ( − jψ n ) ⋅ ⎡⎣−n δ ( x − x (t ))⎤⎦ d z
k k k i 1
x − z1 ≤ R1
(52)
= np ∫ jψk d Ω
Ω( a )
and:
Γψ k = n p ∫ ρψ ( − wk n k ) ⋅ ⎡⎣ −n k δ ( x − xi ( t ) ) ⎤⎦ d z1
x − z1 ≤ R1
(53)
= np ∫ ρψ wk d Ω
Ω( a )

where again the integration is made over the positions of the particle such as the point x is on the
interface. It is customary to rewrite the flux Φψ k as follows:

* * 1
Φψ k = Sk jψk , jψk = ∫ jψk d Ω (54)
Ak Ω( a )
*
where Sk is the area density of k phase interface, jψk is the interfacial average flux of ψ at the k
phase interface. The latter equation evidence the interfacial diffusion process caused by the
molecular diffusion of ψ at the interface.
Likewise, Γψ k can be rewritten as follows:

*
Γψ k = Γ k ψ k (55)
where:

( )
Γ k ≡ Γ1 k = ∫
Ω( a )
ρ wk d Ω and ψk =
* 1
Γ k Ω(∫a )
ρψ wk d Ω (56)

II-17
Submitted to Journal of Physics D: Applied Physics

are respectively the mass transfer rate at the k interface (the solid/liquid mass transfer rate in our
case) and the average interfacial field at the k interface, averaged with respect to the mass transfer
rate. This latter formulation evidence clearly the physical meaning of Γψ k , namely the interfacial
transfer of ψ due to the mass exchange between phases.

3.4. Model closure

Let us now particularize these results with a real binary alloy solidification case. The processes
taking place within an ingot during a rapid solidification are very complex to model since various
coupled phenomena like nucleation, grain growth, coarsening need to be accounted. In view of this
complexity one should first propose physical models at the micro-scale for all these processes and
only then try to compute closure expressions for the unknown terms in the ensembled averaged
equations like (8). Hence, first one has to write down the physical assumptions valid at the micro-
scale.

3.4.1. Micro-scale physical assumtions.

A pure diffusion case. The first assumption we made refers to the fluid flow regime. In the
folowing we will neglet the existence of any flow in the system. So v s and v l will be zero and the
balance equation for momentum will become identical zero. Therefore, the final model will include
only balance equations for mass, energy and solute.

Phase diagram aspects.

During the binary alloy solidification the solute concentration field has a discontinuity at the solid-
liquid interface. Indeed the solute concentration at the liquid interface Cl∗ is linked directly to the
interface temperature through the liquidus curve in the phase diagram. In contrast, due to a lower
solubility of the solute in the solid, the solute concentration at the solid interface Cs∗ is usually
lower than Cl∗ (for the case of ususal metal alloys). Cs∗ can be also linked directly to the interface
temperature through the solidus curve this time. One can approximate the liquidus and the solidus
curves with a linear law and therefore express the two interfacial concentrations as follows:

T − Tf
Cl∗ = , Cs∗ = kCl∗ (57)
m

where m is the liquidus line slope, T f is the fusion temperature of the pure solvant and k is the
partition coefficient (<1).

Solute diffusion controlled solidification

Due to a high value of the Lewis number, i.e., the ratio between the thermal and the solute
diffusivity in the liquid Le = α / Dl , the solidification of a binary metallic alloy is mainly controlled
by the solute diffusion. In this respect the solute conservation at the solid-liquid interface:

∂Cl ∂C
Cl∗ (1 − k ) wn = − Dl + Ds s (58)
∂nl ∂ns

is sufficient for the computation of the solid liquid interface velocity, wn . It is worthwhile to notice
that the solid interface evolution is intrinsically linked to the solute diffusion at the solid/liquid
interface. Therefore, great attention must be paid to the modeling of the solute flux at the solid-
liquid interface. Notice as well that for binary metallic alloys the solute diffusivity in the solid, Ds ,
has a very small value compared with Dl , and it is usually neglected except for some alloys like
steel.

II-18
Submitted to Journal of Physics D: Applied Physics

Local thermal equilibrium

The hypothesis of a local thermal equilibrium was constantly employed in different average
solidification models [6, 7, 14, 15]. In brief, the local thermal equilibrium states that, due to the
high thermal diffusivity characterizing the metallic alloys, one can assume that locally, at the scale
of the grain, the temperature field is uniform: Ts = Tl . The use of such an approximation is based
on the assumption that due to the high alloy diffusivity the diffusion time at length scales of the
grain is small enough to consider a local uniform temperature. Indeed, if l g denotes the length
scale of a dendrite grain (usually ∼ 10−4 − 10−3 m ) and α the thermal diffusivity ( ∼ 10−5 m 2 / s ),
( lg )
2
the thermal diffusion time at the scale of l g becomes td / α = 10−1 − 10−3 s , thus a quite

small one. However, for average or high local cooling rates ∂T / ∂t > 1 − 102 K / s the order of
magnitude of the temperature differences at the scale of the grain ∆Tg ( ( ∂T / ∂t ) td ) becomes
non-negligible ( > 0.1 K ) . Even more, for solidification processes that imply high thermal gradients
G > 103 − 104 , the order of magnitude for temperature differences at the scale of the grain
( ∆Tg Gl g ) may be high enough > 1 K as to invalidate the hypothesis of local thermal
equilibrium.
This hypothesis is of primary importance since, as we will see, it greatly simplifies the
closure expressions. Moreover, as one will notice, the closure expressions obtained given this
hypothesis are very similar with those obtained by Wang and Beckermann in their volume
averaged model [6]. To anticipate, notice that if one does not consider anymore a local thermal
equilibrium at the scale of the grain, the results obtained using the statistical approach are
completely different from those obtained using a volume average model. Even more, the volume
averaged approach gives birth to non-physical closure expressions.

Equiaxed nucleation
The nucleation phenomena are intrinsically linked to the equiaxed grain nucleation within an
undercooled melt. For a pure equiaxed case (no columnar grains) we will approach the nucleation
with an instantaneous phenomena triggering at a temperature TNe . The latter one is usually lower
that liquidus temperature corresponding to the local liquid concentration: TNe = Tliquidus − ∆Tnucl ,
where ∆Tnucl represents the nucleation undercooling. The instantaneous germination model has
been proposed by Stefanescu et al. [16] and was extensively used in average models [6, 7] due to
its simplicity. In this respect, when the local liquid temperature reach the critical temperature TNe ,
solid nuclei appear instantaneously in the undercooled melt with a density equal to ne0 . The value
of ne0 is a input parameter of the model difficult to predict. One could have an insight on this
parameter by means of post-mortem microstructure analysis or, for the case of refined alloys, by
estimating the number of refining particles introduced in the ingot. However, ne0 along with the
nucleation undercooling ∆Tnucl still remains the parameters with the highest degree of incertitude
among all input parameters used in the solidification model.
Another parameter linked to the nucleation process is the radius of the initial nuclei a0 . We
have considered a0 = 10−6 m , similar to the value used in [7]. Numerical test were done with radius
a0 smaller than 10−6 m and no significant influence was observed.

Columnar “nucleation”
As mentioned the nucleation phenomena is linked to the equiaxed solidification only. However, for
an average model describing the columnar solidification as well one will have to model the sudden
appearance of a columnar dendrite tip in a certain point of the ingot. We will denote this particular

II-19
Submitted to Journal of Physics D: Applied Physics

phenomenon with the “columnar nucleation”. Describing the “columnar nucleation” phenomena
will involve, just as for equiaxed nucleation, the definition of the local undercooling conditions at
which the columns tips are supposed to appear in one point and on the other hand of the density of
columns grains nc0 .
As already mentioned, the columnar structure is approximated with an isotropic model (a
spherical model). In this context, the density of columns nc0 is quantified by the average spacing
between two dendrite tips, namely the primary dendrite spacing λ1 , e.g. equation (13). In turn, λ1
can be easily related to local conditions like the local thermal gradient and local isotherms velocity
by means of various models existing in the literature [17, 18].
Defining the undercooling conditions for the “columnar nucleation” would mean, just as
for the case of equiaxe nucleation undercooling, defining the critical parameters at x
corresponding to the sudden appearance of the columnar tips at x . However, this approach would
consider the sudden appearance of a columnar tip at x as a local event, depending on local
parameters only (like temperature, solute concentration, nucleation undercooling). This is not
physically valid since the columnar front position is a function not only of local parameters at the
tip but of the whole history of the tip evolution. In this light and in contrast with the equiaxed
nucleation the “columnar nucleation” can not be treated as a local event.
Thus, to determine the precise moment at which the dendrite tip appears at x one would
need to track at each moment the position of a dendrite tip in the ingot. This could be done with the
help of a front tracking technique which demands in turn the evaluation of the instantaneous
velocity of the columnar tip Vtip with respect to the local undercooling conditions. One can assume
that the tip velocity can be linked at any moment to the local tip undercooling using a standard tip
kinetic theory [19]. The latter hypothesis is not straightforward since the existing kinetic theories of
a dendrite tip are valid for stationary state at the dendrite tip. Therefore, the use of a stationary
solution for Vtip for a case where the local tip undercooling conditions are nonstationary would not
be valid. However if the velocity response time to changes in local undercooling conditions is
much smaller that time scale characterizing the variations of local conditions, the dendrite tip
velocity can be approximated with the steady state solution. Indeed, for rapid solidification
processes involving fast advancing columnar dendrites, the dendrite tip radius rtip is very small
( ∼ 10−6 − 10−5 m ). Knowing that the solute boundary layer at the tip is of the same order as rtip the

( )
2
time scale characterizing the solute transport at the tip would be of order of τ tip ∼ rtip / Dl , thus
small too ( ∼ 10 −1
− 10−3 s ) for a typical Dl ∼ 10−9 m 2 / s . Because the instant tip velocity is
intrinsically linked to the solute transport at the tip, equation (58), the velocity response time of the
tip subject to changes in local undercooling conditions will be of order of τ tip . On the other hand
the local undercooling conditions are subject to changes as well. These changes are produced
mainly by macroscopic processes (fluid flow and heat transfer at the scale of the ingot) and their
characteristic time scale is usually much larger than τ tip . In this respect, one can approach the tip
velocity with the results from a steady state tip kinetic theory like the one proposed by Lipton et al.
[19].
If the implementation of a front tracking technique for a 1D domain is quite simple, the
columnar front tracking for a 2D or 3D case can be extremely difficult and highly time consuming.
Even more, the columnar front tracking technique represents a discrete approach and its use within
an average model whose main advantage is precisely to avoid any interface tracking would not be
consistent. In this context, the “columnar nucleation” represents one of the major problems in
columnar modeling and in modeling the coupling between columnar and equiaxe structures (e.g.
CET). The existence of a model which would help avoiding any front tracking would be of great
interest. As already discussed, one would have to model the “columnar nucleation” as a local event
depending on local parameters only. However for reasons of clarity, we will consider in the
following that the the exact position of the columnar front within the ingot, xcf , is a known

II-20
Submitted to Journal of Physics D: Applied Physics

parameter (e.g. obtained from a front tracking algorithm). In this respect, the columnar nucleation
can be approached as an instantaneous phenomena somewhat similar with the equiaxed case.

Grain growth. Envelope model.

For rapid solidification processes the solid-liquid interface quickly destabilizes after the nucleation
and gives birth to highly complex dendritic grains. To be able to model the precise evolution of the
dendritic interface one should precisely model the solute fluxes within the complex interdendritic
network. Obviously this would be a very difficult task. However, the problem of modeling the
grain evolution can be simplified if one compares the length scales characterizing the interdendritic
and the extradendritic zone (Figure 7). Indeed, the interdendritic spacing λ2 is of order of the
primary arm tip radius and measures typically 10−5 m (in some cases even smaller), whereas the
length scale of the extradendritic zone is of order of the average spacing between equiaxe grains or,
for columnar grains, of the order of λ1 . In both cases this length scale is at least one order of
magnitude larger than the interdendritic spacing, λ2 . This particular feature of the dendritic grain
has enabled Rappaz and Thévoz [14] to admit a scale separation hypothesis between the inter- and
extra-dendritic liquid. A similar hypothesis was used by Wang and Beckermann in their volume
average model [6, 7]. The most important consequence of this hypothesis is that due to extremely
small inter-dendritic spacings the liquid between secondary and/or tertiary arms can be considered
at equilibrium with the liquid interface meaning that its concentration would be equal to the liquid
interface concentration Cl∗ . This hypothesis is valid since the solute diffusion time at the scale of
the inter-dendritic liquid τ dint er ∼ ( λ2 ) / Dl is usually much smaller that the time scale
2

characterizing the variations of local temperature (and consequently of Cl∗ ). In this context one can
separate the inter-dendritic zone and the extra-dendritic one with a virtual surface, enveloping the
external dendrite tips of the grain (Figure 7a). In the following we will denote this surface as the
grain envelope. This envelope separates the liquid phase in two sub-liquid phases: an inter-dednritc
liquid phase in equilibrium with the interface and an extra-dendritic liquid phase which can be in a
non-equilibrium state with respect to the interface. As we will see in more detail this approach is
extremely useful since it encompasses the need of modeling the evolution of the complex inter-
dendritic solid-liquid interface. To compute the solidification rate, one will only have to model the
grain envelope evolution and the solute field in the extra-dendritical zone.

Figure 7: The envelope model (a) and the coresponding solute field distribution (b)

The complex quasi-symmetrical shape of the grain envelope (Figure 7a) is simplified with
an equivalent spherical envelope as in Rappaz and Thévoz [14] and Wang and Beckermann [6].

II-21
Submitted to Journal of Physics D: Applied Physics

Notice that the envelope deviation from a spherical shape can be partially encompassed by means
of a shape correction factor [7]. However, for clarity reasons we will not use in the following this
correction.
Because the envelope joins all the dendrite tips of the grain, the envelope evolution will be
controlled by the dendrite tips local velocity Vtip . As already discussed, one can link the tip
velocity Vtip to local undercooling conditions ( T , Cl∗ and C∞ ) by using a standard tip kinetic
theory. Using the marginal stability criterion and assuming no back diffusion in the solid Lipton et
al. [19] has shown that:

Dl m ( k − 1) Cl∗
Vtip = 2
( Pet )2 (59)
π Γ

Admitting a certain shape for the dendrite tip the Peclet tip number Pet can be linked to the
( )
dimensionless solutal undercooling Ω = Cl∗ − C∞ / Cl∗ / (1 − k ) . For a spherical dendrite tip one
obtains:

Pet = Ω (60)

whereas for a parabolic shape a more complex function is found: Pet = Iv −1 ( Ω ) , where Iv −1
represents the inverse Ivantsov function. This function has no simple analytical solution and
therefore, we will use in the following the fit computed by Wang and Beckermann [7]:

b
⎣ / (1 − Ω ) ⎤⎦
Pet = a ⎡Ω (61)

where a = 0.4567 and b = 1.195 .

Note that equation (59) provides the velocity of a steady-state tip advancing in a semi-infinite
liquid domain with the concentration far away from the tip, C∞ . In reality the grain growth does
not take place within an infinite liquid domain but is constrained by the presence of neighbor
grains. As we have seen, for a local grain distribution having the grain density n the sphere of
influence proper to each grain is equivalent to the cell of radius R1 enveloping the respective grain
(Figure 4). Outside this cell one enters in the sphere of influence of a neighbor particle. One
important remark is that for given the hypothesis of local homogeneity the spherical shell of radius
R1 will play as well the role of a periodic boundary. Moreover, for a pure diffusion case and a local
thermal equilibrium at the scale of the grain, the solute field in the extradendritic liquid will be a
function of r only and the boundary condition at r = R1 will write as follows:

⎛ ∂Cl ⎞
⎜ ∂r ⎟ = 0 (62)
⎝ ⎠ R1

meaning that locally, the solute flux between two adjacent grains is zero.
In view of these results a physically plausible choice for C∞ would be C l , having in mind
however that equation (59) in which C∞ = C l would represent only a first approximation of the
real tip velocity advancing in an finite undercooled liquid domain.
Considering a quasi-stationary solute profile in the extra-dendritic liquid and using an
integral approach, Wang and Beckermann [6] found an analytic expression for Cl ( r ) which
unfortunately has no explicit form and requires a numerical evaluation. More recently, Martorano
et al. [15] proposed an complex explicit analytical approximation to Cl ( r ) . It is important to

II-22
Submitted to Journal of Physics D: Applied Physics

notice that both these solutions do not verify the zero flux condition in (62) and that their deviation
from (62) is not negligible, especially for cases involving small Pet . In this context we propose in
Appendix A an approximate solution for Cl ( r ) which encompasses both these problems: the
obtained result, equation (A.16), has a simple analytical form and verifies the zero flux condition in
(62).

3.4.2. The final averaged equations. For a pure diffusion case the model will contain the mass,
energy and the solute average equations only. The use of the scale separation hypothesis between
the inter- and extra-dendritic liquid divides the liquid in two sub-phases. In the following we will
denote with d and l subscripts respectively the inter- and the extra- dendritic liquid. For a simpler
notation we will introduce the subscript g for the "grain phase"; this is no more than the inter-
dendritic phase plus the solid phase (Figure 7a). The separation of the total liquid ( f ) in two sub-
phases transforms the two-phase problem into a three phase one: two liquid phases d and l , and
one solid phase s . Notice however that the liquid division is a virtual one and in reality the
problem still remains a two-phase problem: f + s . In fact, the three phase approach is interesting
mainly from the point of view of the solute diffusion phenomena.
Let us now compute the closure expressions for each balance equation. We consider at x a
local homogeneous grain distibution (columnar or equiaxed) with a known density n and a known
grain dimension a. We will also consider the solid and liquid densities ( ρ s and ρ f ) as being
uniform at the scale of the grain and within each phase, namely that: ρ s = ρ s and ρ l = ρ d = ρ f .
Moreover for simplicity reasons, we will drop the average sign of all variables defined in Table 1.
Note however that in the following all variables are averaged ones.

The grain balance

The general grain density balance can be written (Drew [9]) as follows:

∂n
+ div ( v s n ) = nc / e (63)
∂t

where v s is the ensemble averaged solid particle velocity (0 in the present case) and nc / e the
colunar/equiaxed grain nucleation rate ⎡ m −3 s −1 ⎤ . Using, as already discussed, the hypothesis of
⎣ ⎦
instantaneous nucleation the equiaxed grains source ne can be written as follows:

ne = ne0δ (T − TNe ) (64)

where δ (T − TN ) is the Dirac delta function, T is the local temperature, TNe the nucleation
temperature and ne0 is the equiaxed nucleation grain density. On the other hand for the columnar
grains the nucleation phenomena can be treated as being instantaneous if one considers a columnar
front tracking algorithm. In this case the columnar grain source in equation (63) can be written as
follows:

(
nc = nc0δ x − xcf ) (65)

where again δ (T − TN ) is the Dirac delta function, x is the local coordinate within the ingot, xcf
is the exact columnar front position (determined for exemple from a front tracking algorithm) and
nc0 is the columnar nucleation grain density determined from the primary arm spacing λ1 (see
equation (13)).

II-23
Submitted to Journal of Physics D: Applied Physics

The mass balance

Particularizing the generic variable ψ (see Table 2) one can easily obtain the solid and liquid mass
balance:

∂ (ε s ρs )
= Γs ,
(
∂ εf ρf ) = −Γ (66)
s
∂t ∂t

where Γ s represents the mass exchange rate between the liquid and the solid ( Γ s > 0 and Γ s < 0
means respectively solidification and fusion at x ). Γ s represents an interfacial averaged variable
and can be expressed analytically using the cell model approximation in equation (53):

Γs = ρs n p ∫ ws d Ω (67)
Ωs

Notice again that in (67) the integration is made over all possible positions of the grain
such as the point x is on the solid-liquid interface. This is equivalent to the integration over a
surface Ω s constructed with all possible positions of the grain in the cell such x is on the solid-
liquid interface. Notice as well that:

As = ∫ dΩ (68)
Ωs

is equal to the area of the dendritical grain interface since Ω s is only a transformation of the real
dendritic interface of a grain centered at x . It is customary to rewrite (67) as follows:

Γ s = ρ s S s ws (69)
where:
1
S s = n p As and ws =
As ∫ ws d Ω (70)
Ωs

represent respectively the solid-liquid interfacial area, as defined in Table 1 and (50), and the mean
solid-liquid interface velocity. Both of these are very difficult to model due to the complexity of the
dendrite shape. However, as we will see, using the scale separation hypothesis between the inter-
and extra-dedritic liquid the mass exchange rate Γ s can be determined from the solute mass
balance in the inter-dendritic liquid.
Since one will need to the track the evolution of the inter-dendritical liquid fraction a mass
balance equation must be written for the d phase:

(
∂ εd ρ f ) =Γ (71)
d
∂t
where:
Γd = n p ∫ ρ ( − wd n d ) ⋅ ⎡⎣ −n d δ ( x − xi ( t ) ) ⎤⎦ d z1 (72)
x − z1 ≤ R1

Due to the particular three-phase approach, the inter-dendritic liquid interacts with two
phases in the same time. Indeed, the inter-dendritic liquid is bounded by both the solid and the
extra-dendritic liquid. In contrast, the solid and the extra-dendritic liquid have an interface with the
inter-dendritic liquid only. This comes from the particular choice of the grain boundary which

II-24
Submitted to Journal of Physics D: Applied Physics

envelops the dendrite tips and has no finite interface with the solid [6]. Therefore, the integration in
(72) can be divided in two, one over the positions of the grain ( z1 ) such as the point x is on the
d − s interface and one over the positions of the grain such as x is on the d − l inteface

⎛ ⎞
Γ d = ρ f n p ⎜ ∫ wd − s d Ω + ∫ wd −l d Ω ⎟ (73)
⎜Ω ⎟
⎝ d −s Ωd −l ⎠

Notice that Ωd − s ≡ Ω s and wd − s ≡ − ws , whereas Ωd −l is no more than the spherical grain

envelope of radius a and centered in x (Figure 6b). Finally wd −l is the local velocity of the grain
dendrite tips and can be computed with the help of (59). Due to a uniform temperature distribution
within the cell (local thermal equilibrium) and to the spherical symmetry of the solute distribution
around the grain envelope the interface velocity wd −l will be constant with respect to the
integration in (73). One obtains finally:

Γ d = −Γ s + Γ g (74)
where:
Γ g = ρ f S g wg (75)

(
is the mass exchange between d and l phase due to grain growth, S g = n p 4π a 2 represents the )
grain interfacial area and:

1 ⎛ Dl m ( k − 1) Cl∗ 2⎞
wg = ∫ wd −l d Ω ⎜=
⎜
( Pet ) ⎟
⎟
(76)
4π a 2 Ωd −l ⎝ π 2Γ ⎠

is the mean velocity of the grain envelope (see as well equations (59), (61)).

The energy equation

To derive the energy balance equations for each phase one has to explicit first the solid and liquid
specific enthalpies hs and h f (Table 2). These will depend on the local temperature as follows:

hf = c f Tf + L ; hs = csTs (77)

where c f and cs are respectively the solid and liquid specific heats and L the latent heat of
fusion. However, using the local thermal equilibrium hypothesis, the solid and liquid temperatures
are locally equal. In this respect, a two-phase approach to the energy equation is no more suited and
a simpler mixture approach is sufficient. The mixture energy equation is obtained by summing the
solid and the liquid energy equations. With Ts = T f = T the mixture energy equation writes as
follows:

( )
∂ ⎡ ε s ρ s cs + ε f ρ f c f T ⎤
⎣
∂t ⎣ s s (
⎦ = div ⎡ ε λ + ε λ ∇T ⎤ + Γ L
f f ⎦ s ) (78)

Notice the presence in the final equation of the heat source Γ s L . This is no more than the
latent heat release due to solidification.

II-25
Submitted to Journal of Physics D: Applied Physics

The solute balance

As already discussed for the solute balance we are using a three phase approach: s + d + l . Using
the data in Table 2 the corresponding three solute balance equations can be written as follows:

∂ ( ε s ρ s Cs )
= div ( ε s ρ s Ds ∇Cs ) + Φ Cs + ΓCs (79)
∂t
(
∂ ε d ρ f Cd ) = div ε ρ D ∇C + ΦC + ΓC
∂t
( d f f d) d d (80)

∂ ( ε l ρ f Cl )
= div ( ε l ρ f D f ∇Cl ) + Φ Cl + ΓCl (81)
∂t

To close this system one has to propose closure expressions for each of the two exchange
terms ΦCk and ΓCk responsible of the solute interactions between phases due to molecular fluxes at
the interface and to the solute exchange at the interface due to phase change respectively. Their
general expression is detailed in Table 1 and their ensemble average formulation can be found
respectively in (52) and (53).
Notice that, using the hypotheses of chemical equilibrium in the inter-dendritic liquid and
of local thermal equilibrium at the scale of the grain, the ensemble average of the inter-dendritic
liquid concentration is obviously equal to the local interface concentration, Cd = Cl∗ .
Let us compute the exchange terms for the solute balance equation in the solid. First, the
term ΦCs is zero since we neglect the solute diffusivity in the solid ( Ds = 0 ). Secondly, since the
temperature is uniform within the grain, the solid interface concentration Cs∗ is uniform too at the
scale of the grain. Using (53) one can easily show that:

ΓCs = Cs∗Γ s (82)

Now, for the extra-dendritic liquid the term ΓCl becomes:

ΓCl = ρ f n p ∫ wl − d Cl − d d Ω
Ωl −d (83)
= −Cl∗Γ g

where Cl − d was computed using the continuity of the solute field at the l − d interface:
( )
Cl − d = Cd −l = Cl∗ .
*
Using (54) the mean solute flux ΦCl can be written as ΦCl = S g jlC . With the help of
analytic results giving the solute field in the extra-dendritic liquid (see Appendix A), one easily
obtains:

ρl S g Dl
ΦCl =
δ l −d
(C ∗
l − Cl ) (84)

where δ l − d is the diffusion length characterizing the solute transfer in the extra-dendritic liquid at
the grain interface (see equation (A.18)). With (83) and (84) the solute balance equation in the
extra-dendritic liquid is now completely determined. However the conservative formulation in (81)
does not evidence the continuous enriching of the extra-dendritic liquid during the solidification

II-26
Submitted to Journal of Physics D: Applied Physics

process. In turn, using the mass balance in the extra-dendritic liquid, ∂ ε l ρ f ( ) ∂t = −Γ g one can
easily rewrite (81) as follows:

∂Cl ⎛ ρ f S g Dl ⎞
εl ρ f
∂t
(
= div ε l ρ f D f ∇Cl + ⎜
⎝ δ l −d
) − Γ g ⎟ Cl∗ − Cl ,
⎠
( ) (85)

Since δ l − d ≤ Dl w g (Appendix A) and Γ g = ρ f S g w g it is easy to notice that the last term on the
RHS of (85) is always positive. Since the macroscopic diffusion div ε l ρ f D f ∇Cl ( ) only
redistributes the solute in the ingot it now evident from (85) that the extra-dendritic liquid is always
enriching during the solidification process.
Let us compute now the exchange terms related to the inter-dendritic liquid. In computing
ΦCd and ΓCd one has to take into account the fact that phase d interacts with both s and l phase.
Indeed, each of the two exchange terms can be expressed as a sum of two terms corresponding to
interactions with s and l phase. Given the local thermal equilibrium and using (55) one can easily
obtain:

ΓCd = ΓCd − s + ΓCd −l ( = −C Γ∗

l s + Cl∗Γ g ) (86)

Similarly, the mean flux ΦCd becomes:

Φ Cd = Φ Cd − s + Φ Cd −l (87)

The second tem in RHS is no more than −ΦCl since jdC−l = − jlC− d . In turn, the exchange term
ΦCd− s , is difficult to model directly due to the complex d − s interface. However, one can express
ΦCl− s from the solute balance at the d − s interface:

(Γ C
s ) (
+ Φ Cs + ΓCd − s + Φ Cd − s = 0 ) (88)

Indeed, using (82) and (86) one easily obtains:

( )
Φ Cd− s = Cl∗ − Cs∗ Γ s (89)

We have managed therefore to close the solute balance equations (79)-(81). Notice that the
unknowns in these three equations are respectively Cs , Γ s and Cl . Thus, by fixing the
concentration in the inter-dendritic liquid one encompasses the need to model Cd obtainig in turn
an equation for the solidification rate Γ s . Indeed the inter-dendritic solute balance equation
becomes an equation for the solidification rate Γ s :

(
∂ ε d ρ f Cl∗ ) = div ε ρ f Sg D f
∂t
( dρf ) (
D f ∇Cl∗ + Cl∗ − Cs∗ Γ s − ) δl −d
(C
∗
l − Cl ) (90)
−Cl∗Γ s + Cl∗Γ g

Using the mass balance in the inter-dendritic liquid, equation (71), one can simplify further the
above equation:

II-27
Submitted to Journal of Physics D: Applied Physics

∂Cl∗ ρ f S g D f
(C ∗
l )
− Cs∗ Γ s = ε d ρ f
∂t
+
δ l −d
( )
Cl∗ − Cl − div ε d ρ f D f ∇Cl∗ ( ) (91)

The latter equation expresses as pointed out the solute balance in the inter-dendritic liquid. Since
the last term on the RHS of (91) can be usually neglected due to the large length scale involved in
the macroscopical diffusion process, the above balance equation can be understood as follows: the
solute rejection at the solid - inter-dendritic liquid interface (LHS) balance the solute accumulation
in the inter-dendritic liquid and the solute flux in the extra-dendritic liquid. It is important to notice
that equation (91) reduces to the well-known Scheil law when the extra-dendritic liquid reaches a
state of perfect solutal mixing, that is when Cl = Cl* . Indeed, using the solid mass balance equation
(66) equation (91)becomes:

∂Cl∗
(C ∗
l − Cs∗ ) ∂∂εt
s
= εd
∂t
(92)

that is equivalent to the Scheil solidification law. It is worthwhile to notice as well the important
role played by the solute flux in the extra-dendritic liquid: the positive solute flux at the grain
boundary intensifies the solidification process ( Γ s ). Hence to a non-equilibrium state of the
extra-dendritic liquid corresponds a more intense solidification process compared with a state of
state of perfect solutal mixing. Even more the solute flux at the grain boundary is strongly
influenced by the grain growth. Basically, a rapid grain growth will intensify the solute transfer at
the grain boundary and therefore the solidification process.

3.4.3. Summary of the model. In the following table a summary of all equations derived until now is
presented. Notice again the three phase approach of the solute balance: three solute balance
equations are written in order to account for the solid, extra- and inter-dendritic liquid. The later
equation is no more than the supplementary equation expressing the solidification rate.
It is worthwhile to notice as well the similarity between the present model formulation and
the volume averaged model of Wang and Beckermann [6]. The equivalence between the two
models is a direct consequence of the use of the local thermal equilibrium hypothesis. Indeed if the
local homogeneity hypothesis is considered, meaning that no local gradients of physical fields
(temperature, solute, grain fraction) are taken into account, then the ensemble average reproduces
the volume average technique [9]. This gives confidence in the present model formulation.
However the ensemble average technique and the cell model approximation to the ensemble
average are fundamentally different from the volume average approach. As we will notice in the
following if the local homogeneity hypothesis is no more valid, the use of the ensemble average
results into closure expression different from the ones obtained with the help of the volume average
technique. Moreover the volume average results will not be physically correct.

Table 3: The summary of the model

Main balance Solid: Liquid:
equations
Mass ∂ ( ε s ρ s ) ∂t = Γ s (
∂ εf ρf ) ∂t = −Γ s
Energy ( ) (
∂ ⎡ ε s ρ s cs + ε f ρ f c f T ⎤ ∂t = div ⎡ ε s λs + ε f λ f ∇T ⎤ + Γ s L
⎣ ⎦ ⎣ ⎦ )
Solute ∂ ( ε s ρ s Cs ) ∂t = ∇ ⋅ ( ε s ρ s Ds ∇Cs ) ∂ C
∗
εl ρ f
∂ t
l
(
= ∇ ⋅ ε l ρ f D f ∇Cl )
+Cs Γ s
(
+ ρ f S g Dl / δ l − d − Γ g Cl∗ − Cl )( )
Supplementary

II-28
Submitted to Journal of Physics D: Applied Physics

balance eq.
Inter-dendritic
mass balance
(
∂ εd ρ f ) ∂t = Γ g − Γ s
Grain balance ∂n ∂t = ne0δ (T − TNe ) or ∂n ∂t = nc0δ x − xcf ( )
Solidification rate
(C ∗
l ) ( )
− Cs∗ Γ s = ε d ρ f ∂Cl∗ ∂t + ρ f S g D f Cl∗ − Cl ( )δ l −d

(
− div ε d ρ f D f ∇Cl∗ )
Auxiliary
expressions
Constraints: εl + ε d + ε s = 1 , ε d + ε s = ε g
Interfacial
concentrations:
⎧⎪kC
Cs∗ = ⎨ l
∗
if Γs ≥ 0 (
Cl∗ = T − T f ) m
⎪⎩Cs if Γs < 0
Cell radius R1 = ( 4 / 3π n )
−1/ 3

Grain radius a = R1ε g1/ 3

Grain envelope S g = 4π a 2 n
interfacial area
Diffusion length see equation (A.18)
Grain growth rate D m ( k − 1) Cl∗
Γ g = ρ f S g wg , wg = l 2
( Pet )2
π Γ

4. The coexistence of columnar and equiaxed structures

We will focus in the following paragraph on the possible coexistence from a statistical point of
view between the equiaxed and the columnar grains. Let us consider an ingot containing at x both
columnar and equiaxed grains each of them having different grain densities (Figure 8). To model
the columnar/equiaxed coexistence one should be able to separate the influence of each structure at
x and at time t . Of course, what happens at ( x,t ) (i.e. the solute field) will be a sum of the
influence of all grains in the ingot, starting with the closest one to x and ending with farthest grain.
However, quantifying the influence at x of all grains in the ingot would be a very difficult task,
increasingly difficult with the increase of the accounted number of grains. Denoting with ψ the
physical field to be studied, we can suppose as a first approximation that only the closest grains to
x would influence significantly ψ ( x,t ) . Notice that this hypothesis is similar with the one
considered for the single grain distribution. Based on this assumption it would be interesting to
focus on the closest grain to x only. Consequently, for a case where both equiaxed and columnar
grains coexist in the ingot one should have an insight on the conditions that the closest grain to x is
a columnar or, on the contrary, an equiaxed grain. Indeed, since equiaxed and columnar structures
can have very different grain densities one expects that the influences of each of the two structures
at x be different as well.

II-29
Submitted to Journal of Physics D: Applied Physics

Figure 8: Coexistence between two different grain distributions ( ne < nc )

Let us now consider nc and ne the two corresponding grain densities for the columnar and
equiaxed structures. Similar to the previous case one can assume the existence of a "master"
distribution function family (Equation (15)) which completely describes statistically the multiphase
process. The only difference with the single grain distribution is that among the N particles in the
ingot one can find this time both columnar and equiaxed grains ( N = N c + N e ). Notice again that
locally, the columnar structures are approached with a spherical model. In this view the modeling
of coexistence between the columnar and equiaxed structures reduces to the coexistence between
two populations of spherical grains having different grain densities.

4.1. Probability distribution functions. The ensemble average

Making for the moment no distinction between columnar and equiaxed grains one can define the
unconditional density function for any sphere center being at z1 , namely f ( ) ( t , z1 ) . If we suppose
1

that the grains are uniformly distributed in space the unconditional probability f ( ) ( t , z1 ) dz1 will
1

represent the probability that at time t we find in the ingot one particle (regardless of its nature)
within dz1 of z1 . Moreover f ( ) ( t , z1 ) is equal to:
1

f ( ) ( t , z1 ) = nc + ne
1
(93)

where nc and ne are the corresponding grain densities. These latter values represent as well the
unconditional density functions for any columnar/equiaxed grain center being within dz1 of z1 ,
that is
:
f c( ) ( t , z1 ) = nc
1
(94)
f e( ) ( t , z1 ) = ne
1

Using a similar reasoning as in paragraph 3.1., the nearest-neighbor distribution function family
(
f N , t , x; z1 , z 2 ,… , z N ) can also be defined (see equation (29)). As already discussed, we will
focus in the following on the closest particle to x only. Thus it is useful to define the unconditional
density function for any sphere, regardless of its nature, being closest to x and centered within
(1)
d z1 of z1 . This is no more than f ( t, x; z1 ) defined already in equation (30). Therefore, the
ensemble average of a generic field ψ ( x,t ) can be written as follows:

II-30
Submitted to Journal of Physics D: Applied Physics

(1)
ψ ( x, t ) = ∫ f ( t, x; z1 )ψ (1) ( t , x; z1 ) d z1 (95)

(1)
where ψ ( t , x; z1 ) is the conditional average of ψ , averaged on the condition that the closest
grain to x , regardless of its nature, is centered within d z1 of z1 . Notice that until now we didn't
make a difference between the two coexisting structures. We have treated the grains regardless of
their type. To separate the effects of columnar or equiaxed structures at x it would be extremely
useful to know whether or not the closest particle to x is a columnar or an equiaxed particle. This
(1)
requires an detailed insight on the distribution function f ( t, x; z1 ) .
Let us consider the two following events: Ec the event that the closest particle to x is
columnar and is centered within d z1 of z1 and Ee the event that the closest particle to x is
equiaxed and is centered within d z1 of z1 . Obviously P ( Ec ∩ Ee ) = ∅ and

(1)
f ( t, x; z1 ) d z1 = P ( Ec ∪ Ee ) (96)
= P ( Ec ) + P ( Ee )

Thus the probability of having the closest particle to x centered within d z1 of z1 is the probability
of having either a columnar or an equiaxed particle closest to x and centered within d z1 of z1 .
The two probabilities in the RHS of (96) can be computed with the help of the two corresponding
(1)
( )
density functions f c, e t , x; z1 that the closest particle to x is columnar/equiaxed and is centered
within d z1 of z1 . One has therefore:

(1)
f ( t , x; z1 ) d z1 = f c(1) ( t, x; z1 ) d z1 + f e(1) ( t, x; z1 ) d z1 (97)
In this context, the ensemble average written in (95), transforms into:

(1) (1)
ψ ( x, t ) = ∫ f c t , x; z1 ψ ( ) ( t , x; z1 ) d z1 + ∫ f (e1) ( t, x; z1 )ψ (1) (t , x; z1 ) d z1 (98)

This result is very important since it states for the fact that the ensemble θ of all possible
realizations of the multiphase flow (see equation (1)) can be approached as the union of two sub-
ensembles: the sub-ensemble θc which corresponds to the ensemble of realizations such as the
closest particle to x , at time t, is a columnar one and the sub-ensemble θ e of realizations such as
the closest particle to x , at time t, is an equiaxed one. Thus:

θ = θc ∪ θe (99)

Notice that integration in (98) is made with respect to the variable z1 , that is over the possible
positions of the closest grain to x . Thus, equation (98) express the ensemble average of ψ as a
sum of two averages corresponding to situations where the closest particle to x is respectively
columnar and equiaxed. To separate the effects of columnar and equiaxed structures at x , it is now
useful to introduce the following two discrete functions (phase functions):

⎧1 if the nearest particle to x is a k type

X k (t , x) = ⎨ k = {c, e} (100)
⎩0 if the nearest particle to x is a k type

II-31
Submitted to Journal of Physics D: Applied Physics

With the help of these two functions two new phases are defined: the columnar (c) and the
equiaxed (e) phase. Hence, corresponding average variables just as those defined in Table 1 can be
attached to the newly defined phases. For example, using (98), the columnar fraction becomes

(1) (1)
( )
ε c = ∫ X c (t , x) f c t , x; z1 d z1 + ∫ X c (t , x) f e t , x; z1 d z1 ( ) (101)

Notice that the second term in the RHS of (101) is identically zero since it expresses an integral
over the ensemble of realizations θ e such as the closest particle to x is equiaxed. The phase field
X c (t , x) is 0 in this case and ε c finally becomes

(1)
(
ε c = ∫ f c t , x; z1 d z1 ) (102)

Likewise, the equiaxed fraction can be expressed as:

(1)
(
ε e = ∫ f e t , x; z1 d z1 ) (103)

and the columnar/equiaxed ensemble average of a generic field ψ ( x,t ) becomes:

⎡ (1) t , x; z + f (1) t , x; z ⎤ψ (1) t , x; z d z

ψ c , e ( x, t ) =
1
ε c, e ∫ X c, e ⎣⎢ f c ( 1 e ) 1
⎦⎥
( ) 1 1 ( )
(104)
(1) (1)
∫ f c, e ( t , x; z1 )ψ ( t, x; z1 ) d z1
1
=
ε c, e

(1)
Notice that in order to compute any average field the two nearest distributions f c , e t , z1 ( )
(1)
and the conditional average ψ ( t , x; z1 ) have to be approached. Analytic expressions for
(1)
( )
f c , e t , z1 are proposed in the following paragraph. Then, similar to the closure approach used for
(1)
the case involving a single grain distribution (see paragraph 3), the conditional average ψ will
be resolved analytically using the hypotheses that a field ψ ( x,t ) is solely influenced by the closest
grain to x .

4.2. The equiaxed/columnar nearest neighbor distribution functions

In the following the coexistence between two grain populations is analyzed. More precisely, the
coupling between a grain population having a finite grain diameter and a second grain population
having infinitely small dimension. This type of coupling can be frequently encountered during
solidification problems especially in processes involving the so-called columnar-to-equiaxed
transition (CET) phenomena. Indeed, CET phenomena are a direct consequence of the coupling
between an advancing columnar front and a developed equiaxed structure, thus between an
infinitely small grain distribution (columnar dendrite tips) and a finite sized grain distribution (the
equiaxed structure).
In this light, the case of a columnar front "penetration" into a developed equiaxed zone will
be analyzed next. Note that the developped theory will not be limited to this particular
phenomenon. Indeed, the results can be straightforward applied to different phenomena involving
the coexistence between two grain distributions, i.e. equiaxed nucleation within a columnar zone or
equiaxed grain advection into an already existing columnar zone.

II-32
Submitted to Journal of Physics D: Applied Physics

Figure 9: A typical directional solidification and the subsequent columnar-equiaxed interaction

Consider the classical case of a directional solidification involving an axial thermal

gradient G and puling a velocity V p (Figure 9). It can be assumed that equiaxed grains nucleate at
the nucleation temperature Tne within the undercooled zone existing ahead of the advancing
columnar tips (between Tliq ( C0 ) and the columnar front temperature T pc ). Therefore under certain
conditions it is likely that columnar tips would encounter in their way a developed equiaxed zone.
Let us consider at x an homogeneous equiaxed grain distribution having a grain density
ne . Let us consider as well that the equiaxed grain are spherical, have a finite diameter ae and that
at moment t the columnar front arrives at x (the exact evolution of the columnar front is assumed
to be known). Note that columnar grains are approached with a spherical model too and their
equivalent grain density nc0 can be directly related to the primary arm spacing λ1 (equation (13)).
Obviously one expects that the statistics characterizing the coexistence state (Figure 10) will be
fundamentally different from the one characterizing separately each structure at x prior to the
columnar “penetration” into the equiaxed zone.

Figure 10: The two grain distributions just before the coexistence state

As pointed out in the preceding section, following the case where the closest particle to x
is either columnar or equiaxed, the ensemble average ψ ( x,t ) can be written, equation (98), as a
sum of the columnar and the equiaxed average. The two averages will measure the contribution of

II-33
Submitted to Journal of Physics D: Applied Physics

each of the two structures to the “total” ensemble average ψ ( x, t ) . To be able to compute these two
(1)
( )
averages the distributions f c / e t , x; z1 , that a c/e grain is centered within d z1 of z1 and is the
closest to x among all particles in the ingot, needs to be determined.

(1)
4.2.1. Equiaxed nearest neighbor distribution function. To compute f e ( t, x; z1 ) it is useful to
consider first the Ae event that the closest equiaxed grain to x, among the equiaxed grains, is
centered within d z1 of z1 . The probability of Ae event can be derived from the distribution
(1)
probability function f e ( t, x; z1 ) that the closest equiaxed particle to x among the equiaxed
particles is centered within d z1 of z1 . It can be proved (see Appendix B) that:

(1)
( t, x; z1 ) = ne ⋅ exp ⎡⎢⎣− 43 π ne ( x − z1 ) ⎤⎥⎦
3
fe (105)

However, after the columnar “penetration” into the equiaxed zone, besides the closest equiaxed
grain at z1 , one may find as well newly “germinated” columnar tips between x and z1 (see Figure
11a).

Figure 11: Computing the nearest neighbor distribution for the equiaxed (a) and columnar (b)
distributions

(1)
Hence, to compute f e , one has to quantify as well the probability that a columnar tip centered
between x and z1 is closer to x than the existing equiaxed grain at z1 . Due to the finite dimension
of the equiaxed grains ( ae > 0 ) two cases have to be analyzed. First, if x − z1 ≤ ae the equiaxed
grain centered at z1 “touches” the point x. Since the columnar tips penetrate into an developed
equiaxed grain and can only “nucleate” within the liquid space between the equiaxed grains (i.e.
the equiaxed extra-dendritic liquid) it is obvious that there is no possibility that a columnar grain
centered between x and z1 be closer that the equiaxed grain at z1 . Hence, the probability that an
equiaxed grain centered at z1 is the closest to x among all particles in the ingot can be identified in
this particular case ( x − z1 ≤ ae ) to P ( Ae ) . One has:

II-34
Submitted to Journal of Physics D: Applied Physics

(1)
( t, x; z1 ) d z1 = f (e ) ( t, x; z1 ) z1 ,
1
f e x − z1 ≤ ae (106)

* *
In contrast, if x − z1 > ae any columnar tips centered at z1 such as x − z1 < x − z1 − ae (see

Figure Figure 11a) will be obviously closer to x than the equiaxed grain at z1 . It is important to
* *
notice that columnar grains centered at z1 such as x − z1 − ae ≤ x − z1 ≤ x − z1 are not

considered to be closer to x than the equiaxed grain at z1 . Indeed, due to its finite dimension, the
equiaxed grain at z1 will influence preponderantly the field ψ ( x, t ) (i.e. solutal field, velocity,
*
pressure, etc.) and not the columnar tip at z1 . Denoting now with Bc the event that there is no
columnar tip centered within the spherical cell of radius x − z1 − ae and centered at x,
(1)
f e ( t, x; z1 ) becomes
(1)
f e ( t, x; z1 ) d z1 = P ( Ae ∩ Bc ) (107)
= P ( Bc Ae ) P ( Ae )

The conditional probability P ( Bc Ae ) can be computed with the help of the conditional probability
(1)
f e ⎛⎜ t , x; z1 Ae ⎞⎟ d z1 that a columnar tip centered at z1 is the closest to x among the columnar
* * *

⎝ ⎠
particles, given the event Ae that the closest equiaxed grain to x is at z1 . It is shown in the
Appendix B that:

(1) ⎡ 4 3⎤
f c ⎛⎜ t , x; z1 Ae ⎞⎟ = nc0 ⋅ exp ⎢ − π nc0 ⎛⎜ x − z1 ⎞⎟ ⎥ , x − z1 < x − z1 − ae
* * *
(108)
⎝ ⎠ ⎢⎣ 3 ⎝ ⎠ ⎥⎦

Now, it is easy to see that

(1)
P ( Bc Ae ) = 1 − f c ⎛⎜ t , x; z1 Ae ⎞⎟ d z1
* *

*
∫ ⎝ ⎠
x − z1 < x − z1 − ae
(109)
⎡ 4
( )
3⎤
= exp ⎢ − π nc0 x − z1 − ae ⎥
⎣ 3 ⎦

(1)
Thus, the distribution function f e becomes

⎧ − 4 π ne ( x − z1 )3
⎪ 3 x − z1 ≤ ae
(1) ⎪ne e if
(
f e t , x; z1 = ⎨ ) (110)
⎪ − 3 π nc0 ( x − z1 − ae ) − 3 π ne ( x − z1 )
4 3 4 3

⎪⎩ne e e if x − z1 > ae

4.2.2. Columnar nearest neighbor distribution function. Notice again that due to the finite
dimension of the equiaxed particles a columnar particle positioned at z1 can be the closest particle

II-35
Submitted to Journal of Physics D: Applied Physics

to x only if there is no equiaxed particle within a range x − z1 + ae around the point x (Figure
11b). By considering the two following events, Ac the event that a columnar tip positioned at z1 is
the closest columnar particle to x and Be the event that there is no equiaxed particle within a range
x − z1 + ae around the point x, one has

(1)
f c ( t, x; z1 ) d z1 = P ( Ac ∩ Be ) (111)
= P ( Ac Be ) P ( Be )

(1)
The Be probability can be easily computed with the help of the f e ( t, x; z1 ) distribution that there
the closest equiaxed particle to x is centered within d z1 of z1 . Indeed, it is clear that

P ( Be ) = 1 −
'
∫
x − z1 < x − z1 + ae
fe
(1)
(t, x; z ) d z
'
1
'
1

(112)
⎡ 4
( )
3⎤
= exp ⎢ − π ne x − z1 + ae ⎥
⎣ 3 ⎦

In turn the probability P ( Ac Be ) can be derived from the conditional probability f c

(1)
(t, x; z B )
1 e

that the closest columnar particle to x is at z1 , given no equiaxed particle within a range
x − z1 + ae around the point x (see Figure 11b). Similar to the second case analysed in the
Appendix B one can obtain

(1)
(t, x; z B ) = n ⎡ 4
( ) ⎤⎥⎦
3
0
fc 1 e c ⋅ exp ⎢ − π nc0 x − z1 (113)
⎣ 3

Hence, the columnar nearest neighbor distribution becomes

(1)
( t, x; z1 ) = nc0 exp ⎛⎜⎝ − 43 π ne ( x − z1 + ae ) ( ) ⎞⎟⎠
3⎞ ⎛ 4 0 3
f c ⎟ exp ⎜ − π nc x − z1 (114)
⎠ ⎝ 3

4.2.3. The double mechanical blockage effect. In Figure 12 the two nearest distributions
(1)
( )
f c ,e t , x; z1 , characterizing the coexistence state along with the f c
(1)
(t, x; z B ) and
1 e
(1)
fe ( t, x; z1 )
distributions characterizing the equiaxed and columnar structures just before their coexistence
(Figure 10) are presented for ae = 810 µ m , ne = 108 grains / m3 and nc0 = 1010 grains / m3 . As a
general trend the coexistence state is characterized by a decrease in the probabilities
(1)
( )
f c,e t , x; z1 d z1 that the closest grain to x is columnar/equiaxed and is centered at z1 .

II-36
Submitted to Journal of Physics D: Applied Physics

Figure 12: The nearest columnar (a) and equiaxed (b) neighbor distribution before (dashed line)
and after (full line) the c/e coexistence ( ae = 810 µ m , ne = 108 grains/m3 and nc = 1010 grains/m3)

Using the two analytical distributions in (110) and (114) the two columnar and equiaxed
fractions as defined in (102) and (103) can now be computed

⎛ 4 ⎞
ε c = exp ⎜ − π ne ae3 ⎟ − I
⎝ 3 ⎠
(115)
⎛ 4 ⎞
ε e = 1 − exp ⎜ − π ne ae3 ⎟ + I
⎝ 3 ⎠
where r = x − z1 and
∞
⎡ 4 3⎤ ⎛ 4 ⎞
I = ∫ ne exp ⎢ − π ne ( r + ae ) ⎥ exp ⎜ − π nc0 r 3 ⎟ 4π ( r + ae ) dr
2
(116)
0 ⎣ 3 ⎦ ⎝ 3 ⎠

is a positive infinite integral having no explicit expression. As expected ε e and ε c sum to one
validating a physically important constraint. Indeed, the probability to find the closest grain to x
(regardless of its nature) centered between x and ∞ must be 1. Note that ⎡1 − exp −4 / 3π ne ae3 ⎤
⎣ ⎦ ( )
is no more than the equiaxed grain fraction ε ge (0.2 for the case in Figure 12). By defining the
characteristic function X ge (equal to 1 if the point x is within an equiaxed grain and 0 otherwise)
and using (98) the equiaxed grain fraction becomes indeed

(
ε ge ≡ X ge = ) ∫ f
(1)
e ( t , x; z1 ) d z1
x − z1 ≤ ae (117)
(
= 1 − exp −4 / 3π ne ae 3
)
Since I > 0 one has thus ε e > ε ge and ε c < 1 − ε ge . These two inequalities express another
important constraint of the coexistence state: the newly germinated columnar phase cannot occupy
a space larger than the available space for its germination existing prior to columnar “nucleation”,
namely the equiaxed extra-dendritic liquid 1 − ε ge . Indeed, due to the finite dimension of the

II-37
Submitted to Journal of Physics D: Applied Physics

equiaxed grains, the columnar tips can only “nucleate” within the equiaxed extra-dendritic liquid
(see Figure 10) and the subsequent columnar fraction will be therefore always smaller than 1 − ε ge .
This particular constraint has another important consequence on the columnar structure. Note first
(1)
that f c ( t , x; x ) is equal as well to the unconditional density function for any columnar tip being at

x, f c( ) ( t , x ) . Remember that, given the hypothesis of a local homogeneous grain distribution,

1

f c( ) ( t , x ) equals as well the local columnar grain density nc , equation (23). Hence, one has
1

(1)
f c( ) ( t , x ) ( ≡ nc ) = f c ( t , x; x )
1
(118)
(
= nc0 1 − ε ge )
that is, the columnar grain density characterizing the coexistence ( nc ) will be lower than nc0 , the
columnar grain density valid just before the columnar “penetration” into the equiaxed zone. Indeed,
due to the finite dimension of equiaxed grains the advancing column tips (Figure 13a) would be
partially stopped by the equiaxed particles. Therefore, only a fraction (1 − ε ge ) of the total number
of columnar tips per unit volume would be able to “penetrate” the equiaxed zone. The presence of
developed equiaxed grains at x will determine thus a rarefaction of columnar tips and equation
(118) mathematically expresses this blocking effect. We will hereafter refer to this effect as to the
first mechanical blockage effect of the coexistence.

Figure 13: The first (a) and the second (b) mechanical blockage: the full line circles represent the
two cells characterizing the mixed c/e zone and the dashed line circles are the cells characterizing
the two structures (c/e) separately, that is just before the coexistence state

Apart from the decrease of the columnar grain density (Equation (118)), one can see from
(1)
Figure 12 that both f c / e distributions become negligible for a smaller r = x − z1 than their
correspondant single grain distributions. Indeed, looking closely to their particular exponential
variation (cf. equations (110) and (114)), one can easily identify a characteristic distance from x,
above which the two distributions decrease rapidly towards 0. Each exponential term of type
(
exp −4 / 3π nr 3 ) remains of order of 1 for r ≤ ( 4 / 3π n )
−1/ 3
and decrease rapidly towards 0 for
−1/ 3
r > ( 4 / 3π n ) . Therefore one can easily identify the two characteristic distances from x as

II-38
Submitted to Journal of Physics D: Applied Physics

R1' c = min ( R1c , R1e − ae )

(119)
R1' e = min ( R1c + ae , R1e )
where
( )
−1/ 3 −1/ 3
R1c = 4 / 3π nc0 , R1e = ( 4 / 3π ne ) (120)

Remember that R1c / e represents half of the average distance between two adjacent c/e grains
before their coexistence (Figure 10) and quantify as well the cell dimension used in the modeling of
the ensemble average for a single grain distribution, e.g., equation (42). Likewise, the distance
R1' c / e would represent the average distance between two adjacent c/e particles characterizing the
coexistence state. In addition R1' c + R1' e would be the average distance between two different
adjacent grains (a columnar and an equiaxed grain, see Figure 13b). Moreover, from the equation
(119), one has

R1' c < R1c , R1' e < R1e (121)

meaning that the coexistence state is characterized by smaller average distances between particles
compared with the single grain distribution state valid just before the coexistence. Consequently,
the two spherical cells describing at x the statistics of the two c/e distributions will be smaller than
the ones valid just before the coexistence (Figure 13b). This trend is physically plausible since
subsequent to the columnar “nucleation an increase number of particles will have to share the same
space. We will hereafter refer to this effect as to the second mechanical blockage effect of the
coexistence. Notice that this effect is of primary importance for the evolution of the coexisting c/e
grains since it’s directly linked to the diffusion lengths characterizing the solute fluxes in the c/e
'
extra-dendritic liquid. Indeed the two diffusion lengths depend strongly on R1e − ae and R1c' ,
respectively. Note as well that

R1' e − ae = R1' c (122)

meaning that the length scales of the coexisting columnar/equiaxed extra-dendritic liquid are equal.
Therefore, the initial equiaxed extra-dendritic liquid, 1 − ε ge , is shared by the two structures such as
the length scales characterizing the two coexisting extra-dendritic liquids have the same order of
magnitude (Figure 13b).
An interesting limit case is the coexistence between two infinite small grain distributions,
that is ae → 0 . For this particular case, one obtains
−1/ 3
ε c / e = nc / e / ( nc + ne ) , ( ) ⎡4 ⎤
R1' c = R1' e = ⎢ π ( nc + ne ) ⎥
⎣3 ⎦
(123)

Note that R1' c = R1' e , meaning that the average distance between two adjacent grains (equiaxed
and/or columnar) is equal. Indeed, due to the infinite small dimension of the two structures and due
to their homogeneous distribution in space, the mixing between columnar and equiaxed grains
would be equivalent from a statistical point of view to a homogenous grain distribution having the
density ( nc + ne ) for which no distinction is made between equiaxed and columnar nuclei. Note
however that the equality in (123) does not mean an equal filling of space with columnar and
equiaxed nuclei. The average distance between two particles is the same but each type of grain will
fill a given volume according to its grain density nc / e . If, for example, nc ne one will find in the
ingot many more columnar nuclei than equiaxed ones. This unbalance is well reflected by the

II-39
Submitted to Journal of Physics D: Applied Physics

fractions (probabilities) ε c / e ; one has ε c ε e in this case, meaning that it is much more likely that
the closest particle to x be a columnar particle than an equiaxed one.

4.2.4. The cell model. The use of the exact formulations (110) and (114) to compute an ensemble
averages, equation (98), would be extremely complex. Moreover, the c/e fractions ε c and ε e , key
parameters within an average model, have no explicit form for ae > 0 . Therefore, a simpler form
(1)
for f c / e , resulting in an explicit formulation for ε c / e , would be of great interest.
Note first that a coexistence state between equiaxed and columnar grains is always
succeeding a state where equiaxed and columnar structures grains evolve separately. Thus, one can
logically assume that this previous state is modeled using the cell model approach developed for a
(1)
single grain distribution , throughout paragraph 3 In this respect, the exact forms for f e ( t, x; z1 ) ,
(1) (1)
( )
f c ⎛⎜ t , x; z1 Ae ⎞⎟ and f c t , x; z1 Be , equations (105), (108) and (113), could be replaced with
⎝
*

⎠
their corresponding cell model approximations

⎧ne if x − z1 ≤ R1e
(1)
fe ( t, x; z1 ) = ⎪⎨ (124)
⎪⎩0 if x − z1 > R1e
and
⎧nc0 if x − z1 ≤ R1c
fc
(1)
( )
⎛
⎝
(1)
( ⎞ ⎪
t , x; z1 Ae ⎜ = f c t , x; z1 Be ⎟ = ⎨
⎠ ⎪0
) if x − z1 > R1c
(125)
⎩

where R1c and R1e are already defined in (120). Using now these approximations in (106), (109)
and (113) the approximate nearest-neighbor distributions, characterizing the coexistence state,
become respectively

⎧ne if z1 − x ≤ ae
⎪
⎪⎪ ⎡ 4
(1)
( ) ( ) n ⎤⎥⎦ (
3
f e t , x; z1 = ⎨ne ⎢1 − π z1 − x − ae 0
if z1 − x ∈ ae ; R1' e ⎤ (126)
c ⎦
⎪ ⎣ 3
⎪0 if z1 − x > R1' e
⎪⎩
and
⎧ 0⎡ 4
( ) n ⎤⎥⎦
3
(1) ⎪nc ⎢1 − π z1 − x + ae if z1 − x ≤ R1' c
( )
e
f c t , x; z1 = ⎨ ⎣ 3 (127)
⎪0 if z1 − x > R1' c
⎩

where R1c' and R1e' are already defined in (119). A graphic comparison between the exact
distributions and their corresponding approximation is given in Figure 14 for ae = 810 µ m , ne = 108
grain/m3 and nc = 1010 grain/m3. As one can see, (126) and (127) approximate reasonably well the
exact distributions in (110) and (114). Notice the difference between the exact and the approximate
(1)
columnar distributions at r = 0 , that is f c ( t , x; x ) ( ≡ nc ) . This is a normal consequence of the
difference between the exact and the approximate equiaxed grain fraction ε ge . Indeed, in contrast
with the exact approach, equation (117), the cell approximation to ε ge is

II-40
Submitted to Journal of Physics D: Applied Physics

ε ge = 4 / 3π ne ae3 (128)

Note as well that the approximate distributions integrate the same lengths scales R1c' and R1e'
which characterized the exact distributions.

Figure 14: The exact (dash), the approximate (full) and the cell model approximation form (dash-
(1) (1)
dot) for f c (a) and f e (b) distributions, ( ae = 810 µ m ,ne = 108 grain/m3 and nc = 1010 grain/m3)

Using (131) the c/e fraction can be now computed explicitly. One obtains

⎧ ⎡
( ) ⎤⎥
1
2

⎪ 1 ⎢ 10 n + 36 ( n ) 3 n ε
0 3

( )
e e c ge
⎪1 − ε ge − ⎢ ⎥ if R1c ≤ R1e 1 − ε ge1/ 3
20nc0 ⎢
( )
2

εc ⎪ 1
+45 ( ne ) 3 nc0ε ge 3 ⎥
⎪ ⎣ ⎦
= (129)
(1 − ε e ) ⎨⎪
0 ⎡( ε ) − 20ε + 45 ( ε ) 3 ⎤
2
2
n
⎪ c ⎢
⎪ 20n ⎢
ge ge ge
⎥
⎥ if (
R1c > R1e 1 − ε ge1/ 3 )
( ge ) 3 + 10
1

⎪⎩ e −36 ε
⎢⎣ ⎥⎦

where ε ge is the equiaxed grain fraction computed with the cell model approximation, equation
(128). A plot of (129) is given in Figure 15. Note that, in contrast with the discrete distributions in
(126) and (127), the resulting c/e fraction is a continuous function. Notice as well that the
constraint ε c < 1 − ε ge is verified too.

II-41
Submitted to Journal of Physics D: Applied Physics

Figure 15: The columnar fraction ε c with respect to nc0 and ε ge , given ne = 108 grain/m3

(1)
However, we are looking for an even more simpler form for f c ,e , that is for their “top hat”
approximation (cell approximation)

⎧nc / e if x − z1 ≤ R1*c / e
(1) ⎪
f c/e ( )
t , x; z1 = ⎨ , (130)
⎪⎩0 if x − z1 > R1*c / e

( )
where nc = nc0 1 − ε ge . The two length scales R1*c / e can be determined by ensuring that

(1)
∫ f c, e ( t , x; z1 ) d z1 = ε c,e , (131)

One easily obtains for R1*c,e

−1/ 3
⎛4 ⎞
R1*c ,e = ⎜ π nc ,e ⎟ ε c,e1/ 3 (132)
⎝ 3 ⎠
(1) (1)
The cell approximations to f c and f e are now fully determined from (130) and (132).
A plot of these two “top-hat” distributions is shown in Figure 14 along with their exact and
approximated forms. It is important to see that, even if equations in (130) represent a double
approximation for the exact distributions (110) and (114), they integrate in a much simpler form all
the constraints and features characterizing the exact distributions. First, ε c + ε e = 1 . Then the
inequality ε c < 1 − ε ge is verified as well. Moreover, the length scales, R1*c ,e , characterizing the c/e
cells are of the same order of magnitude with R1' c ,e , previously determined for the exact
distributions. Indeed it can be verified that

0.63 < R1*c / R1' c < 1

(133)
0.79 < R1*e / R1' e < 1

Finally, the length scales of the c/e extradendritic liquid characterizing the coexistence are of the
same order of magnitude. It can be verified that

II-42
Submitted to Journal of Physics D: Applied Physics

(
0.63 < R1*c / R1*e − ae < 1.34 ) (134)

(1)
With the help of the cell approximations to f c ,e one is now able to easily compute c/e
averages using the general equation (98). For a generic field ψ one obtains

∫ f c,e ( t , x; z1 )ψ
(1) (1)
ψ c , e ( x, t ) =
1
ε c ,e
( t, x; z1 ) d z1
nc,e (135)
(1)
=
ε c ,e ∫ ψ ( t, x; z1 ) d z1
x − z1 ≤ R1*c , e

It is worthwhile to notice that the use of the cell model transforms the infinite integrals in (98) into
definite ones. The equiaxed and columnar averages reduce to an integral over all possibile positions
( z1 ) of the closest columnar/equiaxed grain to x within the two spherical cells of radius R1*c ,e and
centered at x (Figure 16). This result is of great importance since it quantifies the two sub-
ensembles θ c,e , defined in (99) (ensembles of realizations such as the closest particle to x , at time
t, be a columnar/equiaxed one). Indeed, in the light of the cell model θ c become the ensemble of
realizations such as the closest particle to x is columnar and is centered within the cell of radius
*
R1c centered at x. Likewise, θe is the ensemble of realizations such as the closest particle to x is
*
equiaxed and is centered within the sphere of radius R1e centered at x.

Figure 16: The cell model approximation to the columnar (a) and the equiaxed (b) ensemble
average

Using the cell model, one can easily express as well e/c interfacial averages like Ske / c ,
Φψ ke / c or Γψ ke / c . Their mathematical formulation will be very similar to the one used for the
case of single grain distribution, equations (52) and (53), with the difference that all integrations
are made over all possible positions of the equiaxed/columnar particle within the cell of radius
R1*e / c such as the point x is on the e/c solid interface. For exemple, the generic exchange term Φψ k
written for the equiaxed strucure would write as follows

II-43
Submitted to Journal of Physics D: Applied Physics

Φψ ke = ne ∫ ( − jψ n ) ⋅ ⎡⎣−n δ ( x − x (t ))⎤⎦ d z
ke ke ke i 1
x − z1 ≤ R1*e
(136)
ψ ⎛ ψ ⎞
*
= ne ∫ jke d Ω ⎜ = Ske jke ⎟
⎝ ⎠
Ω( ae )

that is, an integral over all posible positions ( z1 ) of the equiaxed grain within the equiaxed cell of
*
radius R1e and such as the point x is on the k phase interface. Hence, using the two nearest-
neighbor distributions two spherical cells have been delimited around x, characterizing the statistics
of the coexisting c/e structures. Moreover the cell of radius R1*c ,e enveloping a c/e grain will delimit
locally the “zone of influence” of that grain. Any point outside this sphere will be under the
influence of a neighbor particle.

4.3. Model closure

The physical assumptions valid at the micro-scale remain exactly the same as those defined for a
single grain distribution. Note however that the scale separation between the inter- and the extra-
dendritic liquid divide now the total liquid into four sub-phases: the equiaxed inter- (de) and extra-
(le) dendritic liquid and the columnar inter- (dc) and extra- (lc) dendritic liquid. Therefore, the
solute field around an equiaxed or a columnar grain has to be reviewed since one has to take into
account the fact that the solute diffusion in the e/c extra-dendritic liquid is limited within the cell of
radius R1*e / c enveloping each e/c grain. However using the hypotheses of the local thermal
equilibrium at the scale of the grain the solute field in the e/c extra-dendritic liquid will be a
function of r only. Moreover since

R1*c R1*e − ae (137)

the e/c cell boundary will play the role of a symmetry surface between two adjacent grains. Hence,
locally, the solute flux between two adjacent grains is zero, namely

⎛ ∂Cle / c ⎞
⎜ ∂r ⎟ * = 0 (138)
⎝ ⎠ R1e / c

In this respect the solute diffusion field already determined for the case of a single grain
distribution (Appendix A) can be used for the two coexisting structures too. In (A.18) one needs
only to replace R1 with R1*e / c and a with ae,c in order to obtain the corresponding e/c solute
fields.

4.3.1. Final averaged equations. Given the separation between the inter- and extra-dendritic liquid,
for the case of equiaxe/columnar coexistence six distinct phases can be identified: the c/e
(columnar/equiaxed) solid (sc, se), the c/e inter-dendritic liquid (dc, de) and finally the c/e extra-
dendritic liquid (lc, le). Basically the problem is a three-phase one: se + sc + total liquid (f).
However, the present model needs to quantify for the solute balance within the c/e extra-dendritic
liquid and therefore one has to account as well for the virtual separation of the total liquid in four
sub-phases ( f = de + le + dc + lc ).
Before writing the final averaged equation first one has to define the characteristic functions X k ,
equation (4), for each of the six phases, k ∈ {se, de, le, sc, dc, lc} . Then using the averaged variables
defined in Table 1 the generic balance equation, (8) can be particularized to obtain respectively the
mass, energy and solute balance for each phase.

II-44
Submitted to Journal of Physics D: Applied Physics

Grain balance equations

The equiaxed grain balance remains unchanged with the one determined for the single grain
distribution case

∂ne / ∂t + div ( v se ne ) = ne0δ (T − TNe ) (139)

where δ is the Dirac delta function, T the local temperature, TNe the equiaxed nucleation
temperature and ne0 is the germination nuclei density.
As pointed out, in contrast with the equiaxed germination, the “columnar nucleation” can
not be treated as a local event since the positions of the columnar front at a precise moment cannot
be linked to local parameters only. In the absence of a precise local criterion defining the columnar
front, one has to track exactly the position of the columnar front at each moment. In this light, the
columnar nucleation can be approached with a local instantaneous germination

∂nc / ∂t = nc' δ ⎡⎣ x − xcf (t ) ⎤⎦ (140)

where again δ is the Dirac delta function, xcf (t ) the known columnar front position at time t and
nc' the columnar germination nuclei density. It was proven that the columnar tips penetrating into a
developed equiaxed structure will be partially blocked by the equiaxed grains (first mechanical
( )
blockage), equation (118). Hence, nc' = nc0 1 − ε ge where ε ge is the local equiaxed grain fraction

( ε se + ε de ) and nc0 is the local columnar grain density valid just before the columnar “nucleation”
within the equiaxed zone. In turn, nc0 is linked, equation (13), to the local primary arm spacing λ1 .
Note that with the help of nc0 , ne , ε ge , the columnar and equiaxed fractions as defined in (129)
* *
and the two cell radii R1c , R1e can now be computed. The cell model to the ensemble average can
therefore be applied as detailed in the precedent section, equations (135), (136).

Mass balance
The mass balance for the two solid phases (sc and se) and for the total liquid (f) become
respectively

∂ ( ε se ρ s ) / ∂t = Γ se , ∂ ( ε sc ρ s ) / ∂t = Γ sc
(141)
( )
∂ ε f ρ f / ∂t = − ( Γ se + Γ sc )

where Γ se and Γ sc are the mass exchange rates between the total liquid and the equiaxed and
columnar solid respectively. Both Γ se and Γ sc can be computed from the solute balance in the e/c
inter-dendritic liquid. To account for the inter/extra- dendritic separation, the mass balances for the
e/c inter-dendritic liquid have to be considered too

( )
∂ ε dk ρ f / ∂t = Γ gk − Γ sk , k = {e, c} (142)

where Γ gk is the mass exchange rate at the dk − lk interface as a result of the e/c grain growth.
Using the cell model approximation to the ensemble average one easily obtains

Γ gk = ρ f S gk w gk , k = {e, c} (143)

II-45
Submitted to Journal of Physics D: Applied Physics

(
where S gk = nk 4π ak 2 ) is the grain interfacial area of the e/c grain envelope, a k is the radius of

the e/c grains and w gk is the mean velocity of the e/c grain envelope. Similar to the single grain
distribution case, w gk can be computed using a standard tip kinetic law, equation (59), by
replacing Ω with the corresponding local e/c undercooling

( )
Ωe / c = Cl∗ − Cle / c / Cl∗ / (1 − k ) (144)

Energy equation
Similar to the single grain distribution and given the local thermal equilibrium, the energy balance
can be approached with a unique mixture equation. Since the density and the thermal conductivities
are common to both equiaxed/columnar liquid and solid phases, the energy balance equation is
identical with the one obtained for the precedent case, equation (78) in which one has evidently

ε s = ε se + ε sc , ε f =1− εs , Γ s = Γ se + Γ sc (145)

Solute balance
A total of six balance equations have to be written. Indeed, for each of the two structures (e/c) one
has three corresponding solute balance equations: for the solid and for the inter- and extra-dendritic
liquid

∂ ( ε sk ρ s Csk )
= div ( ε sk ρ s Ds ∇Csk ) + Φ Csk + ΓCsk
∂t
(
∂ ε dk ρ f Cdk ) = div ε ρ D ∇C + ΦC + ΓC
∂t
( dk f f dk ) dk dk k ∈ {c, e} (146)

∂ ( ε lk ρ f Clk )
= div ( ε lk ρ f D f ∇Clk ) + Φ Clk + ΓClk
∂t

Using the cell approximation to the c/e ensemble average, the above equations can be closed using
a similar procedure as for the single grain distribution, see paragraph 3.4.2. The main difference
lays in the choice of the integration cell with respect to the columnar/ equiaxed structure. Indeed,
each one of the exchange terms in (146) is determined as an integral over all posible positions of
the c/e grain within the cell of radius R1*c / e centered at x, i.e. equation (136). Thus, it is easy to
verify that

ΦCsk = 0 ΓCsk = Cs∗Γ sk

ρ f S gk Dl , k = {e, c} (147)
ΦClk =
δ lk − dk
(C
∗
l − Clk ) ΓClk = −Cl∗Γ gk

and finally
ρ f S gk Dl
ΦCdk + ΓCdk = Cl∗Γ gk − Cs∗Γ sk −
δ lk −dk
(C
∗
l − Clk , ) k = {e, c} (148)

where δ lk − dk is the diffusion length characterizing the solute transfer at the e/c grain interface. As
pointed out, this diffusion length can be computed from equation (A.18) by replacing R1 with
R1*e / c and a with ae,c .
Using the e/c solid mass balance equation, the solute balance in the e/c extra-dendritic
liquid can be written as follows:

II-46
Submitted to Journal of Physics D: Applied Physics

∂Clk ⎛ ρ f S gk Dl ⎞
ε lk ρ f
∂t
(
= ∇ ⋅ ε lk ρ f D f ∇Clk + ⎜ )
⎝ δ lk − dk ⎠
(
− Γ gk ⎟ Cl∗ − Clk , k = {e, c} ) (149)

Using the special property of δ lk − dk , equation (A.19), it is easy to see that the last term of the RHS
of (149) is always positive. Since the macroscopic diffusion process is only redistributing solute in
the ingot, ∂Clk / ∂t ≥ 0 and therefore the extra-dendritic liquid is always enriching. Notice that the
solute diffusion around an e/c grain is spatially limited to a maximum distance R1*e / c around each
grain. This together with zero solute flux condition at the cell boundary, equation (138) will
determine a continuous enriching of the extra-dendritic liquid which will finally arrive at a state of
complete solute mixing. This would be equivalent with a complete stop of the grain growth. For
this reason the grain growth stop will always take place for an e/c grain fraction lower than the
corresponding e/c fraction one. Hence, the maximum e/c grain fraction will always be smaller than
ε e / c , verifying thus an important physical constraint.

4.3.2. Summary of the model. In the following table the equations of the model are sumarized. It is
important to notice that it involves a six-phase approach to the solute balance. Indeed one can
identify two solute balance equation in the e/c solid, two solute balances in the e/c extra-dendritic
liquid and finally two equations stating for the solute balance in the e/c inter-dendritic liquid. The
last two equations are presented as auxiliary equations since they are used to compute the e/c solid
mass transfers, Γ se and Γ sc . Indeed, by fixing the concentration in the inter-dendritic liquid one
encompasses the need to model Cd obtaining in turn an equation for the solidification rate Γ se / c .

Table 4: Summary of the model

Main balance Solid: Liquid:
equations k = {e, c} k = {e, c}
Mass ∂ ( ε sk ρ s ) ∂t = Γ sk (
∂ εf ρf ) ∂t = −Γ se − Γ sc
Energy
( ) ( )
∂ ⎡ ε s ρ s cs + ε f ρ f c f T ⎤ ∂t = div ⎡ ε s λs + ε f λ f ∇T ⎤ + ( Γ se + Γ sc ) L ,
⎣ ⎦ ⎣ ⎦
(ε s = ε se + ε sc )
Solute ∂ ( ε sk ρ s Csk ) ∂t = ∇ ⋅ ( ε sk ρ s Ds ∇Csk ) ε ρ ∂Clk = ∇ ⋅ ε ρ D ∇C
lk f
∂t
lk f f lk ( )
+Cs∗Γ sk
(
+ ρ f S gk Dl / δ lk − dk − Γ gk Cl∗ − Clk )( )
Supplementary
balance eq.
Inter-dendritic
mass balance
(
∂ ε dk ρ f ) ∂t = Γ gk − Γ sk
Grain balance ∂ne ∂t = ne0δ (T − TNe )

( )
∂nc / ∂t = nc0 1 − ε ge δ ⎡⎣ x − xcf (t ) ⎤⎦
Solidification rate
(C ∗
l ) ( )
− Cs∗ Γ sk = ε dk ρ f ∂Cl∗ ∂t + ρ f S gk D f Cl∗ − Clk ( )δ lk − dk

(
− div ε dk ρ f D f ∇Cl∗ )
Auxiliary
expressions
Constraints: ε le + ε de + ε se = ε e ε de + ε se = ε ge
, εc + εe = 1,
ε lc + ε dc + ε sc = ε c ε dc + ε sc = ε gc

II-47
Submitted to Journal of Physics D: Applied Physics

Interfacial
concentrations:
⎪⎧kC
Cs∗ = ⎨ l
∗
if Γs ≥ 0 (
Cl∗ = T − T f ) m
⎪⎩Cs if Γs < 0
Equiaxed/column
ar fractions
( )
ε e , ε c = f ne , nc0 , ε ge , see equation (129)
Cell radius R1*k = ( 4 / 3π nk )
−1/ 3
ε k1/ 3
Grain radius ak = R1k ε gk1/ 3
Grain envelope S gk = 4π ak 2 nk
interfacial area
Diffusion length δ lk −dk , see equation (A.18)
Grain growth rate Dl m ( k − 1) Cl∗ 2
Γ gk = ρ f S gk w gk , w gk = 2
⎡⎣ Pet ( Ω k ) ⎤⎦
π Γ

5. Differences between the volume average and the ensemble average approach
As already anticipated one can expect to observe significant differences between the results
provided by the ensemble average approach on one hand and the volume average on the other
hand. In paragraph 3.4.3. the similarity between the present approach and the volume average
model of Wang and Beckermann [6] was remarked. However the similarities between the two
models are a direct consequence of the local homogeneity hypothesis of all fields, meaning that no
local gradients at the scale of the grain are considered. In the following we will show that given the
hypothesis of local non-homogeneity the ensemble and the volume average closure techniques give
birth to very different results especially when the computation of the closure expressions for
interfacial averaged fields are considered.

Figure 17: a) The cell approximation to the ensemble average; b) The cell approximation to the
volume average.
Let us consider that no local thermal equilibrium can be considered at x. Hence a non-
negligible thermal gradient exist at the scale of the grain. Let us denote it with ∇T (see Figure 17).
Let us consider as well that the far concentration from the grain interface is constant, C0 . In this
respect, let us compute for example the interfacial average Γ g (the grain growth) using both
ensemble and volume approach. The use of the ensemble average and more precisely of the cell
approximation to the ensemble average, equation (53), would give the following results:

II-48
Submitted to Journal of Physics D: Applied Physics

Γg = n ∫ ρ ( − wg n g ) ⋅ ⎡⎣ −n g δ ( x − xi ( t ) ) ⎤⎦ d z1
x − z1 ≤ R1

= n ρl ∫ wg ( x ) dS (150)
z1∈S ( a )

= n 4π a 2 ρl wg ( x )

Dl m ( k − 1) Cl∗ 2
where the velocity wk = 2
⎡⎣ Pet ( Ω ) ⎤⎦ is always evaluated at x since the above
π Γ
integration is the made over the position of the grain in the cell such as the point x is always on the
grain interface, that is over the spherical shell S (a ) of radius a and centered at x. Hence the local
Cl* (x) − C0
undercooling is evaluated at x too, Ω = .
Cl* (x) − kCl* (x)
On the other hand the use of the cell approximation to the volume average to compute Γ g
would mean an integration of wg over the grain interface S (a ) , given a fixed grain centered at x
(see Figure 17b). Hence one obtains:

Γg = n ∫ ρ wg z1 dS( )
z1∈S ( a ) (151)
2
= n 4π a ρl w g
where
wk =
1
4π a 2
∫ ( )
wg z1 dS (152)
z1∈S ( a )

is the mean growth velocity of the grain. It is important to notice that due to the non-linearity of
wg with respect to Cl* and consequently with respect to T, the mean velocity w g will be different
from the velocity wg ( x ) in (150). Hence, different closure expressions are obtained using the
ensemble and the volume average respectively. Moreover the result in (151) is incorrect because
the mean interfacial velocity w g at x should be an average of the velocity wg evaluated at x and
not on the grain interface S (a ) as in Figure 17b. The error comes from the fact that the cell
approximation to the volume average considers a fixed grain centered at x. Therefore the average
fields computed in the cell (including the liquid and the interfacial fields) are attached to the point
x, a point that is always located in the solid.
Similar findings are obtained by Drew [9] when he compared the volume and the ensemble
average of the interfacial pressure force acting on the grain interface. Indeed if a pressure gradient
∇p is to be considered at the scale of the grain the ensemble average of the interfacial pressure
force would be zero since:

Fp = n ∫ p ( x ) n g dS
z1∈S ( a ) (153)
=0

On the other hand the volume average approach would give:

II-49
Submitted to Journal of Physics D: Applied Physics

Fp = n ∫ ( )
p z1 n g dS
z1∈S ( a ) (154)
≠0

that is a non-zero buoyancy force because the pressure force is integrated over the grain interface
S (a ) , given the grain centered at x. The result in (154) is physically incorrect since the buoyancy
force would appear in the momentum equation twice, along with the already existing force due to
the fluid pressure gradient. This emphasizes the difficulties inherent to the volume average
approach.

6. Conclusions
In the present paper a new multi-phase Eulerian model valid for both columnar and equiaxed
dendritic solidification has been developed in the purely diffusive case. The mean conservation
equations are derived by means of a statistical phase averaging technique, and the mathematical
formulation of the model can be used for both columnar and equiaxed solidification. The use of
such an averaging procedure has several advantages compared with the volume averaging
technique which has been widely used during the last decade. Firstly, the statistical approach does
not need the definition of a representative elementary volume. Secondly, it is consistent with the
random aspect of the equiaxed grain germination and motion as well as the possible turbulent
behavior of the fluid flow. Finally, and the most important, owing to the statistical nature of the
model, we are able to treat rigorously the coexistence of equiaxed and columnar structures and
consequently the CET phenomena.
First the pure columar/equiaxed case is analyzed. With the help of the nearest neighbor
probability distribution functions one is able to rigorously compute the unknown terms and tensors
existing in the unclosed averaged balance equations. Indeed, using the cell approximation to the
(1)
nearest neighbor distribution, f ( t, x; z1 ) , the average fields can be easily computed by means of
simple integration within a characteristic spherical cell. The main physical hypothesis refers to the
scale separation between the inter- and the extra-dendritic liquid. In a similar way as Rappaz and
Thévoz [14] and Wang and Beckermann [6], an envelope model is used to parameterize the length
scale smaller than the size of the dendrite, that is the interdendritic spacing. The dendrite envelope
is defined as the virtual surface surrounding the c/e dendrite grain and connecting all c/e dendrite
tips. This leads us to distinguish two types of liquid, the inter- and the extra- dendritic one, and this
for each one of the two grain structures. The inter-dendritic liquid is supposed to be in a state of
perfect solutal mixing, its solute concentration, Cl* , being uniquely determined from the local
temperature and the phase diagram. The cell model together with the physical meaningful
hypotheses valid at the micro-scale enables one to completely close the complex system of
averaged balanced equations. The final model valid either for the pure columnar or equiaxed case is
similar with the one obtained by Wang and Beckermann [6]. The equivalence between the two
models is a direct consequence of the use of the local thermal equilibrium hypothesis. Indeed if the
local homogeneity hypothesis is considered, meaning that no local gradients of physical fields
(temperature, solute, grain fraction) are taken into account at the scale of the grain, then the
ensemble average reproduces the volume average technique. However the ensemble average
technique and the cell model approximation to the ensemble average are fundamentally different
from the volume average approach. Indeed, if the local homogeneity hypothesis is no more valid,
the use of the ensemble average results into closure expression different from the ones obtained
with the help of the volume average technique. Moreover the volume average results are not
physically sound. In this respect it is shown that the use of the cell appoxmation to the ensemble
average can be successfully used to incorporate the effects of the inhomogeneities of the various
scalar fields, e.g., the solute and temperature gradients on the closure expressions.
Secondly the case of the coexistence between the columnar and equiaxed grains is
analyzed from the perspective of the ensemble average technique. More precisely the case of a
columnar front "penetration" into a developed equiaxed zone is analyzed. This type of coupling can
be frequently encountered during solidification problems especially in processes involving the so-

II-50
Submitted to Journal of Physics D: Applied Physics

called columnar-to-equiaxed transition (CET) phenomena. The main advantage of the ensemble
( )
averaging is that, with the help of the nearest neighbor distributions f c / e t , x; z1 , one can quantify
in a rigorous way the c/e influence at x. Using the cell approximation to f c / e , two spherical cells
* *
centered at x and of radius R1c and R1e were identified, each of them characterizing in a simple
way the statistics of the c/e grain distributions (see Figure 16). Using the two spherical cells it is
possible to separate the influence at x of the c/e grains by defining corresponding c/e ensemble
averages which will quantify the average effects of the two structures at x. Moreover, any ensemble
average (i.e. c/e fractions, solid/liquid fractions, average solute fields, interfacial averages, etc.) can
be easily computed by means of relative simple integrations over all possible positions of the c/e
grain within the respective c/e cell. Thus, following the case where the closest grain to x is
columnar or equiaxed, the model leads us to distinguish two main types of ensemble averages,
columnar and equiaxed ensemble averages respectively. In this respect, two important fractions are
identified, the columnar and equiaxed fraction, ε c and ε e respectively. These two averages
quantify the way how the two structures share the space at x. Since the columnar and equiaxed
grains share a same space at x their corresponding fractions will obviously sum to one, i.e.,
ε c + ε e = 1 . Moreover each structure will be characterized by its proper solid (s) and liquid (f)
phase and therefore the c/e fraction will be a sum of their corresponding c/e solid and liquid
fractions, ε c / e = ε se / c + ε fe / c . Another important feature of the ensemble average is that it enables
one to quantify the mechanical blockage of the equiaxed grains on the columnar tips. It was proven
that the columnar and the equiaxed grain distributions interact “mechanically” following a two-fold
mechanism. The first mechanical blockage effect of the equiaxed grains will determine a
rarefaction of the columnar zone when entering (“nucleating”) into a developed equiaxed zone.
Indeed, due to the finite equiaxed grain fraction the columnar tips have only a limited space
(1 − ε ge ) within which they “nucleate”. Hence only a fraction, 1 − ε ge , of the total number of
columns, nc0 , will be able to enter (to nucleate within) the equiaxed zone. The second mechanical
blockage effect of the coexistence reflects the fact that for a mixed zone (i.e. columnar + equiaxed),
the two coexisting structures are forced to share a same space. Consequently, the space available
for each structure will be different and smaller with respect to the corresponding space available
just before the coexistence state. Basically the second mechanical blockage states that the cell radii
R1*c / e characterizing the coexistence statistics are smaller than the corresponding length scales
characterizing the c/e state separately just before their coexistence, R1c / e . This effect is of primary
importance because R1*c / e can be very different form the original length scales R1c / e and
consequently the evolution of the mixed zone can be significantly different from that of the pure
columnar or equiaxed zone. Using the cell model together with the physical assumptions valid at
the micro-scale we are able to close rigorously the whole system of balance equations. Finally, a
complete model describing the coexistence of equiaxed and columnar case grains is obtained.

Appendix A: Solute diffusion field in the extradendritical liquid

As pointed out in section 3.4.1., by assuming a spherical shape for the dendrite grain, a thermal
equilibrium at the scale of the grain and the local homogeneity hypothesis, the solute field around
the grain will be a function of r only. As well, the solute flux at the cell boundary ( r = R1 ) is zero,
meaning that ( ∂Cl ∂r )r = R = 0 . In this respect the solute diffusion process within the extra-
1

dendritic liquid will be controlled by the following boundary moving equation:

II-51
Submitted to Journal of Physics D: Applied Physics

∂Cl Dl ∂ ⎛ 2 ∂Cl ⎞
= 2 ⎜ r ∂r ⎟
∂t r ∂r ⎝ ⎠
Cl ( t , r = a ) = Cl* , ( ∂Cl ∂r )r = R = 0
1
(A.1)
da
= w ge
dt

where w ge is the growth velocity of the grain envelope, computed from (59) where C∞ would be
equal to the mean extra-dendritic concentration C l :

R1
1 2
Cl = ∫ Cl 4π r dr (A.2)
4 / 3π ( R13 −a 3
) a

Figure A1: The solute field in the extra-dendritic liquid

By integrating the diffusion equation in equation (A.1) on the extra-dendritic liquid domain and
using the well known Leibnitz integral rule one can easily obtain the solute balance equation in the
extra-dendritic liquid:

R1
⎛ ∂C ⎞
∂
∂t ∫ Cl 4π r
2
( )
dr + Cl* 4π a 2 w ge = − 4π a 2 Dl ⎜ l ⎟
⎝ ∂t ⎠ r = a
( ) (A.3)
a

By denoting with 2δ the solutal boundary layer existing around the spherical grain (Figure A1)
one can rewrite the above solute balance as follows

Rg + 2δ R
Cl* − C0
∂
∂t ∫ Cl 4π r dr +2∂ 1
∂t a +∫2δ
C0 4π r 2
dr + Cl
*
4π a 2
w ge = 4π a 2
(
Dl
δ
) ( ) (A.4)
Rg

The first term in the RHS of the above equation is no more than the solute variation of the solute
within the solutal boundary. This can be rewritten as:

∂ a + 2δ ∂
∂t a∫
Cl 4π r 2 dr =
∂t
Vsb C sb ( ) (A.5)

II-52
Submitted to Journal of Physics D: Applied Physics

where Vsb is the volume of the boundary layer and C sb the average concentration within the same
boundary layer. The second term in the RHS of (A.4) represent the variation of the solute within
the space outside the solutal boundary untouched by the solute diffusion wave. One has therefore:

R
∂ 1 ∂
∫ C0 4π r 2 dr = ⎡⎣(Vl − Vsb ) C0 ⎤⎦ (A.6)
∂t a + 2δ ∂t

where Vl is the volume of the extra-dendritic liquid and C0 the uniform concentration of the liquid
untouched by the diffusion wave. By assuming now a solute profile within the boundary layer one
can obtain from (A.4) an ordinary differential equation controlling the evolution of δ . This
technique is no more than the integral approach to a boundary layer problem as detailed in
Burmeister [20]. Note that this approximate integral approach in (A.4) is simpler than the one used
by Rappaz and Thévoz [14] since it requires the solution of a relative simple ODE in order to
obtain the δ evolution. In contrast in [14] the full solution of the PDE equation in (A.1) is
computed. However, solving an additional ODE for each cell of the spatial mesh would increase
greatly the computational time of the already complex problem. To simplify and similar with the
approach adopted by Wang and Beckermann [6] we will consider that the solute diffusion problem
within the extra-dendrite liquid is quasi-stationary. This hypothesis would be valid only if the
solute boundary layer around the grain readapts more rapidly to the external conditions ( a, Cl* )
than the later ones are actually changing. In this respect one expects that the variation of the
parameters characterizing the solutal boundary layer (both Vsb and C sb ) with respect to time is
negligible. It is worthwhile to notice that Rappaz and Thévoz [21] made a similar assumption in his
approximate approach of the solute diffusion problem in the extra-dendritic liquid. Given this
hypothesis, from (A.4) one finally obtains:

Cl* − C0
( ) ( ) (
−C0 4π a 2 w ge + Cl* 4π a 2 w ge = 4π a 2 Dl ) δ
(A.7)

The solute boundary layer characterizing the quasi-steady state becomes:

Dl
δ st = (A.8)
w ge

Notice that the same result was obtained by Rappaz and Thévoz in [21]. It would be however
interesting to analyze the validity domain of the quasi-steady hypothesis. To do this let us quantify
first the transition time from an initial solutal state characterized by the following initial and
boundary conditions:

Cl* = ct (T = ct ) ⎫⎪
⎬ ⇒ w ge = ct
Cl* > C0 ⎪⎭ (A.9)
a(t = 0) > 0

to a quasi-stationary state . This transition time would be of order of the time scale tst during
which the solute boundary around the grain extends to a dimension corresponding to the quasi-
steady state already detailed above. Knowing the travel time of a diffusion wave is tdiff = l 2 / 4 / Dl
[22] one easily obtains for tst :

II-53
Submitted to Journal of Physics D: Applied Physics

( 2δ st )2 Dl
tst = = 2
(A.10)
4 Dl w ge

So, one can approximate the solute diffusion problem with a quasi-steady state if the readaption
time of the solutal boundary to the new condition is smaller than the time scale characterizing the
changes of external parameters like a, w ge , T . For the case of a grain distribution characterized by
an average spacing between particles 2 R1 the time scale characterizing the change of external
conditions could be approximated with the total growth time of the grain:

R1 − a
t* (A.11)
w ge

In this respect the quasi-steady approximation would be valid if

tst << t * ⇔ δ st << ( R1 − a ) (A.12)

It is interesting to note that the above condition is also coherent with the result in (A.8). Indeed, due
to the spatially limited diffusion around the grain, the stationary boundary layer δ st have a
meaning only if 2δ st << ( R1 − a ) (see Figure A1) or:

w ge ( R1 − a )
Peg = >> 2 (A.13)
Dl

If on the contrary Peg << 2 a quasi-steady state cannot be considered. In fact, Peg = 2 would
correspond to a case where the solute boundary layer around the grain have already reached the cell
boundary. Above this limit, due to the zero flux condition at r = R1 , the enriching of the extra-
dendritic liquid becomes non-negligible and the far concentration C0 starts to increase. For this
non-stationary regime, the solute distribution in the extra-dendritic liquid becomes highly time
dependent and the precise modeling of the solute transfer at the grain boundary would require a
more complex model for δ . However, as a first approximation, one could consider that for
Peg < 2 a pertinent solutal boundary layer would be 2δ = R1 − a . One would have therefore:

⎧⎪ Dl / w ge if Peg ≥ 2
δ =⎨ (A.14)
⎪⎩( R1 − a ) / 2 if Peg < 2

It is importnant to notice that the above solutal boundary layer model enable one to express the
interfacial solute transfer in the extra-dendritic liquid ΦCl , equation (81). However, in an average
model one knows a priori the mean concentration in the extra-dendritic liquid, C l , and not the far
concentration C0 . Hence one should express the solute flux at the grain interface with respect to
C l and not C0 :

Cl* − C0
ϕ g = Dl
δ
(A.15)
Cl* − Cl
= Dl
δ l −d

II-54
Submitted to Journal of Physics D: Applied Physics

One must therefore compute an analytical expression for δ l − d (Figure A1). This can be done if one
assumes a solute profile within the solute boundary layer. A straightforward profile would be the
second order approximation:

Cl (r ) = br 2 + cr + d , r ∈ ( a, a + 2δ ) (A.16)

The three constants b, c, d can be easily found by imposing that:

⎛ ∂Cl ⎞
Cl (a ) = Cl* ; Cl (a + 2δ ) = C0 ; ⎜ ∂r ⎟ =0 (A.17)
⎝ ⎠a + 2δ

It is worthwhile to notice that equation verifies as well the constraint: ( ∂Cl / ∂r )a + 2δ = δ . Using
(A.2), (A.15) and (A.16) one finally obtains:

⎡ ⎤
Cl* − C l ⎢ 10a 2δ + 10aδ 2 + 4δ 3 ⎥
ϕ g = Dl where δ l −d = δ 1− (A.18)
δ l −d ⎢
⎣ (
5 R13 − a3 ⎥
⎦ )
Coupled with the equation (A.14) the above expression form a complete analytical model
describing the solute flux at the grain interface. Note as well that δ l − d verifies an important
constraint relative to the solute diffusion process in the extra-dendritic liquid, namely that:

Dl
δ l −d < (A.19)
w ge

Indeed, as noticed by Wang and Beckermann [6], due to the continuous enriching process of the
extra-dendritic liquid δ l − d has to verify (A.19) as to be physical meaningful.

Appendix B: Probability distribution functions. The coexistence case.

(1)
The distribution function f e ( t, x; z1 )
(1) (1)
In order to find an expression for f e ( t, x; z1 ) one must see that fe ( t , x; z1 ) d z1 is the probability
that there is no sphere within r = x − z1 of x and there is one sphere within d z1 of z1 . Note as
well that the probability that there is no sphere within r of x and the probability that the nearest
sphere is within r of x sum to one. Hence one has:

⎛ ⎞
(1)
fe ( )
⎜
t , x; z1 d z1 = ⎜1 −
⎜ x − z1' < x− z1
(1) '
(
' ⎟ (1)
)
∫ f e t , x; z1 d z1 ⎟ fe z1 d z1
⎟
( ) (B.1)

⎝ ⎠

( )
where f e(1) z1 is the unconditional density functions for any equiaxed grain center being within
d z1 of z1 , that is equal with equiaxed grain desnity ne . Hence, one can notice from (B.1) that
(1)
fe ( t, x; z1 ) is a function of r = x − z1 only. By replacing d z1 with 4π r 2 dr one can observe that
the left side of (B.1) is the differential of the integral on the right. One has therefore:

II-55
Submitted to Journal of Physics D: Applied Physics

d ⎛ r
(1) ⎞
ln ⎜1 − 4π ∫ f e ( t , r ')( r ') dr ' ⎟ = −4π ne r 2
2
(B.2)
dr ⎜⎝ 0
⎟
⎠
Integrating,
(1)
( t, x; z1 ) = C1 exp ⎛⎜⎝ − 43 π ne x − z1
3⎞
fe ⎟ (B.3)
⎠

where C1 is a constant which is evaluating by requiring that

(1)
∫ fe (t , x; z1 )d z1 = 1 (B.4)

Hence one obtains:

(1)
( t , x; z1 ) = ne exp ⎛⎜⎝ − 43 π ne x − z1
3⎞
fe ⎟ (B.5)
⎠

(1)
The distribution function f c ⎛⎜ t , x; z1 Ae ⎞⎟
*

⎝ ⎠
Similar with the previous case one can state that:

⎛ ⎞
f
(1) ⎛
⎝
*
c ⎜ t , x; z1
⎞
⎠
* ⎜
Ae ⎟ d z1 = ⎜ 1 −
⎜ '
∫ *
x − z1 < x − z1
f
(1) ⎛ '
c ⎜ t , x; z1
⎝
⎞
(
' ⎟ (1) *
) *
Ae ⎟ d z1 ⎟ f c z1 Ae d z1
⎠ ⎟
(B.6)

⎝ ⎠

(
where f c(1) z1 Ae
*
) is the conditional density functions for any columnar grain center being within
* *
d z1 of z1 , given the event Ae that the the closest equiaxed grain to x is centered at z1 where
*
x − z1 < x − z1 . It is important to notice that f c(1) z1 Ae ( *
) is equal with the columnar nucleation
grain density nc0 . Similar with the previous case one finally obtains

(1) ⎛ 4 * 3⎞
f c ⎛⎜ t , x; z1 Ae ⎞⎟ = C1 exp ⎜ − π nc0 x − z1 ⎟
*
(B.7)
⎝ ⎠ ⎝ 3 ⎠

(1) ⎛ * ⎞
Requiring that ∫ fc ⎜ t , x; z1 Ae ⎟d z1 = 1 one has
⎝ ⎠

(1) ⎛ 4 * 3⎞
f c ⎛⎜ t , x; z1 Ae ⎞⎟ = nc0 exp ⎜ − π nc0 x − z1 ⎟
*
(B.8)
⎝ ⎠ ⎝ 3 ⎠

II-56
Submitted to Journal of Physics D: Applied Physics

Nomenclature

Symbol Meaning Unit

v velocity field [m/s]

C solute concentration field [wt. %]
T temperature field [K]
n grain density [ m −3 ]
ρ density field [ kg .m −3 ]
ψ the generic physical field ---
εk the “volume” fraction of phase k ---
wk the interface velocity of the k phase interface [m/s]
Ck* the interfacial concentration at the k phase interface [wt. %]
Xk the k phase characteristic phase function ---
nc0/ e the columnar/equiaxed nucleation grain density [ m −3 ]
λ1 columnar primary arm spacing [m]
Sk the interfacial area density of phase k [ m −1 ]
δk− p the solute diffusion length at the k-p interface [m]
xcf position of the columnar front [m]
Tne the equiaxed nucleation temperature [K]
Tf the fusion temperature of the pure solvent [K]
m the slope of the liquidus line [ K .( wt % ) ]
−1

k the partition coefficient ---

ck heat capacity for the k-phase [ J .kg −1.K −1 ]
λk the phase k thermal conductivity [ W .m −1.K −1 ]
L the latent heat [ J .kg −1 ]
Ω the dimensionless local undercooling ---
Pet the dendrite tip Peclet number ---
Dk the solute diffusivity within the phase k [ m 2 .s −1 ]
ac / e the local columnar/equiaxed grain radius [m]
R1c / e the cell radius corresponding to the single columnar/equiaxed [m]
grain distributions
R1*c / e the cell radius corresponding to the coexisting [m]
columnar/equiaxed grain distribution
zi the position in the ingot of the ith grain [m]
f ( ) ( t , z1 ) the unconditional density function that at time t in the ingot [ m −3 ]
1
one can find a particle centered within dz1 of z1
zi the position in the ingot of the ith closest grain to x [m]
(1)
(
f c / e t , x; z1 ) the unconditional density function that the closest grain to the
point x is centered within d z1 of z1 and is a
[ m −3 ]

columnar/equiaxed grain
Γk the k phase mass transfer rate [ kg .m −3 .s −1 ]
Over lines
ensemble average

II-57
Submitted to Journal of Physics D: Applied Physics

relative to the nearest neighbor distributions

Subscripts
c columnar phase
e equiaxed phase
c/e columnar / equiaxed phase
s solid phase
l extra-dendritic liquid phase
d inter-dendritic liquid phase
f total fluid phase ( = l + d )
g grain phase ( = d + s )
kc columnar k phase
ke equiaxed k phase

References

1. Bennon W D and Incropera F P 1987 Int. J. Heat Mass Transfer 30 2161

2. Ganesan S and Poirier D R 1990 Metallurgical Transactions B 21B 173

3. Poirier D R, Nadapurkar P J and Ganesan S 1991 Metallurgical Transactions B 22B 889

4. Ni J and Beckermann C 1991 Metallurgical Transactions B 22B 349

5. Felicelli S D, Heinrich J C and Poirier D R 1991 Metallurgical Transactions B 22B 847

6. Wang C Y and Beckermann C 1993 Metallurgical and Materials Transactions A 24A 2787

7. Wang C Y and Beckermann C 1996 Metallurgical and Materials Transactions A 27A 2754

8. Drew D A 1983 Ann. Rev. Fluid Mech. 15 261

9. Drew D A and S L Passman 1999 Theory of Multicomponent Fluids (New York: Springer-
Verlag)

10. Furmanski P 2000 Computer assisted mechanics and engineering sciences 7 402

11. Fautrelle Y, Lehmann P, Quillet G, Medina M, Durand F and Du Terrail Y 2000 Proc.
EUROMECH Colloquium 408 “Interactive dynamics of convection and solidification” ed. By
Kluwer Pub 87

12. Ciobanas A I, Baltaretu F and Fautrelle Y 2004 Proc. of Solidification and Gravity, Miskolc

13. Kataoka I, 1986 Int. J. Multiphase Flow 12.5 745

14. Rappaz M and Thévoz Ph 1987 Acta Metallurgica 35.7 1487

15. Martorano M A, Beckermann C and Gandin Ch-A 2003 Metallurgical and Materials
Transactions A 34A 1657

16. Stefanescu D M, Upadhya G and Bandyopahyoy D 1990 Metallurgical Transactions A 21A

997

17. Hunt J D 1979 Solidification and casting of metals – Metals Society, London 3

II-58
Submitted to Journal of Physics D: Applied Physics

18. Kurz W and Fisher D J 1981 Acta Metallurgica 29 11

19. Lipton J, Glicksman M E and Kurz W 1984 Materials Science and Engineering 65 57

20. Burmeister L C 1993 Convective Heat Transfer (New York: Wiley)

21. Rappaz M and Thévoz Ph 1987 Acta Metallurgica 35.12 2929

22. Bejan A 2000 Shape and Structure, From Engineering to Nature (New York: Cambridge
University Press)

II-59
Référence III
Submitted to Journal of Physics D: Applied Physics

Ensemble averaged multi-phase eulerian model for

columnar/equiaxed solidification of a binary alloy
Part II. Simulation of the columnar-to-equiaxed transition
A. I. Ciobanas1, Y. Fautrelle1
1
EPM/CNRS laboratory, ENSHMG, BP 95, 38402 Saint Martin d’Hères Cedex, France

E-mail: aciobanas@yahoo.com

Abstract
A new multi-phase Eulerian model for the columnar and equiaxed dendritic solidification has been
developed. In this paper we first focus on the numerical simulation of quasi-steady solidification
experiments in order to obtain corresponding CET maps. We have identified three main zones on
the CET map: the pure columnar, the pure equiaxed zone and finally the mixed columnar+equiaxed
zone. The mixed c/e zone was further quantified by means of a columnar fraction ε c which
quantifies in a rigorous the way the two coexisting structures. Since it intrinsically includes the
solutal and the mechanical blocking effects, the new ensemble model unifies the semi-empirical
Hunt’s approach (pure mechanical blocking mechanism) and the Martorano et al. approach (pure
solutal blocking mechanism). Secondly the present model was used to simulate unidirectional
solidification experiments. It was found that the columnar front evolved in a quasi-steady state until
a time very close to the critical CET moment. It is also found that the equiaxed nucleation
undercooling is close to the maximum columnar dendrite tip undercooling and that the CET is
virtually independent of the equiaxed zone ahead of the CF. If the equiaxed zone is not taken into
account it is observed that the CF velocity exhibits a sudden increase at the beginning of the
solidification followed by a quasi-plateau corresponding to a quasi-state at the columnar tips and
finally, above a critical time, an oscillatory evolution. The beginning of the oscillatory evolution of
the CF was well correlated with the CET position measured in the experiments. We find as well
that this oscillatory evolution of the CF is very favorable for the fragmentation of the columnar
dendrites and thus for the CET. In this respect, it seems that the unsteady regime of the CF with
respect to the local cooling conditions represent the main cause for the CET phenomena, at least for
the non-refined alloys.

PACS: 81.30.Fb Solidification, 82.20.Wt Computational modeling; simulation

1 Introduction
The present paper is devoted to the application of a model derived in a previous publication [1].
That model is aimed at predicting the macroscale solidification of a binary alloy. It is based on the
ensemble averaging technique, which allows the treatment of various types of solidification, i.e.,
columnar or/and equiaxed. The use of an ensemble averaging technique for deriving a micro-macro
scale model has some important advantages compared with the volume averaging technique which
has been widely used during the last decade (see for example [2, 3]) . Firstly, the statistical
approach does not need the definition of a representative elementary volume. Secondly, it is
consistent with the random aspect of the equiaxed grain germination and motion as well as the
possible turbulent behavior of the fluid flow. Finally, and the most important, owing to the
statistical nature of the model, one is able to treat rigorously the coexistence of equiaxed and
columnar structures and consequently the CET phenomena. Knowing that the CET phenomena are
direct consequence of complex mechanical and solutal interactions between an advancing columnar
front and an equiaxed zone ahead of the columnar dendrite tips, the existence of a model that could
integrate mathematically these interactions would be of great interest.
The paper comprises four sections. In Section 2, the mathematical is briefly described, and
the overall equations are summarized. Section 3 is devoted to the analysis of the ability of the
model to predict the CET mechanism in a simple 1D directional solidification situation. In the

III-1
Submitted to Journal of Physics D: Applied Physics

fourth section the model is tested on previous basic CET experiments [4, 5] devoted to the study of
the CET.

2 Mathematical model

2.1 The ensemble averaged model

The model is based on the ensemble averaging procedure coupled with the so-called cell model.
The model is described in detail in a first part [1]. We only summarize here the main features of the
model. Given two coexisting grain populations (i.e. columnar and an equiaxed) characterized by
two different grain densities, nc / e (c/e stands for either the columnar or the equiaxed structure),
their influence at a certain point x may be different as well. The main advantage of the ensemble
averaging is that, with the help of the nearest neighbor distributions if ( )
t , x; z 1 , one can quantify
c/e

( )
in a rigorous way the c/e influence at x. Indeed, if c / e t , x; z 1 measures the density of probability
that one c/e grain centered at z 1 is the closest grain to x among all the grains existing in the ingot.
These two distributions play a vital role in determining the influence of c/e grains at x since one
expects that what happens at x would be mainly influenced by the closest grain to x. Using the cell
approximation to if c / e , two spherical cells centered at x and of radius R1c *
and R1e*
can be
identified, each of them characterizing in a simple way the statistics of the c/e grain distributions
(see Figure 1).

Figure 1: The cell model approximation to the columnar (a) and the equiaxed (b) ensemble average

Indeed, in the light of the cell model the ensemble θ c / e of realizations such as the closest
particle to x be a columnar/equiaxed one is equivalent with the ensemble of realizations such as
the closest particle to x is c/e and is centered within the cell of radius R1*c / e centered at x. Using the
two spherical cells it is possible to separate the influence at x of the c/e grains by defining c/e
ensemble averages which will quantify the average effects of the two structures at x. Moreover, any
ensemble average (i.e. e/c fractions, solid/liquid fractions, average solute fields, interfacial
averages, etc.) can be easily computed by means of relative simple integrations over all possible
positions of the c/e grain within the respective c/e cell. Hence, following the case where the closest
grain to x is columnar or equiaxed, the model leads us to distinguish two main types of ensemble
averages, the columnar and equiaxed ensemble averages respectively. Accordingly, two important
fractions are identified, the columnar and equiaxed fraction, ε c and ε e respectively. These two
averages quantify the way the two structures share the space at x. Since the columnar and equiaxed
grains share a same space at x their corresponding fractions will obviously sum to one, ε c + ε e = 1 .
Moreover each structure will be characterized by its proper solid (s) and liquid (f) phase therefore

III-2
Submitted to Journal of Physics D: Applied Physics

the c/e fraction will be a sum of their corresponding c/e solid and liquid fractions,
ε c / e = ε sc / e + ε fc / e .
In a similar way as Rappaz and Thévoz [6] and Wang and Beckermann [2], an envelope
model is used to parameterize the smallest scale of the dendrite, that is the interdendritic spacing.
The dendrite envelope is defined as the virtual surface surrounding the dendrite grain and
connecting all dendrite tips. This leads us to distinguish two types of liquid, the inter- and the extra-
dendritic one, and this for each of the two grain structures. The inter-dendritic liquid is supposed to
be in a state of perfect solutal mixing, its solute concentration, Cl* , being uniquely determined from
the local temperature and the phase diagram. On the other hand, due mainly to its relative large
diffusion length scale, the extra-dendritic liquid may be in a non-equilibrium state with respect to
the local solid-liquid interface, meaning that its average concentration Clc / e can be different from
Cl* . The consequent local undercooling Cl* − Clc / e will drive the c/e grain growth which in turn
will strongly influence the solute transfer in the extra-dendritic liquid. This complex coupled solute
transfer behavior is taken into account by the present model by introducing the diffusion length
scale δ lc / e− dc / e which quantifies the solute flux at the grain – extra-dendritic liquid interface. This
is a key parameter in modeling the c/e grain evolution since it controls the solute rejection in the
limited extra-dendritic liquid enveloping each c/e grain. The extra-dendritic liquid will undergo a
solutal enriching process reducing the local undercooling Cl* − Clc / e , to finally arrive at a state of
complete solute mixing and consequently to a complete stop of the grain growth. In this light, the
correct modeling of the diffusion length is of great importance for the correct simulation of the c/e
grain evolution and of the solidification process as a whole.
It is important to notice that the length scale δ lc / e− dc / e is strongly influenced by the length
scale of the c/e extra-dendritic liquid, that is R1*c / e − ac / e . So special care has to be taken in
modeling the c/e cell dimensions R1*c / e . It is noticeable that the present ensemble average model
gives a pertinent answer, allowing a rigorous approach to the mixed columnar/equiaxed zone. In
contrast with the present model, the Martorano et al. model [7] can deal with the columnar and the
equiaxed zone in a separate way only. The mixed zone which can eventually develop behind the
advancing columnar front is artificially approximated with a pure columnar zone. This implies as
well that the resulting mixed zone is resolved using the length scales characterizing uniquely the
columnar structure. This approximation leads to discontinuities regarding the variations of the grain
dimensions at the columnar front and, as one will see next, produce non-negligible
underestimations in the evaluation of the CET position.
Let us consider a simple directional solidification involving a constant thermal gradient G
and a constant pulling velocity V p (see Figure 2). Admitting for the moment that the columnar
front has reached a quasi-steady state, its velocity, Vcf would be therefore equal to V p . The local
undercooling at the columnar tip ∆Tc = Tliq − Tcf will result in an undercooled liquid melt ahead of
the columnar front (Figure 2), zone which is favorable to an eventual equiaxed nucleation. In this
respect, the columnar front will be forced to advance through a developed equiaxed zone, leaving
behind a mixed columnar/equiaxed zone. Let us consider for example that the grain density of the
equiaxed zone is ne and that of the columnar zone is nc0 > ne .

III-3
Submitted to Journal of Physics D: Applied Physics

Figure 2: A typical directional solidification and the subsequent columnar-equiaxed interaction

Using the Martorano et al. model [7] the mixed zone behind the CF is approached with an
equivalent pure columnar zone. In their model the position of the CF is solely needed in order to
decide what grain density to use in modeling the two zones. Hence, the columnar front will play the
role of a discontinuity surface for the variation of the grain density n , separating the pure equiaxed
zone ahead from the pure columnar zone behind. On the other hand all other average variables (i.e.
solid and grain fractions) are considered to be continuous. Since all length scales (i.e. grain
diameter a, interfacial area S g ) involved in the modeling of the solidification depend on the grain
density n , at the CF the variation of all these length scales will be discontinuous too. As we will
see further on, beside from being physically non-realistic, this approximation can lead to important
errors in the modeling of the mixed zone behind the CET.
In contrast with this approximate approach the present model takes into account the fact
that within this mixed zone columnar and equiaxed grains coexist sharing a same space. As pointed
* *
out previously, the coexistence state is quantified by means of two length scales, R1c and R1e , each
of them characterizing at x the amount of space available for a columnar or an equiaxed grain
respectively. The c/e grains evolution can be accounted separately and average variables can be
attached to each one of the two structures. Using this new approach, the columnar front is not
viewed anymore as a discrete separation between two different grain distributions because the
mathematical formulation of the model is valid for the pure equiaxed zone ahead of the CF as well
as for the mixed zone behind it. Indeed, the position of the columnar tips is only needed in order to
trigger the columnar “nucleation” within the equiaxed zone, this in turn being intrinsically
accounted by the model. Hence, no particular condition has to be used to commute from a single
grain distribution zone to a mixed one. It is important to notice that, since the equiaxed grain
density ne is conserved at the columnar front, the equiaxed grain diameter ae and the interfacial
area S ge will be conserved too. As one will note next, the use of ensemble model enables one to
quantify in a more precise way the limited solute diffusion processes in the c/e extra-dendritic
liquid. Finally it provides the user with accurate results on the mixed c/e zone behind the columnar
front.

2.2 The mathematical formulation

The mathematical model used in the following study is based entirely on the ensemble averaged
model thoroughly detailed in [1]. However, a simpler mathematical formulation is retained here
(see Table 1) due to several simplifying assumptions.

Table 1: Equation summary

Main balance Solid: Liquid:
equations

III-4
Submitted to Journal of Physics D: Applied Physics

k = {e, c} k = {e, c}
Mass ∂ ( ε sk ρ s ) ∂t = Γ sk (
∂ εf ρf ) ∂t = −Γ se − Γ sc
Energy
( ) (
∂ ⎡ ε s ρ s cs + ε f ρ f c f T ⎤ ∂ ⎡ ε s λs + ε f λ f ∂T ∂x ⎤
⎣ ⎦= ⎣ ⎦ + Γ +Γ L )
( se sc )
∂t ∂x
Solute ∂ ( ε sk ρ s Csk ) ∂Clk ⎛ ρ f S gk Dl ⎞
∂t
= Cs∗Γ sk ε lk ρ f
∂t
=⎜
⎝ δ lk − dk
− Γ gk ⎟ Cl∗ − Clk
⎠
( )
Supplementary
balance eq.
Inter-dendritic
mass balance
(
∂ ε dk ρ f ) ∂t = Γ gk − Γ sk
Grain balance ∂ne ∂t = ne0δ (Tliq − T − ∆Tne )

∂nc / ∂t = nc0 (1 − ε ge ) δ ⎡⎣ x − xcf ( t ) ⎤⎦

Inter-dendritic
solute balance
(C ∗
l ) ( )
− Cs∗ Γ sk = ε dk ρ f ∂Cl∗ ∂t + ρ f S gk D f Cl∗ − Clk ( )δ lk − dk

Auxiliary
expressions
Statistical ε le + ε de + ε se = ε e ε de + ε se = ε ge
constraints: , εc + εe = 1,
ε lc + ε dc + ε sc = ε c ε dc + ε sc = ε gc
Interfacial
concentrations:
⎧⎪kC ∗ if
Cs∗ = ⎨ l
Γs ≥ 0 Cl∗ = T − T f( ) m
⎪⎩Cs if Γs < 0
Equiaxed/column
ar fractions
( )
ε e , ε c = f ne , nc0 , ε ge , see equation (18)
Cell radius R1*k = ( 4 / 3π nk )
−1/ 3
ε k1/ 3
Grain radius ak = R1*k ε gk1/ 3
Grain envelope S gk = 4π ak 2 nk
interfacial area
Diffusion length ⎡ ⎤
⎢ 10ak 2δ k + 10ak δ k 2 + 4δ k 3 ⎥
δ lk − dk = δk 1 − where
⎢
⎣ (
5 R1*k 3 − ak 3 ⎥
⎦ )
⎧ D / w gk
⎪ l
δk = ⎨
if ( R − a ) ≥ 2( D / w )
*
1k k l gk

⎩ (
⎪ R1*k − ak / 2 if ) ( R − a ) < 2( D / w )
*
1k k l gk

Grain growth rate Dl m ( k − 1) Cl∗ 2

Γ gk = ρ f S gk w gk , w gk = ⎡ Iv −1 ( Ω k ) ⎤
π 2Γ ⎣ ⎦

First, since we will deal with unidirectional solidifications, only the 1D formulation will be
used hereafter. Moreover, in the solute balance equations, the macroscopical diffusion terms
( )
∇ ε pk D p ∇C pk will be neglected transforming the solute balance equations from PDE equations
into non-linear ODE equations. This can be quite easily justified [2] observing that the length
scales involved in the macroscopical diffusion terms are much larger than the ones involved in the
corresponding interfacial exchange terms. Moreover, this approximation allows a direct
comparison with the results of Martorano et al. model [7].

III-5
Submitted to Journal of Physics D: Applied Physics

Hence, for the case where no convection is considered, the model will include only mass
energy, solute and grain balance equations. Since the model takes into account a mixed zone,
separate balance equations are written for the columnar and equiaxed structure. First, concerning
the mass balance, two balance equations are written for the c/e solid phase (sc, se), two for the c/e
inter-dendritic liquid (dc, de) and one for the total fluid (f). Note that all these five equations are
simple ODE equations. Secondly, given the hypothesis of local thermal equilibrium the energy
balance can be approached with the help of a single mixture equation. Next the solute balance
equations are written for the solid, inter- and extra-dendritic liquid respectively and this for each
one of the two structures. Hence, a total of six equations need to be considered. Notice however
that all those six equations are ODE ones as well and that the solute balance in the c/e inter-
dendritic liquid is used solely for computing the c/e solidification rate, Γ sc , Γ se . Finally two grain
balance equations quantifying the equiaxed and columnar grain conservation have to be considered.
Notice as well that for the no convection case, these two equations are only quantifying the
corresponding c/e nucleation since no grain transport must be accounted. On one hand, the
equiaxed nucleation is modeled as a local event: the equiaxed grains are supposed to nucleate
instantaneously at a given undercooling ∆Tne . On the other hand and in contrast with the equiaxed
nucleation, the columnar “nucleation” (the sudden arrival of the columnar dendrite tips at x) cannot
be treated as a local event depending solely on local parameters. Indeed, the presence at x of the
columnar tips will be a function of its evolution history and therefore a front tracking technique
must be used in order to determine the precise position of the columnar front (referred hereafter as
to CF) at x and at time t. The model has to be completed therefore with the following equation,
describing the precise evolution of the CF

dxcf
dt
( )
= Vcf xcf (1)

where xcf is the columnar dendrite tips position and Vcf is their corresponding local velocity. The
later is modeled in a similar way as in reference [7] considering that

Vcf = w ge (2)

meaning that Vcf is equal to the local equiaxed tip velocity. Indeed, the columnar tips “nucleate”
within an equiaxed zone and more precisely within the equiaxed extra-dendritic liquid (1 − ε ge ).
Therefore the latter liquid will be shared between the two coexisting grain distributions. One can
logically assume that subsequent to the columnar nucleation the newly formed equiaxed and
columnar extra-dendritic liquids (le, lc) will share as well the same average concentration.
Accordingly, at the columnar tips one has

Clc = Cle (3)

the subsequent columnar and equiaxed undercoolings being equal as well. As already noted by
Martorano et al. [7] this is one of the most important feature of the model since it couples the CF
evolution with the equiaxed zone ahead of the columnar tips. If the equiaxed zone ahead of the CF
eventually arrives to a state of complete mixing, the columnar tips will be stopped, triggering the
CET. This is no more than the solutal interaction mechanism, as defined in [7].
It is important to notice that the formulation of the model in Table 1 can be used for the
pure columnar or equiaxed zone as well as for the possible mixed c/e zone. Moreover, no particular
condition has to be used to commute form one zone to another (i.e. from a pure equiaxed
formulation to a mixture one). In fact, the pure c/e formulation is only the limit of the model in
Table 1 for nc / e → 0 and ε c / e → 0 . The exact position of the CF, xcf , is used only to trigger the

III-6
Submitted to Journal of Physics D: Applied Physics

columnar “nucleation” within the equiaxed zone. A mixed columnar+equiaxed zone will therefore
be implicitly modeled behind the CF.
Finally the model has to be completed with several auxiliary relations, expressing
statistical and phase diagram constraints as well as with the length scales involved in the cell model
approach: c/e cell dimensions R1*c / e , grain diameters ac / e , interfacial areas S gc / e and diffusion
length scales δ lk − dk . Moreover, the c/e grain growth rates Γ gc / e and the local c/e tip velcities need
to be considered.

2.3 Numerical procedure

The mathematical model to be solved contains five mass balance equations, one energy equation,
six solute balance equations and finally two grain balance equations. Despite the increase number
of equations to be solved, one should note however that only the energy equation represents a PDE
equation, the rest of them being relative simple ODE equations. To the model in Table 1 one
should add the columnar front tracking equation (1), an ODE too. The equations are discretised
using a fully implicit control volume method [8]. The implicit first order Euler method is used for
the time discretisation. The system of equation in Table 1 together with equation (1) is solved in a
segregated manner: first the energy equation is treated and the resulting tri-diagonal system is
solved using a direct method by means of a simple LU decomposition. It must be noted that due to
the latent heat release, the energy equation has a highly non-linear character and therefore, special
care must be paid to the discretisation of the latent heat source. This one is implicitied using a
special numerical implicitation algorithm as described in Patankar [8]. Then the conditions for the
equiaxed and columnar nucleation are analyzed and the grain balance equations are solved together
with the computation of all parameter needed for the modeling of an eventual mixture zone. Next
the solute balance equations are solved. Note that the solute diffusion equations in the extra-
dendritic c/e liquid have as well a high non-linear character In addition, for problems involving
high cooling rates of small grain densities (large R1c / e ), they become stiff equations too. Hence,
special care must be paid to the discretisation of the interfacial sources in the extra-dendritic solute
balance equations. On one hand the interfacial sources are always treated in an implicit way. On the
other hand the stiff behavior of these two solute balance equations requires very small time steps in
order to converge properly. To limit the computational time an adaptive time scheme is used in the
following study. First, the time line is discretised uniformly using an initial time step ∆t (see
Figure 3).

Figure 3: The adaptive time step scheme used in the current study

If the problem does not converge at the nth time iteration, the current time step
∆t n = t n +1 − t n is divided into Nt smaller time steps and the simulation is retaken form the last
converged time position, that is t n . Since this algorithm is recursive, meaning that the time refining

III-7
Submitted to Journal of Physics D: Applied Physics

procedure can be applied to an already refined time step, a minimum time step ∆tmin is imposed,
below which no time refining is allowed. A typical Nt = 5 was found to give satisfactory results
for the problems analyzed subsequently.

Figure 4: The adaptive mesh scheme used in the current study

As pointed out, a front tracking technique has to be used in order to determine the precise
position of the columnar front. The approach used in this study is very similar with the one used by
Martorano et al. [7]. The equation (1) is discretised as follows

( xcf )i = ( xcf ) ( )i ∆t
n n −1 n
+ Vcf (4)

where n is the current time step and i the current iteration. Note that, since Vcf is strongly
depending on local parameter like temperature and solute concentrations, the above equation is
solved iteratively together with all other balance equations. In addition a mesh refining technique is
used for the control volumes close to the columnar front. This procedure is required due to the fact
that the columnar-equiaxed interactions result into high solutal and solid/grain fraction gradients
close to the columnar front. The mesh adaptation scheme is identical with the one used in reference
[7]: the three closest control volumes to the CF are each refined into N x smaller cells. This is done
in a dynamic way (see Figure 4) as the columnar front is moving from one control volume to
another.
Finally the whole numerical algorithm is implemented and solved within Matlab®
computing environment, taking thus advantage of its powerful matrix computing procedures.

3 The quasi-steady state CET - The CET maps.

Before attempting to fully model the unidirectional solidification experiments [4, 5], the model
described in the previous section is used first to analyze the simpler case of a quasi-steady state
CET. Hence, a directional solidification, involving a constant vertical gradient (G) and a constant
pulling velocity ( V p , equal as well to the isotherm velocity Vis ) will be studied in the following, in
an attempt to compare the new ensemble average model with the Hunt [9] and Martorano et al. [7]
model.
By imposing a constant thermal gradient and a constant pulling velocity the temperature
field in the ingot will be imposed too. Indeed the temperature distribution becomes

T ( x, t ) = T0 − G.V p .t + G.x (5)

III-8
Submitted to Journal of Physics D: Applied Physics

where T0 is the initial temperature at the bottom of the ingot, that is T ( 0,0 ) . Since the ingot
cooling starts from a pure liquid state, T0 must be larger than Tliq (the liquidus temperature
corresponding to the initial uniform liquid concentration C0 ). Since for this particular case an
eventual overheating of the liquid would not influence the subsequent solidification process it is
customary to consider [7] that T0 = Tliq . Knowing a priori the temperature field, the energy
equation in Table 1 doesn’t need to be considered anymore.
It is assumed as well that the columnar front starts forming at the bottom wall at the
liquidus temperature, that is at t = 0 . Since its velocity Vcf is initially close to 0 and logically

( )
smaller than the local isotherm velocity Vis ≡ V p , the temperature at the columnar front Tcf will
progressively decrease increasing consequently the local undercooling ∆Tcf = Tliq − Tcf and the
columnar tip velocity Vcf . If no equiaxed grains are supposed to nucleate within this increasing
undercooled zone ahead of the CF, the columnar tips will finally reach a quasi-steady state when
their velocity equals the isotherm velocity Vis . Indeed, when Vcf = Vis , the temperature at the
columnar tips does not change anymore and the corresponding undercooling ∆Tcfsteady remains
constant. Therefore, the columnar front is advancing undisturbed and no CET phenomenon
appears. However, when equiaxed grains are allowed to nucleate within the undercooled zone
ahead of the CF, the interaction between this equiaxed zone and the advancing columnar front
needs to be analyzed in order to determine if whether or not CET occurs in the ingot. Let us
consider that equiaxed grains nucleate at a given undercooling ∆Tne below the local liquidus
temperature. As pointed out, as the cooling of the liquid progresses ( t > 0 ), the columnar front
velocity and the undercooling at the columnar tips ∆Tcf increase as well. If ∆Tcf becomes larger
that ∆Tne the conditions for the equiaxed nucleation within the undercooled melt ahead of the CF
are established and the columnar tips will be forced to advance through a developed equiaxed zone.
Given some particular conditions the latter will eventually stop the advancing of the CF and trigger
the CET. One should therefore determine these particular conditions producing the CET
phenomena.
In this context, the ensemble model will be used to simulate this particular unidirectional
solidification and more precisely the complex solutal and mechanical interactions between the
advancing CF and the equiaxed zone ahead. Since we are interested in obtaining uniquely CET
transition maps specifying whether or not a CET takes place given G, V p and ∆Tne , one does not
need to model the whole ingot evolution but only the evolution of the columnar front and of the
eventual equiaxed zone ahead. Indeed, by imposing the temperature field and by neglecting the
macroscopic solute diffusion terms in the solute balance equations one decouples the equiaxed
zone ahead of the CF from the columnar/equiaxed mixed zone behind the CF. This will greatly
simplify the modeling task since by transforming the solute balance equations into ODE ones, the
pure equiaxed zone ahead of columnar tips can be uniquely determined from the local temperature.
Thus one has to model first the evolution of a pure equiaxed zone for a given ∆Tne , cooling rate
CR and grain density ne . The complete state of the equiaxed zone ( ε ge , ε se , Cle ) will be therefore
known for any temperature T below Tliq . Since the local temperature at the CF, Tcf , is known
from the exact position of the columnar front xcf , the equiaxed state at the columnar front can be
computed as well, knowing Tcf .
Therefore, one has to resolve first the pure equiaxed zone in front of the CF knowing that
locally the temperature distribution is given by

Teq ( t ) = T0 + CR.t (6)

III-9
Submitted to Journal of Physics D: Applied Physics

A complete description of the equiaxed structure is therefore obtained, depending solely on CR,
∆Tne , ne and on the local temperature Teq :

ε ge , ε se , Cle , w ge = f (Teq , ne , ∆Tne , CR ) (7)

To model now the columnar front evolution and an eventual CET, one has only to solve the front
tracking equation (1) in which the local columnar tip velocity is linked directly to the local
equiaxed tip velocity through the local temperature Tcf

(
Vcf = w ge Tcf , ne , ∆Tne , CR ) (8)

3.1 The equiaxed zone ahead of the columnar front

Let us first analyze the purely equiaxed case. As already discussed, the pure equiaxed zone ahead
of the CF is only a limiting case of the complete model in Table 1 for nc = 0 and ε c = 0 . The
columnar averaged equations and the energy equations being discarded one will only have to solve
the equiaxed solute and mass balance equations where ε e = 1 . As pointed out in [1], this set of
equations is equivalent with the Wang and Beckermann model [2] excepting the solute diffusion
length δ l − d which in the ensemble averaged model is approached from an approximate model. In
this light a validation of the new proposed diffusion length has to be assessed first. For reasons of
clarity, the comparison with the Wang and Beckermann model [2] is presented in the Appendix A.
It is worthwhile to notice that despite its simplicity, the approximate model gives very accurate
results. The agreement with the integral approach in [2] for cases involving high cooling rates or
small grain densities is very satisfactory. In contrast with the latter case, for small cooling rates or
high ne , the results deviate considerably from the Wang model [2]. This is due mainly to the fact
that the integral approach in [2] does not respect the zero solute flux condition at the cell boundary,
and therefore overestimates the enriching process of the extra-dendritic liquid.

Figure 5: Extra-dendritic liquid solute concentration and tip velocity evolution for the equiaxed
grains ahead of the CF (Al-3wt%Cu, CR=0.1K/s, ∆Tne = 1 K , ne = 107 , 109 , 1012 m−3 )

As noted in Appendix A the equiaxed grain density ne has significant influence on the
subsequent equiaxed solidification, given a constant CR and ∆Tne . Therefore it is instructive to
analyze in more detail the influence of ne on the pure equiaxed zone evolution and consequently
on the CET. The equiaxed solidification of an Al-3wt%Cu alloy is analyzed in the following for a

III-10
Submitted to Journal of Physics D: Applied Physics

constant cooling rate of 0.1 K/s, a equiaxed undercooling of 1K and for three different grain
densities: ne = 107 , 109 , 1012 m−3 . In Figure 5 and Figure 6 the main parameters of the equiaxed
grains are plotted against the local temperature of the undercooled zone ahead of the CF. Notice
that with the use of (6) one can directly obtain their time evolution as well. Two main sub-zones
can be identified: a pure undercooled liquid melt between the equiaxed nucleation temperature
Tne = Tliq − ∆Tne and Tliq within which no equiaxed nucleate and the developed equiaxed zone for
T < Tne . Note again that for the moment we ignore the columnar zone and we analyze only the
equiaxed structure. Hence, when the local temperature reaches Tne , equiaxed nuclei germinate
homogeneously with a grain density ne . As pointed out in [1], this would be also equivalent with
−1/ 3
an average distance between nuclei of R1e = ( 4 / 3π ne ) .

Figure 6: Solid and grain fraction evolution for the equiaxed zone ahead the CF (Al-3wt%Cu, CR =
0.1 K/s, ∆Tne = 1 K , ne = 107 , 109 , 1012 m−3 )

Let us focus for the moment on the first case involving the smallest grain density
( ne = 107 m −3 ), and thus the largest distance between adjacent grains, R1e  2.9 mm . Since the local
temperature undergoes a continuous decrease in time, the local solutal undercooling ∆Ce = Cl* − Cl
will increase as well due to the fact that the mean concentration in the extra-dendritic liquid still
remains close to the initial one C0 (see Figure 5). This in turn is due to the fact that the newly
formed nuclei reject only small quantities of solute in the extra-dendritic liquid in view of their
small exchange interfacial area ( S ge = ne 4π ae 2 ). The increasing local undercooling results into
increasing local tip velocity (Figure 5) and rapidly, the initial fully solid nuclei evolve into highly
dendritic grains for which ε se < ε ge (Figure 6). Note however, that for this first case the
concentration of the extra-dendritic liquid remains close to C0 even for high grain fractions and
consequently for relative high interfacial exchange areas. This can be explained by the fact that for
high dendrite tip velocities, the local Peg ( Peg = w ge ( R1e − ae ) / Dl ) is higher than 1 even for large
ε ge . As pointed out in [1] on has

δ l −d 1
 (9)
R1e − ae Peg

III-11
Submitted to Journal of Physics D: Applied Physics

meaning that the solutal boundary layer surrounding the grain envelope is much smaller than the
length scale of the extra-dendritic liquid. Basically this means that the solutal interactions between
two adjacent grains are negligible. Given these conditions, the liquid will remain close to C0 as
long as δ l − d  R1e − ae . Only when the solutal boundary layer around the grain arrives at the cell
boundary, that is when

δ l − d  R1e − ae (10)

the enriching of the equiaxed extra-dendritic liquid becomes non-negligible. Since for high cooling
rates δ l − d  R1e − ae (see equation (9)) one obtain that the enriching process becomes important
for grain diameters ae  R1e , that is for equiaxed grain fractions ε ge close to unity. Indeed looking
at Figure 6 one can see that the enriching of the extra-dendritic liquid becomes important below a
temperature T * corresponding to a grain fraction ε *ge of order of 1. Above this limit the extra-
dendritic liquid undergoes a rapid enriching process (Figure 5) having as consequence a rapid drop
in the local undercooling. The latter will finally vanish ( Cl  Cl* ) and the grain growth is stopped.
Subsequently the solidification will continue within a state of perfect solutal mixing, that is
following a Scheil law. It is important to notice as well the particular variation of the local tip
*
velocity which reaches a maximum value, w ge at the critical temperature T * (Figure 5). Since the
local equiaxed velocity is equal with the columnar tip velocity if a columnar tip would be present
there, it is clear that the particular variation of the equiaxed tip velocity would also influence the
columnar front evolution and eventually the CET phenomena.
If the second case is analyzed ( ne = 109 m −3 ; R1e  620 µ m ) one can observe that due to
somewhat larger interfacial areas S ge , the extra-dendritic liquid is enriching faster than the
previous case. Basically due to a larger ne and smaller average distance between grains, R1e , the
solutal interactions between adjacent grains are more important. This will determine smaller
characteristic Peg numbers and consequently a faster solutal enriching process. The latter becomes
non negligible for smaller equiaxed grain fractions ( ε *ge ) and is characterized by a smaller
*
maximum tip velocity w ge (Figure 5).
A limit case is the third one ( ne = 1012 m −3 ; R1e = 62 µ m ) which corresponds to a highly
grain refined alloy. The solutal interactions between two adjacent grains are so strong that the
extra-dendritic liquid is enriched from the very beginning of the equiaxed solidification. As seen in
Figure 5, Cle equals Cl* soon after the equiaxed nucleation. Moreover, the equiaxed grains solidify
in a globulitic manner ( ε ge = ε se ) following a Scheil law (Figure 6).
Knowing now the exact evolution of the pure equiaxed zone ahead the columnar front with
respect to the CR , ne and ∆Tne one is able to analyze its influence on the advancing CF and
consequently on the CET. As discussed earlier, the equiaxed grains reject solute in their extra-
dendritic liquid and decrease consequently the local undercooling at the dendrite tip. This would
influence of course the evolution of the columnar tips if they were present within the equiaxed
zone. This effect is no more than the solutal blockage mechanism of the equiaxed grains on the
columnar tips, just as it was identified in [7].
On the other hand it is expected that due to the finite equiaxed grain dimension, the
equiaxed structure would partially obstruct the columnar tips in their advance, producing a kind of
“mechanical blockage” effect on the columnar front, somewhat similar with what Hunt [9]
predicted. Before presenting the full modeling of the columnar front advance, it would be
instructive to analyze in detail the solutal and the mechanical blocking effects.

III-12
Submitted to Journal of Physics D: Applied Physics

3.1.1 Solutal blockage effect. Now we consider both types of solidification, e. g., columnar and
equiaxed. Basically we are interested on the particular conditions describing the unidirectional
(
solidification V p , G , ne , ∆Tne ) under which a CET can be triggered. First, we have assumed that
the columnar tips and the equiaxed zone at the CF share a same undercooling and consequently a
( )
same tip velocity Vcf = w ge . Secondly, by neglecting the macro-diffusion terms in the solute
balance equations, the pure equiaxed zone ahead of the CF is completely decoupled from the
evolution of the zone behind the CF. Hence, knowing the local temperature at the columnar tips
Tcf , one can completely describe the equiaxed state at the CF. Since the equiaxed zone ahead of
the CF is completely determined from the CR , ne and ∆Tne , one has to analyze only the influence
of the pulling velocity V p on the CET phenomenon. In turn, the thermal gradient is directly
resolved from the CR, G = CR / V p . Let us focus on the second equiaxed case analyzed in the
previous section ( ne = 109 m −3 ). It is assumed first that the pulling velocity V p is smaller than the
*
maximum equiaxed velocity w ge (Figure 7).

Figure 7: Solutal blockage effect of the equiaxed zone ahead the CF (Al-3wt%Cu, CR=0.1 K/s,
∆Tne = 1 K , ne = 109 m −3 )

As pointed out, the columnar grains nucleate at the bottom wall when the local temperature
reaches the liquidus one Tliq , that is at t = 0 , given the initial conditions in (5). Initially the CF
velocity will be close to 0 and obviously smaller that the local isotherm velocity Vis = V p . ( )
Consequently the columnar tip temperature Tcf will decrease and the local undercooling ∆Tcf and
Vcf at the CF will increase. At the very beginning the undercooling ∆Tcf is lower than the
equiaxed nucleation undercooling ∆Tne and no equiaxed grains exist in front of the CF. The CF is
advancing in a pure undercooled melt (Figure 7, see the columnar evolution path coupled with the
w ge evolution). The CF undercooling increases further until it reaches the equiaxed nucleation
undercooling ∆Tne . At this precise moment equiaxed grains nucleate instantaneously at the CF. Let
0
us denote with w ge the equiaxed tip velocity corresponding to the nucleation undercooling ∆Tne .

( 0
)
Since Vcf = w ge is still smaller than the isotherm velocity, Tcf decreases further and the equiaxed
zone ahead of the CF becomes larger too. The columnar tips will be forced to advance through a

III-13
Submitted to Journal of Physics D: Applied Physics

developed equiaxed zone (Figure 2). Notice that for the moment Tcf > T * , and therefore to a
decrease of the CF temperature it corresponds an increase of both the local undercooling ∆Tcf and
tip velocity Vcf . This is due to the fact that equiaxed grains ahead the CF don’t reject yet enough
solute in their extra-dendritic liquid and consequently Cle remains close to C0 . In turn, the
increase in the CF velocity is continuously decreasing the difference between Vcf and the local
isotherm velocity V p . When Vcf becomes equal to V p the CF reach a quasi-steady state. Indeed
following this moment the CF temperature does not change anymore and the local undercooling at
the tip remains unchanged too. Hence, the CF velocity reaches the quasi-steady state.
Therefore for this particular choice of the pulling velocity, the CF reaches a quasi-steady
state and, neglecting for the moment the mechanical blocking effect of the equiaxed grains, a
mixed columnar/equiaxed structure will be identified in the ingot.
0
Let us analyze now the case V p < w ge . Using a similar reasoning as for the previous case,
one can couple the CF evolution with the one of the equiaxed tip velocity (Figure 7). Following the
CF evolution path one can observe that the condition for the existence of a quasi-steady state at the
columnar tips, Vcf = V p , is reached before the equiaxed grains have the possibility to nucleate
within the undercooled melt ahead of the CF. In this case the undercooled zone ahead of the CF is
not sufficiently large to allow for equiaxed nucleation. Thus, the CF will advance undisturbed in a
pure undercooled melt and no CET can be identified. A fully columnar structure will be identified
in the ingot.
*
Now let us consider a third case for which V p > w ge . Following again the evolution path of
the CF (Figure 7) it is observed that the CF velocity increases until it reaches the maximum
*
equiaxed velocity w ge . Since Vcf is still smaller than the local isotherm velocity V p , Tcf will
continue to decrease. However, at this level the equiaxed grains would have already started to
enrich the extra-dendritic liquid and diminish the local undercooling (Figure 7). Therefore, to a
decrease in the local CF temperature will correspond now a decrease in local tip velocity (Figure
7). The CF velocity will therefore decrease below w ge < V p
*
( ) and the CF temperature, Tcf , will
undergo an even faster decrease. The columnar tips are slowed down by the equiaxed zone ahead
and will be finally stopped, Vcf → 0 . Hence, no quasi-steady state establishes at the CF and a CET
occurs. Basically, the equiaxed zone is solutally blocking the columnar front due to the important
solutal enriching of the extra-dendritic liquid. It is important to notice that this effect is no more
than the solutal blockage mechanism as it was identified by Martorano et al. [7]. Finally a fully
equiaxed zone will be observed in the ingot.
Therefore for a given CR , ne and ∆Tne the CET is strongly influenced by the choice of
the pulling velocity V p . Three cases were identified:
0
ƒ a fully columnar structure in the ingot for V p < w ge ;
0 *
ƒ a mixed c/e structure for w ge ≤ V p ≤ w ge ;
*
ƒ a fully equiaxed structure for V p > w ge .
0
It is important to notice that w ge depends solely on the choice of the equiaxed undecooling ∆Tne ,
* *
whereas w ge depends strongly on CR and ne . In Figure 8 the dependence of w ge on the choice of
*
CR is analyzed. Basically, decreasing the CR would decrease the w ge too, limiting thus the
* 0
difference w ge − w ge and conditions for obtaining a quasi-steady mixed zone. This is due mainly to
the fact that for small cooling rates the solute diffusion process at the grain envelope has enough

III-14
Submitted to Journal of Physics D: Applied Physics

time to enrich the extra-dendritic liquid. Hence, the enriching of the extra-dendritic liquid will
become non-negligible for larger local temperature (Figure 8).

Figure 8: Influence of the cooling rate on the equiaxed zone ahead the CF (Al-3wt%Cu,
∆Tne = 1 K , ne = 109 m −3 )

3.1.2. Mechanical blockage effect. As seen given certain solidification conditions such that
0 *
w ge ≤ V p ≤ w ge , a particular CET can be identified, involving a mixed columnar/equiaxed state at
the columnar tips. To characterize more precisely this zone the mechanical blockage effect of the
equiaxed zone on the CF should be taken into account. Basically, one would have to quantify the
mechanical interactions between the advancing CF and the developed equiaxed zone ahead. As
pointed out in [1], due to its statistical approach, the new ensemble model can take into account
these types of interactions and characterizes in a rigorous way the mixed c/e zone behind the CF.
Let us consider an advancing columnar front characterized by a grain density nc0 (linked to
the primary arm spacing) and a developed equiaxed zone ahead having a grain density ne and a
non-zero grain diameter ae . This would correspond to an equiaxed grain fraction
( )
ε ge = ne 4 / 3π ae3 . As pointed out in [1] the columnar and the equiaxed grain distributions
interact “mechanically” following a two fold mechanism.

First mechanical blockage

The first mechanical blockage effect of the equiaxed grains will determine a rarefaction of the
columnar zone when entering (“nucleating”) into a developed equiaxed zone. This will decrease the
initial columnar grain density nc0 with a factor if (1 − ε ge ), (see as well the columnar grain balance
in Table 1). Indeed due to the finite equiaxed grain fraction the columnar tips have only a limited
space (1 − ε ge ) within which they can “nucleate”. Hence only a fraction of the total number of
columns will be able to enter the equiaxed zone.

III-15
Submitted to Journal of Physics D: Applied Physics

Figure 9: The first mechanical blockage: observe the rarefaction process of the columnar zone.

It is clear now that, given no redensifying mechanism for the columnar tips, the advancing
columnar front will undergo a continuous rarefying process, its grain density nc constantly
decreasing in time. To quantify this process it would be instructive to consider the simple situation
in Figure 9. Since the equiaxed zone ahead of the CF has an uniform grain density ne , the average
−1/ 3
space between two adjacent equiaxed grains is no more than R1e = ( 4 / 3π ne ) . If the CF arrives
at a point x within the equiaxed zone the columnar “nucleation”, as defined in [1], triggers within
the developed equiaxed zone. A mixture c/e zone is therefore formed. As pointed out the columnar
grain density which corresponds to the columnar “nucleation” event can be computed as follows:

(
nc ( x ) = nc0 1 − ε ge ) (11)

where nc0 is the nucleation grain density of the columnar grains. This is of course valid when it is
assumed that a homogeneous spherical grain population nucleate within the equiaxed zone. In
reality the columnar grains are not spherical ones but highly anisotropic grains developed along the
x direction. However, as pointed out in [1] one can approach, to a first approximation, the columnar
zone with an equivalent spherical model for which the two main characteristics, the average
spacing between the particles λ1 and the grain fraction ε gc are inherited from the real columnar
zone. In this way a pure columnar zone can be locally approximated with a homogeneous spherical
grain distribution, having a grain density computed from the primary arm spacing

−1
⎛4 ⎞
nc0 = ⎜ π λ13 ⎟ (12)
⎝3 ⎠

This is somewhat equivalent with transforming the columnar zone into an equivalent equiaxed one.
However this is a particular “equiaxed” zone since one must have in mind the fact that between
these columnar spherical particles there is a strong statistical correlation: first no mobility of the
grains is allowed and secondly an axial gradient of the local grain fraction ε gc would be equivalent
with a local variation of the columnar grain diameter along x direction. Notice that in (12) λ1 can
be expressed with respect to the local thermal gradient G and isotherm velocity Vis by means of
various models existing in the literature [10, 11].
Returning to the particular case of a columnar front advancing within a developed equiaxed
zone, the question to be answered is what the pertinent value for nc0 in equation (11) would be. In
reference [1] it was proven that, from a statistical point of view, nc0 represents the conditional

III-16
Submitted to Journal of Physics D: Applied Physics

probability of having a columnar grain nuclei at x given no equiaxed grain centered within the
sphere centered at x and of radius ae , that is, given no equiaxed grain touching the point x. In this
light nc0 appear to be correlated with the columnar grain density nc characterizing the columnar
zone at a certain distance behind the nucleation site x and outside the radius of influence of the
closest equiaxed grain to x. Knowing the length scale characterizing the equiaxed zone at x, i.e. the
average distance between equiaxed particles R1e , one expects that nc0 to be correlated with the
grain density characterizing the columnar zone at its previous position x − R1e (Figure 9), that is
just before entering the equiaxed zone at x. One expects thus that

nc0 ( x )  nc ( x − R1e ) (13)

Due to the isotropic approximation used to characterize the columnar zone, the above
equation must be seen only as a first approximation to the columnar nucleation grain density nc0 . A
more detailed analysis could be made if we consider the interactions between an advancing
anisotropic columnar grains and the developed isotropic equiaxed zone ahead of the columnar
front. However, this approach would be more complex since one would have to deal with the
statistical coupling between a cylindrical model (2D approach) and a spherical one (3D approach).
Note that, despite the simplified approach used above, two important aspects are intrinsically
embedded in equations (11) and (13). Firstly, due to the finite dimension of the equiaxed grains the
columnar grain density decrease with a factor proportional with the equiaxed grain fraction ε ge .
Secondly, the length scale characterizing the distance over which nc0 is decreasing with a factor
equal to 1 − ε ge is no more than the equiaxed length scale R1e . These two information help us to
quantify the decrease rate of the columnar grain density, that is to say the rarefaction rate of the
columnar zone. Indeed, for a given equiaxed grain fraction ε ge and a constant columnar front
velocity Vcf advancing through the equiaxed zone the decrease rate of nc becomes

[ nc ( x) − nc ( x − R1e )] / nc ( x − R1e ) = −ε Vcf

( R1e / Vcf ) ge
R1e (14)
= −ε geVcf ( 4 / 3π ne )
1/ 3

Notice that the decrease rate of nc is computed with respect to the time scale τ = R1e / Vcf .
From (14) it is easy to see that the columnar density nc will decrease more rapidly for a denser
equiaxed zone (larger ne ), for a more developed equiaxed zone (larger ε ge ) and for a faster
advancing CF (larger Vcf ).

Second mechanical blockage effect

As pointed out in reference [1], the second mechanical blockage effect of the coexistence reflects
the fact that for a mixed zone (i.e. columnar + equiaxed), the two coexisting structures are forced to
share a same space (Figure 10). Consequently, the space available for each structure will be
different and smaller too from the corresponding space available just before the coexistence state.

III-17
Submitted to Journal of Physics D: Applied Physics

Figure 10: The second mechanical blockage: the full line circles represent the two cells
characterizing the mixed c/e zone and the dashed line circles are the cells characterizing the two
structures (c/e) separately, that is just before the coexistence state

As discussed already, the space characterizing the mixed c/e structure is quantified in the
ensemble model by means of two corresponding spherical cells of radius R1*c / e (Figure 10).
Basically the second mechanical blockage states that the length scales R1*c / e are smaller than the
corresponding length scales characterizing the c/e state separately before their coexistence, namely
R1c / e . Notice again that R1*c / e together with their corresponding grain densities nc / e are of primary
importance for the modeling of the mixed c/e zone. Indeed these parameters characterize
completely the statistics of the coexistence state, enabling one to define average variables (i.e. c/e
fractions, average concentrations, etc) for each one of the two structures. Moreover, since R1*c / e
can be very different form the original length scales R1c / e the evolution of the mixed zone can be
significantly different from that of an equivalent pure columnar or equiaxed zone.

3.2 Numerical results

Let us now fully resolve the columnar front evolution for an Al-3wt%Cu alloy, a cooling rate of
CR = 0.1 K / s , an equiaxed grain density ne = 109 m −3 and an equiaxed nucleation undercooling of
1 K. As pointed out, the equiaxed zone ahead of the columnar tips is fully determined for a given
CR, ne and ∆Tne (Figure 7). Let us first consider the case where the pulling velocity
*
V p = 8 ⋅ 10−5 m / s is smaller than w ge . As pointed out one has only to solve the CF tracking
equation (1) knowing that the columnar front velocity is equal to the local tip velocity
characterizing the equiaxed grains at the CF, equation (2). Moreover, the two mechanical blockage
effects of the equiaxed grains, as detailed in the previous section need to be considered as well.

III-18
Submitted to Journal of Physics D: Applied Physics

a) b)
Figure 11: a) The CF temperature and the CF velocity evolution; b) The evolution of the equiaxed
grain fraction, the columnar fraction and the columnar grain density at the CF (Al-3wt%Cu, CR =
0.1 K/s, ∆Tne = 1 K , ne = 109 m −3 , V p = 8 ⋅ 10−5 m / s )

The evolution of several parameters at the CF is presented in Figure 11. One can see
(Figure 11a) that the CF velocity Vcf start to increase from the very beginning ( t = 0 ). Indeed we
supposed that the columnar grains nucleate on the bottom wall when the local temperature reaches
Tliq , that is at t = 0 given the boundary conditions. Due to the fact that its initial velocity is smaller
than the local isotherm velocity ( Vis = V p ), Tcf decreases and consequently Vcf increases. As
observed, initially (Figure 11b), the columnar structure conserves its initial grain density nc0 (cf.
equation (12)), since no equiaxed grains exist ahead of the CF. For the same reason, initially
(Figure 11b) the columnar fraction ε c remains equal to unity, meaning that the columnar tips are
advancing within a pure undercooled melt; no coexistence with another structure exists. This first
stage in the evolution of the CF continues until Tcf becomes equal to Tne , the equiaxed nucleation
temperature. Starting with this moment ( tne ) an equiaxed zone develops in front of the CF and the
columnar tips will be forced to advance within an increasingly developed equiaxed zone. Note
again that the CF velocity Vcf is computed form the local equiaxed tip velocity, equation (2). As
expected, since Vcf is still smaller than Vis , Tcf decreases further and Vcf increases. This increase
is due to the fact that the equiaxed grains have not yet enriched the extra-dendritic liquid
( Cle  C0 ). The tip velocity continues to increase until Vcf reaches the isotherm velocity Vis .
Above this moment ( tst ) the columnar tip velocity and the tip temperature Tcf don’t change
anymore (Figure 11a). Indeed, from the point of view of Tcf and Vcf the CF reaches a quasi-steady
state. Note again that this quasi-steady state is possible due to the negligible solutal interaction
between the equiaxed zone and the CF.
On the other hand the mechanical interactions cannot be neglected. Starting with the
moment tne when Tcf = Tne the CF will enter into an increasingly developed equiaxed zone.
Indeed, the equiaxed grain fraction ε ge at the CF (Figure 11b) is continuously increasing
obstructing the columnar tips. As discussed earlier two mechanical blockage effects of the equiaxed
grains on the CF can be identified. Firstly, the finite size equiaxed grains will lead to a rarefaction
of the columnar tips, equation (11). As observed (Figure 11b) nc is continuously decreasing in
time for t > tne and if no other redensifying mechanism is considered, the columnar zone will
finally fade away into the equiaxed one.

III-19
Submitted to Journal of Physics D: Applied Physics

Secondly, the columnar “nucleation” within an equiaxed zone leads to a sharing of space
between the two coexisting structures (columnar and equiaxed). As discussed, the way the two
structures share a same space is reflected by the two fractions ε c / e and the two lengths scales
R1*c / e . Basically the coexistence state is characterized by a smaller average distance between
particles ( R1*c / e ) compared with the ones valid if no coexistence between particles would be
considered ( R1c / e ). As observed (Figure 11b) at tne , that is when the columnar tips enters into the
newly formed equiaxed zone a drop in the columnar fraction takes place. Indeed, at the tne freshly
new equiaxed nuclei and the columnar tips are forced to share a same space. Consequently the
columnar fraction ε c , a function of ne , nc0 and ε ge = 0 (Table 1), becomes smaller than 1. Since
the CF is advancing into an increasingly developed equiaxed zone (larger ε ge ) ε c will decrease
further. Notice that this effect is amplified by the fact that nc0 decreases too due to the first
mechanical blockage effect.
The first mechanical blockage effect has another important consequence on the CET. As
pointed out due to negligible solutal interaction between the CF and the equiaxed zone ahead Vcf
and Tcf reach a quasi-steady state when Vcf becomes equal to Vis . For t > tst all parameters
characterizing the equiaxed state at the CF becomes constant with respect to time. Despite this fact,
due to the first mechanical blockage effect nc will continue its decrease even after tst . Hence,
given no other redensifying mechanism for the columnar grains, no quasi-steady state at the CF can
be identified. If the ingot is sufficiently long nc will become negligible with respect to ne , ε c will
drop to zero and the columnar zone will finally disappear favoring the pure equiaxed structure.
Finally a fully equiaxed structure will be observed in the ingot.

a) b)
Figure 12: a) The CF temperature and the CF velocity evolution; b) The evolution of the equiaxed
grain fraction, the columnar fraction and the columnar grain density at the CF (Al-3wt%Cu, CR =
0.1 K/s, ∆Tne = 1 K , ne = 109 m −3 , V p = 1.5 ⋅ 10−4 m / s )

*
Let us now consider a pulling velocity larger than w ge , i.e., V p = 1.5 ⋅ 10−4 m / s . As noted
before, due to initial negligible solutal interactions between equiaxed grains and due to the fact that
the local CF velcity is initially smaller than V p , Tcf decrease and Vcf increases in time (Figure
*
12a). Vcf increases until it reaches the maximum equiaxed velocity w ge , that is at t = tmax (Figure
*
12a). Since w ge is still smaller than the local isotherm velocity V p , Tcf will decrease further.

III-20
Submitted to Journal of Physics D: Applied Physics

However, at this level, due to the non-negligeble solutal interaction between equiaxed grains, to a
decrease in the local CF temperature it corresponds a decrease in local tip velocity (Figure 7). The
*
( )
CF velocity will therefore decrease below w ge < V p . The columnar tips are slowed down by the
equiaxed zone ahead and will be finally stopped ( Vcf → 0 ) at t = tCET (Figure 12a). Thus no quasi-
steady state establishes at the CF; and a CET occurs. A fully equiaxed zone will further occupy the
ingot. Notice that the constant equiaxed grain fraction for t > tCET (Figure 12b) reflects in fact the
fully developed equiaxed zone at the CF. Since the CF is stopped, the columnar grain density at the
CF remains constant in time as well (Figure 12b).

3.2.1. Redensifying mechanism of the columnar zone. In reality the columnar tip density nc is not a
simple imposed parameter as ne is, but derives from a complex behavior of the columnar tips
advancing into an undercooled melt. As already pointed out the primary arm spacing λ1
characterizing a quasi-steady columnar front is mainly controlled by the local thermal gradient G
and the local isotherm velocity Vis at the columnar tips. Moreover, if the columnar tips are subject
to changes in their local cooling conditions (G and Vis ), λ1 will readapt to the new conditions.
However this readapting process is not at all instantaneous but follows a particular dynamics. This
readapting process is analyzed in detail by Ciobanas et al. in [12]. In fact the primary arm spacing
is viewed as the spacing characterizing a stable dendrite array configuration. In [12] we have
quantified in a simple way the two mechanisms responsible for the selection of the primary arm
spacing: the dendrite division mechanism and the dendrite overgrown mechanism. First, in the
upper dendrite spacing range the mechanism of dendrite division is identified. Indeed, for a large
average spacing between dendrites, the solutal interactions between adjacent dendrites are weak
and the liquid behind the dendrite points will remain close to initial concentration C0 due to the
thin solutal boundary layers around the columnar tips. Secondary or even tertiary arms developed
behind the columnar tips may evolve into vertical primary arms reducing thus the primary arm
spacing. Hence this configuration will not be stable with respect to the dendrite division
mechanism. The division of columnar dendrites has been quantified in [12] by means of the
characteristic time trs representing the response time of the columnar tips to a small perturbation
from the quasi-steady state. Basically this time quantifies the laps of time during which a tip
slightly deviated from the quasi-steady state reach back the isotherm Tts for which Vt = Vis , that is
reach back the quasi-steady state. It was found that

−1
⎛ ∂V ⎞
trs  −2 ⎜ t ⎟ (15)
⎝ ∂x ⎠ xst

where xst is the position of the isotherm Tts corresponding to the quasi-steady state at the tip. Note
as well that trs depends strongly on the thermal gradient G at the CF and the isotherm velocity Vis .
Secondly, in the lower dendrite spacing range the mechanism of dendrite overgrown holds. Indeed
if the spacing between the dendrites array is small enough, due to strong solutal interactions
between adjacent dendrite tips, some of the dendrites are overgrown by the surrounding elements
of the array. This configuration will be thus unstable with respect to the dendrite overgrown
mechanism. The solutal interactions between adjacent tips has been quantified in [12] with the help
of the time scale tdw .
As pointed out in [12] one expects that at the intersection of these two unstable regimes to
find a stable dendrite array configuration defined by the primary arm spacing λ1 . It is shown that
the stable primary arm spacing can be linked to the characteristic time trs as follows:

λ1 = 3 ( Dl trs )
1/ 2
(16)

III-21
Submitted to Journal of Physics D: Applied Physics

where Dl is the solute diffusivity in the liquid. A detailed comparison with various directional
experiments [12] shows that, despite its simplicity, equation (16) provides results which agree
remarkably well with the experimental data.
It is well known that when the dendrite array is subject to changes in local cooling
conditions (G and Vis ) or optimal spacing λ1 (by means of a blocking mechanism, i.e. the first
mechanical blockage effect) the dendrite spacing will always readapt to the new conditions. Notice
again that this readapting process is not instantaneous but follows a particular dynamics. An
important conclusions of the study in [12] is that the time scale characterizing this dynamics is no
more than trs . Indeed trs , besides from quantifying the dendrite division mechanism, will also
quantify the time during which the dendrite array will reach back the quasi-steady state
corresponding to the new conditions at the tip. This time scale is of great interest in determining if
whether or not the columnar tips can be considered as evolving in a quasi-steady state with respect
to the non-stationary local conditions at the tip (G, Vis and λ1 ). Indeed, if the time scale
characterizing the changes in local conditions at the tip is much larger than trs that means that the
readaptation time of the dendrite array to the new local conditions will be much more rapid than the
conditions at the tip are actually changing. Given these conditions, in spite of non-stationary local
conditions at the tip, the columnar array evolution should be very close to the quasi-steady state.
As detailed previously (Figure 11) if no redensifying mechanism of the columnar tips is
considered, even if Vcf and Tcf remain constant at the CF, no quasi-steady state of the CF can be
obtained ( nc and ε c are constantly diminishing for t > tst ). However, as pointed out earlier, the
columnar tips will react to changes in local conditions (i.e. nc ) trying to reach back the quasi-
steady state. Indeed, by diminishing the columnar grain density nc , the effect of the first
mechanical blockage is to rarefy the dendrite array. However, the natural tendency of the columnar
grains would be to reestablish the optimum primary arm spacing by means of the dendrite division
mechanism. Indeed, the decrease of nc deviates the columnar tips from the optimum state
characterized by the primary arm spacing λ1 . Therefore, the columnar division will act to
reestablish the optimal primary arm spacing. This mechanism is not instantaneous but have a
characterized time scale, namely trs . In this respect it will be interesting to compare trs with a time
scale characterizing the first mechanical blockage. Let us denote this latter time scale with tm1 . For
a case where the CF advances with a constant velocity through the equiaxed zone Vcf = Vis ( )
(Figure 11a, t > tst ) the time scale tm1 can be easily estimated. Indeed, as pointed out in equation
(14), the columnar grain density nc decrease with a factor ε ge in a distance of order of
−1/ 3
R1e = ( 4 / 3π ne ) . Hence the time scale characterizing the decrease of nc is simply

tm1 = R1e / Vcf (17)

Now, by comparing tm1 with trs one can determine if whether or not the redensifying
mechanism of the columnar dendrite tips is rapid enough to neglect the effects of the first
mechanical blockage. Indeed, if trs  tm1 the first mechanical blockage is acting faster than the
rebranching mechanism. As a consequence the columnar zone will be rarefied continuously during
the CF advancing through the equiaxed zone. The columnar tips will finally disappear leaving
place to a fully equiaxed structure.
However, if trs  tm1 the rebranching mechanism is much faster than the equiaxed grains
are actually blocking the columnar tips. Hence, one can approximate the columnar nucleation
density as being constant and equal to the value nc0 computed from the optimum primary arm
spacing, equation (16).

III-22
Submitted to Journal of Physics D: Applied Physics

Figure 13: The evolution of the equiaxed grain fraction, the columnar fraction and the columnar
grain density at the CF (Al-3wt%Cu, CR = 0.1 K/s, ∆Tne = 1 K , ne = 109 m −3 , V p = 8 ⋅ 10−5 m / s ,
no first mechanical blockage effect)

Let us for example reanalyze the case shown in Figure 11. If one neglects the first
mechanical blockage effect (admit that trs  tm1 ), the columnar grain density nc and the columnar
fraction ε c at the CF for t > tst will remain constant (Figure 13). Hence, all parameters
characterizing the state at the CF are stationary. A quasi-steady state condition at the CF can be
obtained and a mixed c/e zone will further occupy the ingot.

3.2.2 CET maps. The CET maps represents in the ( G − Vis )-space the different possible quasi-
steady CET, i.e. pure equiaxed, pure columnar or mixed columnar-equiaxed. To built a CET
diagram one should model the CF evolution for different G − Vis pairs and see if whether or not a
pure columnar, equiaxed or mixed CET is obtained. However this direct approach would be very
tedious since many G − Vis pairs must be considered to cover a typical ( G − Vis )-space.
Fortunately, the problem can be greatly simplified by considering the following hypothesis:
i) the equiaxed zone ahead of the CF is decoupled from the zone behind the CF;
ii) for a given CR , ∆Tne and ne the equiaxed zone ahead of the CF can be fully
determined from the local temperature;
iii) as pointed out in the previous section, the pure columnar, pure equiaxed or mixed CET
depends solely on the pulling velocity V p ( = Vis ) .
Therefore, to realize a CET map one has to fix first the equiaxed nucleation undercooling ∆Tne and
the equiaxed grain density ne . If the CR is further fixed, the equiaxed state ahead of the CF can be
fully determined. By varying the CR the pure equiaxed state ahead of the CF is fully determined for
the whole range G − Vis . Finally for a given CR, one has only to vary the pulling velocity V p in
order to identify the CET type. In fact for a given CR , ∆Tne and ne the variation of the equiaxed
tip velocity can be directly linked to the local temperature (see Figure 7). As discussed, two
0 *
characteristic velocity have been determined, w ge and w ge (Figure 7). The first one is linked to
the equiaxed undercooling ∆Tne and fixes the limit between the pure columnar solidification and
the mixed one. The second velocity marks the limit above which the CF will be solutally blocked
by the equiaxed zone ahead, that is the limit between the mixed and pure equiaxed solidification.
0
Basically one has a pure columnar solidification for V p < w ge , a pure equiaxed solidification for

III-23
Submitted to Journal of Physics D: Applied Physics

* 0 *
V p < w ge and a possible mixed solidification for w ge ≤ V p ≤ w ge . We say “possible” mixed
solidification because a quasi-steady state can be obtained at the CF only if the redensifying
mechanism is more effective than the first mechanical blockage effect ( trs  tm1 ). Otherwise, a
pure equiaxed regime must be considered since the columnar zone will finally vanish into the
equiaxed one leaving place to a fully equiaxed structure.
Let us neglect for the moment the first mechanical blockage. In this respect, for a pulling
0 *
velocity between w ge and w ge , a quasi-steady mixed solidification will be obtained. In turn this
state can be characterized with a column fraction ε c which can be expressed as follows:

⎧ ⎡
( ) ⎤⎥
1
2

1 ⎢10ne + 36 ( ne ) nc ε ge
0 3
⎪ 3

⎪1 − ε ge − ⎢
20nc0 ⎢
⎥ if (
R1c ≤ R1e 1 − ε ge1/ 3 )
( )
2

εc ⎪ + 45 ( n )
1
0
3 n ε
3 ⎥
⎪ ⎣ e c ge ⎦
=⎨ (18)
(1 − ε e ) ⎪
0 ⎡( ε ) − 20ε + 45 ( ε ) 3 ⎤
2
2
⎪ nc ⎢ ge
⎪ 20n ⎢
ge ge
⎥
⎥ if (
R1c > R1e 1 − ε ge1/ 3 )
( ge ) 3 + 10
1

⎪⎩ e −36 ε
⎢⎣ ⎥⎦

where ε ge is the equiaxed grain fraction at the CF corresponding to the quasi-steady state (see
Figure 13 t > tst ), nc0 is the columnar nucleation grain density which depends on the local primary

( )
−1/ 3 −1/ 3
arm spacing λ1 and finally R1c = 4 / 3π nc0 , R1e = ( 4 / 3π ne ) . Since we neglect for the
moment the first mechanical blockage effect, λ1 can be approached with the quasi-steady solution
of Equation (16).

Figure 14: CET map for the Al-3wt%Cu alloy ( ∆Tne = 0.75 K and ne = 109 m −3 )

By varying now the CR and resolving for each CR the equiaxed zone, the two
0 *
corresponding critical velocities w ge and w ge , can be determined. Hence, the CET map can be
quite easily built, by marking the two limits between the pure columnar and mixed solidification
and between the mixed and pure equiaxed regimes. Moreover, the mixed zone can be further
characterized by specifying the iso-columnar fraction curves. In this way the mixed zone predicted
semi-empirically by Hunt [9] can now be rigorously characterized.

III-24
Submitted to Journal of Physics D: Applied Physics

Two typical CET maps for the Al-3wt%Cu alloy, a constant ∆Tne of 0.75K and two
different equiaxed grain density ( ne = 109 m −3 and respectively ne = 105 m −3 ) are presented in
Figure 14, Figure 15. One can observe the two important limits corresponding to the two critical
0 *
velocities w ge and w ge (blue lines). The first one (the horizontal line) marks the limit between the
0
pure columnar and the mixed regime. Indeed, if the pulling velocity is below w ge the maximum
undercooled zone ahead of the CF will be always smaller than ∆Tne . Hence, no equiaxed grains
can nucleate in the pure undercooled melt ahead the CF. Finally a fully columnar structure is
identified in the ingot.
The second limit separates the mixed from the pure equiaxed regime. Above this limit the
solutal blockage mechanism as defined by Martorano et al. [7] becomes non-negligible and the
equiaxed grains will finally stop the CF. As pointed out, for a given CR this limit correspond to a
*
pulling velocity equal to w ge , i.e., the maximum equiaxed velocity (Figure 7). Any V p above this
limit corresponds to a sufficiently large undercooled melt ahead the CF in such a way as the
equiaxed grains have the time to enrich the extra-dendritic liquid and to finally block the columnar
*
tips. For V p > w ge , a pure equiaxed structure will be finally obtained in the ingot. Notice again that
this solutal blocking limit is the same as the CET limit identified by Martorano et al. in [7].Notice
however that due to the different approach regarding the estimate of the diffusion length δ l − d
characterizing the solute flux at the grain interface, the Martorano et al. model underestimates the
solutal limit with respect to the present model. This is a consequence of the fact that for average CR
Martorano et al. approach predicts a faster enriching process of the extra-dendritic liquid than the
present model does (see Appendix A). Note as well the important effect of ne on the solutal
blocking limit. The decrease in ne is significantly shifting this limit towards the low-G regime
(Figure 15), increasing therefore the mixed c/e zone. This can be expected since smaller ne means
weaker solutal interactions between equiaxed grains delaying the solutal enriching process of the
*
extra-dendritic liquid. Indeed, for a given CR a smaller ne will determine an increase of w ge
(Figure 5). Hence, the solutal blocking effect becomes non-negligible at smaller thermal gradients.

Figure 15: CET map for the Al-3wt%Cu alloy ( ∆Tne = 0.75 K and ne = 105 m −3 )

As pointed out, between these two limits, a quasi-steady mixed solidification regime is
obtained if the first mechanical blockage is neglected. This is quantified by means of the columnar
fraction at the CF ε c computed with (18). Consequently iso-columnar fractions can be plotted to
better describe the mixed CET zone (Figure 14, Figure 15). An important remark is that by

III-25
Submitted to Journal of Physics D: Applied Physics

decreasing the CR the mixed zone is shrinking as well. This is a normal consequence of the fact
0 *
that to a smaller CR it corresponds a smaller interval w ge - w ge (Figure 8). As observed (Figure 14,
* 0
Figure 15, low-G regime), there is a limit value of the CR below which w ge − w ge = 0 . Indeed,
below a critical CR the equiaxed extra-dendritic liquid remains in a state of perfect solutal mixing
from the beginning of the solidification. No mixed zone exists in that case, and the solutal limit can
0
be identified with the horizontal line corresponding to w ge .
It is of interest to compare now these CET maps with the simplified Hunt’s model [9] for
which a LGK approach is used to determine the dendrite tip kinetics (Hunt-LGK model). By
( )cf , Hunt’s model
imposing empirically two limits on the equiaxed grain fraction at the CF, ε ge

manages to delimit as well three zones on the CET map: a pure columnar zone if ( ε ge ) < 0.1 , a
cf

pure equiaxed zone if ( ε ge ) > 0.49 and a mixed one if 0.1 ≤ ( ε ge ) ≤ 0.49 . The two limits
cf cf

corresponding to ( ε ge ) = 0.1 and respectively ( ε ge ) = 0.49 are superimposed on the two CET
cf cf
maps. One observes that in the high-CR regime these two limits converge almost perfectly with the
iso-columnar fractions of 0.9 and 0.5 respectively. This result has a two-fold explanation. On one
hand for the high-G regime, due to very small solutal interactions between equiaxed grains the
extra-dendritic remains close to C0 until very close to the moment when the equiaxed grain
fraction reaches the unity. Given these conditions, the Hunt’s model which neglects completely the
solutal interactions between grains would predict a similar equiaxed grain evolution as the present
model. On the other hand the equivalence between the two equiaxed grain limits fixed by Hunt (0.1
and 0.49) and the iso-columnar fractions (0.9 and 0.5) can be explained by the fact that for high
CR, and basically for relatively high G and Vis , the primary arm spacing λ1 becomes very small in
comparison with the average distance between equiaxed grains, meaning that nc0  ne . As a
consequence when the columnar tips enters within equiaxed zone, the sharing of space between the
equiaxed and the columnar structures will be highly unbalanced. Indeed for nc0  ne , the columnar
tips will occupy much more effectively the available equiaxed extra-dendritic liquid (1 − ε ge )
compared with the equiaxed grains. This is well reflected by equation (18) in which for nc0  ne
( )
one obtains a columnar fraction ε c → 1 − ε ge . However, in the middle range of CR regime, the
Hunt’s model deviates sensibly from the present predictions, overestimating the ε ge ( )cf = 0.49
limit. The same has been identified by Martorano et al. [7] when comparing the solutal limit and
( )cf
the ε ge = 0.49 limit. This non-agreement is a consequence of the fact that for middle range
values of CR, the solutal interactions between grains becomes non-negligible. Therefore, by
neglecting the solutal enriching process of the equiaxed extra-dendritic liquid, the Hunt’s model
will overestimate the moment when the equiaxed grain fraction reaches the 0.49 value. This
disagreement is somewhat less important for the case involving small equiaxed grain density
( ne = 105 m−3 ). This is due to the fact that for large average distances between equiaxed grains,
even for relatively small CR the solutal interactions between grains is negligible. The enriching of
the extra-dendritic liquid takes place suddenly when the equiaxed grain fraction approaches unity.
Finally, in the low-CR regime the two Hunt’s limits [9] converge to a same horizontal line as the
present model does. As pointed in [7] this should be expected since in the low-G regime the CET
should be independent on ne and depends solely on the isotherm velocity. The good agreement
between the Hunt and the present model in the low and high-CR regime is further establishing
confidence in the present numerical calculations.

III-26
Submitted to Journal of Physics D: Applied Physics

Remember that in characterizing the mixed CET zone we have neglected so far the first
mechanical blockage, that is we have supposed that trs  tm1 . It would be instructive now to
superimpose on the two CET maps the limit trs = tm1 (Figure 14, Figure 15, red line). In this way
we will divide the mixed CET zone in two sub-domains. For the one corresponding to trs < tm1
(between the red line and the horizontal blue line) one can expect to obtain a quasi-steady mixed
zone since the redensifying mechanism of the columnar tips is rapid enough to counter the first
mechanical blockage effect. On the other hand for the sub-domain corresponding to trs > tm1
(between the red line and the solutal blocking limit), the columnar structure will finally vanish into
the equiaxed zone leaving place to a fully equiaxed zone. It is also interesting to notice that for high
cooling rates the limit trs = tm1 almost converge with the solutal blocking limit. Finally three
distinct zones have been identified in the CET maps: the pure columnar CET, the quasi-steady
mixed c/e CET and the pure equiaxed zone.
The use of the new ensemble averaged model enabled us to quantify rigorously both the
solutal and the mechanical blockage mechanism producing the CET phenomena. Basically, since it
intrinsically includes the solutal and the mechanical blocking effects, the new ensemble model
unifies the semi-empirical Hunt approach [9] (pure mechanical blocking mechanism) and the
Martonero approach [7] (pure solutal blocking mechanism).

4 Simulation of unidirectional solidification experiments

In the following section the present model is used to simulate various unidirectional solidification
experiments in an attempt to validate the new ensemble averaged model. The series of directional
experiments are taken from the studies of Mahpatra and Weinberg [4] and Ziv and Weinberg [5].
Notice as well that the same set of experiments has been used by Wang and Beckermann [13] to
validate their CET volume averaged model. Hence a comparison with this model will be assessed
too.

4.1 Experimental conditions

For the first experimental study [4], a Sn-Pb alloy is poured into a 100 mm long cylindrical mold
which is solidified vertically from the bottom by means of a water-cooled cooper chill (Figure 16).
Various experiments (Table 2) are carried out covering a wide range of chill heat transfer
coefficients ( hext ), pouring superheats ( ∆T0 ) and lead concentrations (5, 10, 15 wt%).
Unidirectional solidification conditions are achieved by minimizing the upper and lateral heat
losses.

Figure 16: The solidification configuration used in [4] and [5]

The heat transfer coefficients, hext , are determined by the authors by fitting the numerical results
obtained with an approximate heat transfer model with the experimental temperature
measurements. Note however that the 1D heat transfer model used in [4] is an approximate one and

III-27
Submitted to Journal of Physics D: Applied Physics

therefore the fitted values for hext may have a relative high degree of uncertainty. Unfortunately
the temperature measurements are not provided by the authors and a direct comparison with the
simulation results is not possible. However CET positions are observed and measured by sectioning
the ingots longitudinally and by polishing and etching the obtained sections. Note however that
values for grain size are not provided as well. As one can see in the following, by supposing that
fragmentation of columnar grains is the main cause for the CET phenomena, the grain size
measurements will not be needed anymore.
The second study [5] uses the same solidification configuration as the previous one.
However the solidified alloy is Al-3wt%Cu this time. Again various chill heat-transfer coefficients
are covered and the CET positions is measured for each experiment. Note that the authors have
carried out as well experiments with grain refined alloy (TiB2 additions) obtaining in some cases a
mixed equiaxed-columnar zone behind the CET frontier, confirming therefore the existence of
mixed c/e zone behind the CF. In contrast with the Sn-Pb experiments, detailed grain size
measurements are provided. For the non-refined experiments the equiaxed grain size varied slightly
with the heat transfer, between 5.4-6.2 mm, corresponding thus to an equiaxed grain density equal
to ne ≈ 5 × 106 m −3 . An important missing parameter for this set of experiments is the pouring
superheat ∆T0 . In the absence of a value for ∆T0 we used the same approach as Wang and
Beckermann [13], by considering that a reasonable approximation for ∆T0 would be 20o C .
A summary of experiments results obtained in [4, 5] is presented in Table 2. A detailed
comparison of these data with the model results is assessed in the following.
4.2 Numerical results and discussion
Since the temperature field is not anymore imposed as for the quasi-steady solidifications in the
previous section, the fully transient formulation for the energy equation has to be used. Basically,
this would be equivalent with solving the full transient 1D model detailed in Table 1.

Table 2: Summary of the unidirectional experiments in [4] and [5]

Exp. Superheat hext Measured xcfcr cr
trs tthcr G cr
No. [K] CET
⎡ W ⎤ [m] [s] [s] [K/m]
⎢ m 2 .K ⎥ position
⎣ ⎦ [m]
Sn – 10 wt pct Pb
1 11.5 63 0.038 0.0508 882.1 1549.1 9.16
2 11.5 113 0.060 0.0591 626.3 1003.9 10.91
3 11.5 134 0.065 0.0606 466.9 868.2 14.14
4 19 63 0.040 0.0511 804.7 1558.2 10.09
5 19 84 0.045 0.0535 656.8 1222.6 11.81
6 19 134 0.065 0.0599 344.8 858.5 18.95
7 31 63 0.035 0.0545 841.4 1661.3 9.30
8 31 96 0.050 0.0590 597.9 1179.1 11.84
9 31 113 0.057 0.0587 526.4 997.4 13.20
10 31 134 0.065 0.0617 476.8 883.5 13.72
11 36 96 0.050 0.0585 623.5 1170.7 11.45
12 36 134 0.065 0.0614 380.6 879.8 17.46
13 36 209 0.080 0.0677 280.2 621.7 21.04
Sn – 10 wt pct Pb
14 7.5 71 0.052 0.0460 686.7 1244.8 8.33
15 12 92 0.060 0.0520 599.7 1086.2 8.86
16 22 113 0.070 0.0556 452.4 944.4 10.86
Sn – 15 wt pct Pb
17 21.5 151 0.048 0.0630 443.4 800.6 17.89
18 21.5 167 0.055 0.0647 418.5 744.3 18.55

III-28
Submitted to Journal of Physics D: Applied Physics

19 21.5 272 0.070 0.0709 282.0 500.6 23.49

20 36 105 0.040 0.0613 550.1 1120.8 15.46
21 41 146 0.050 0.0638 405.5 839.3 19.69
Al – 3 wt pct Cu
22 20 95 0.072 0.0469 883.1 1678.0 6.66
23 20 65 0.053 0.0421 1103.6 2201.3 6.15
24 20 63 0.048 0.0397 1143.6 2139.6 5.93
25 20 53 0.045 0.0375 1273.2 2406.7 5.69
26 20 50 0.037 0.0368 1219.2 2500.3 6.01
27 20 37 0.034 0.0342 1563.9 3144.3 4.85

The 1D simulations are carried out on a N = 200 cells grid. As pointed out the three cells
near the columnar front are always refined into N ref smaller cells. Several test were realized and a
typical value of N ref = 41 was found to be highly sufficient for the typical solidification conditions
in Table 2. The CET, cooling curves, and the other average parameters converge in a satisfactory
way for N > 100 and N ref > 21 . A time step of ∆t = 0.05 s was found to capture well enough the
kinetics of the solidification experiments in Table 2. In addition the use of the adaptive time
scheme was found to reduce considerably the influence of ∆t on the final results. The upper
thermal boundary condition is considered adiabatic and at the lower boundary a heat flux is
imposed:

φbot = hext (Tbot − Text ) ⎡W / m 2 ⎤

⎣ ⎦
(19)

where Text = 293 K is the ambient temperature, Tbot the bottom wall temperature and hext the heat
transfer coefficient. The typical CPU time for 1000s of real solidification was of order of 45
minutes on a AMD Athlon 2000+ standard PC.
It is worthwhile to notice that in the following the columnar nucleation grain density nc0 is
directly linked to the local cooling conditions G and Vis using Equation (16). Therefore we
considered that the first mechanical blockage effect on the columnar tips is negligible, meaning that
trs  tm1 .

4.2.1 Base test case. A first test case is discussed in more detailed in the present paragraph. This is
the 22nd case in Table 2 that is an Al-3wt%Cu alloy, poured in a superheated liquid ( ∆T0 = 20o C )
and cooled from the bottom ( hext = 95 W .m −2 .K −1 ). We firstly consider in the following a
vanishing equiaxed nucleation undercooling, ∆Tne = 0o C and an equiaxed nucleation grain density
of ne = 105 m −3 . The main numerical results are presented in Figure 17-23. In Figure 17 the cooling
curves at five different position along the ingot are presented. As expected, the recalescence
phenomena are more pronounced close to the bottom wall due mainly to a high heat extraction rate.
Far from the bottom wall, the cooling curves exhibit thermal plateaus. As Wang and Beckermann
[13] noticed, by taking into account the mixed columnar and equiaxed solidification the
recalescence phenomena reduces in intensity compared with a pure equiaxed solidification.

III-29
Submitted to Journal of Physics D: Applied Physics

Figure 17: Cooling curves for the Al-3wt%Cu test case experiment ( hext = 95 W .m −2 .K −1 ,
∆T0 = 20o C , ∆Tne = 0o C , ne = 105 m−3 ) corresponding (from the left to the right) to a distance of
x = 1, 10, 20, 30, 40, 50 mm from the bottom of the ingot

a) b)
Figure 18: a) the CF position and the CF velocity and b) the thermal gradient and the equiaxed
grain fraction at the CF evolution for the Al-3wt%Cu test case experiment ( hext = 95 W .m−2 .K −1 ,
∆T0 = 20o C , ∆Tne = 0o C , ne = 105 m −3 )

It is of interest to focus now on the columnar-equiaxed interactions. In Figure 18a the time
evolution of the CF position with respect to time is presented along with the corresponding CF
velocity. As it can be observed at tCET = 840 s the CF is stopped, its velocity drops suddenly to 0
and the CET occurs at xCET = 0.068 m . Since no first mechanical blockage effect is considered, the
CET is due in this case to the solutal blocking effect of the equiaxed grains ahead the CF which
reject enough solute in their extra-dendritc liquid to finally block the CF. Note as well that the CET
occurs very rapidly, similar to what Martorano et al. [7] observed. It would be interesting to look in
more detail the evolution of the equiaxed grains at the CF In Figure 18b, the equiaxed grain and the
thermal gradient at the CF are plotted against time. As observed the equiaxed grain fraction
(ε ge )cf is continuously increasing. Indeed since ∆Tne = 0o C , the equiaxed grains nucleate in front
of the CF from the very beginning of the solidification. As the CF velocity increases in the
beginning of the process, the undercooling at the CF ∆Tcf increases. Hence, the extend of the
undercooled zone ahead of the CF increases as well and this increase is further amplified by the

III-30
Submitted to Journal of Physics D: Applied Physics

decrease of the thermal gradient at the CF. The decrease of Gcf is a normal consequence of the
particular cooling conditions of the ingot. Indeed, the lower wall temperature decreases in time and
the flux extracted at the bottom diminishes as well. The decrease of Gcf continues until the CET
occurs and increases afterwards since, even if the CF is stopped, the cooling of the ingot continues.
Basically, the increased undercooled region ahead of the CF favors the equiaxed growth. The
equiaxed grains reject more and more solute in front of the CF and will finally solutally block the
CF. As discussed already, besides the solutal blocking effect, the equiaxed grains ahead the CF are
also mechanically blocking the columnar grains.

a) b)

Figure 19: a) the columnar fraction and b) the columnar nucleation grain density ( ∝ λ1−3 )
variations along the ingot for the Al-3wt%Cu test case experiment ( hext = 95 W .m −2 .K −1 ,
∆T0 = 20o C , ∆Tne = 0o C , ne = 105 m −3 )

As seen in Figure 19a the columnar fraction is continuously decreasing along the ingot in
favor of the equiaxed structure. This is due to the second mechanical blockage effect as it was
( )cf
detailed in the previous section. Indeed, due to the finite value of ε ge and the differences

between nc0 and ne (Figure 19b), the two structures are constrained to share in a different way a
same volume. This is quantified in the model by means of the c/e fractions ε c / e . As seen in Figure
19a, at the very beginning of the solidification the columnar structure occupy all the space available
( ε c  1 ) despite the presence of equiaxed grains ahead of the CF. This is due to a high nc0
compared with ne (Figure 19b) and to the initial small equiaxed grain fraction at the CF (Figure
18b). However, as the CF advances the equiaxed zone ahead becomes larger, ε ge ( )cf increases and

consequently nc0 decreases further. This, of course, is reflected by the rapid drop of the columnar
fraction ε c far from the lower boundary. Finally the ingot will be filled with a relative shallow
columnar structure at the bottom of the ingot, followed by a mixed c/e zone which makes the
transition to a fully equiaxed structure.

III-31
Submitted to Journal of Physics D: Applied Physics

Figure 20: The c/e fraction, the c/e solid fraction and the c/e grain fractions along the ingot at
t = 625 s (Al-3wt%Cu test base case)

a) b)
Figure 21: a) The total solid fraction ( ε se + ε sc ) evolution along the ingot and b) the c/e extra-
dendritic liquid concentration and the total solid fraction near the CF at t = 625 s (Al-3wt%Cu test
base case)

a) b)
Figure 22: a), the grain density and the grain diameter near the CF and b) the solid and extra-
dendritic liquid concentration near the CF at t = 625 s ( t = 625 s (Al-3wt%Cu test base case,
Martorano et al. [7] approach)

III-32
Submitted to Journal of Physics D: Applied Physics

The mixed c/e zone which form behind the CF will undergo as well a complex evolution.
This is well reflected by Figure 20 where the c/e solid and grain variations along the ingot are
plotted at t = 625 s , before the CET moment ( tCET ≈ 840 s ). Note the correlation between the
decrease of ε c (increase of ε e ) and the decrease (increase) of the corresponding c/e solid and grain
fractions. In addition the constraint ε c / e ≥ ε gc / e ≥ ε sc / e is verified as well due to the particular
mathematical formulation of the model in Table 1 and especially of the length diffusion length
δ lk − dk . Indeed, since the solute diffusion process around the c/e equiaxed grain is limited to a
distance R1*c / e around the grain center, the grain growth will be always stopped before it reaches
R1*c / e . Since R1*c / e are no more than the dimensions of the c/e cells characterizing the statistics of
the mixture zone, the constraints enounced above are intrinsically verified. In Figure 21b the total
solid fraction ( ε se + ε sc ) and the c/e extra-dendritic liquid concentration variation near the CF are
presented at t = 625 s . As seen the solutal blocking effect of the equiaxed zone is still negligible,
( Cle )cf  C0 . However, just behind the CF, the enriching of the liquid becomes important. This is

due to the second mechanical blockage which reduces sensibly the characteristic lengths R1*c / e ( nc0
is larger than ne ). Note the smooth transition zone (about 1mm ) from the non-enriched state at the
CF to the perfect mixing state behind. Note as well the small differences between Clc and Cle ,
result which is coherent with the physical fact that both dendrite grains share a same space
therefore a same extra-dendritic liquid. The equivalence between Clc and Cle is establishing
confidence in the model formulation. This is an important aspect of the ensemble averaged model.
In fact, even if we have delimited a columnar and an equiaxed extra-dendritic liquid, due to the
equivalence between the length scales characterizing each extra-dendritic liquid,

R1*c  R1*e − ae (see [1]) (20)

the evolution of Clc and Cle prior to the coexistence moment will be similar. Note as well the
equivalence between the total solid fraction evolution and the Scheil law for points found in state of
perfect solutal mixing. Indeed, despite the separate handling of the columnar and equiaxed
structures, the c/e solid fraction evolution behind the CF is such that when the state of perfect
mixing is reached, ε se + ε sc evolution identifies with the Scheil regime (Figure 21a). This
establishes further confidence in the mathematical formulation of the model since any solid
configuration found in a state of perfect mixing should evolve following a Scheil law. By taking
into account the coexistence between the two structures one has a more precise approach on the
solidification phenomena taken place behind the CF. In this respect it would be interesting to
compare the above results with the approach used by Martorano et al. [7]. In that study, the authors
approximate the mixed zone behind the CF with an equivalent columnar zone of density nc0 .The
present model was modified to mimic the Martorano et al. approach by considering the zone
behind the CF as pure columnar and enforcing the continuity of average values at the CF. For the
present base test case, the resulting CET position was found to be x = 0.064 m that is almost 6%
lower than the one obtained with the new model. Even if the difference is not significant it is
interesting to look in more detail what could cause this underestimation of the CET position. Since
one expects that the two approaches will give similar results for the pure equiaxed zone ahead the
CF and the perfect solutally mixed zone behind the CF (Scheil law), one should focus on the
transition zone to the perfect solutally mixed state behind the CF. In Figure 22b the solid and the
extra-dendritic concentration are presented close to the CF at t = 625 s In contrast with the smooth
transition to the perfect solutally mixed state in Figure 21b, Figure 22b shows an extremely rapid
increase of Cl to Cl* . The thickness of the transition zone in this case is not greater than 250 µ m ,
that is, almost four times smaller than the one obtained with the present model. This is mainly due

III-33
Submitted to Journal of Physics D: Applied Physics

to the discontinuity of the grain density ne at the CF (Figure 22a). Indeed, since the continuity of
average variables at the CF ( ε ge , ε se , etc) is enforced one obtains discontinuities in the evaluation
−1/ 3
of real dimensional parameters like grain diameter a = ( 4 / 3π n ) ε g1/ 3 or interfacial areas
S g = 3ε g 2 / 3 ( 4 / 3π n )
1/ 3
) (Figure 22a). Since usually nc0  ne the interfacial area will undergo a
sudden increase. This in turn will lead to a very rapid enriching process of the extra-dendritic liquid
behind the CF (Figure 22b) and a significant modification of the solidification rate at the CF
leading finally to an underestimation of the CET position. The relative small difference between
the two approaches is due mainly to the fact that we have neglected the macroscopic solute
diffusion terms in the solute balance equations transforming them into local ODE equations. Far
from the CF the axial gradients of Cl are indeed small. However, behind the CF these ones
become important and could modify the solute fluxes at the tip, underestimating further the CET.
On the other hand for the case analyzed above one had nc0  ne . The inclusion of relative dispersed
grains into the highly dense columnar structure would not change significantly the solidification
conditions behind the CF with respect to a pure columnar zone of density nc0 . In contrast, for
nc0  ne the mixed zone behind would evolve similarly with an equivalent pure equiaxed zone of
density ne and not with a pure columnar zone of density nc0 as the Martorano et al. [7] approach
would have predicted. Even if, given the present solidification conditions, the differences between
the CET position predicted by the two approaches is not significant, the use of the present model is
of great interest since it gives a rigorous description of the mixed c/e zone behind the CF.
Note that so far we have computed the primary arm spacing λ1 (and nc0 consequently)
using equation (16). The latter equation is valid for a quasi-steady state CF, that is when Vis = Vcf .
It should be interesting to investigate if whether or not the CF evolves in a quasi-steady state with
respect to the local cooling conditions ( G , Vis ). By knowing the precise position of the CF one can
compute locally the isotherm velocity since Vis = CR / G . In Figure 23 the columnar tip velocity
and the local isotherm velocity at the tip are plotted against time. Basically one can delimit the CF
evolution in three typical zones. The first one ( t < t1 ), corresponding to the very beginning of the
solidification corresponds to a transition stage in the evolution of the CF. Indeed the columnar tip
velocity Vcf increases quite rapidly from 0 to reach finally the local isotherm velocity Vis . This
stage is similar with the transition zone identified in the previous section for the case involving a
fixed G and Vis (Figure 11a). The second phase corresponds to a CF velocity almost equal with
the local Vis ( t1 < t < tCET ), that is to a quasi-steady state at the CF In this case the use of the
equation (16) to compute λ1 would be entirely justified. As observed, the isotherm velocity
undergoes small oscillations around a mean value. Several tests were carried out to better
understand this phenomenon. It was observed that the mesh influenced slightly the nature of these
oscillations, and it appears that their origin is mainly a consequence of the particular law used for
both columnar and equiaxed instantaneous nucleation. The discrete modeling of the nucleation
phenomena generates small thermal perturbations at the CF which in turn gives the noisy aspect of
the isotherm velocity. This second phase in the evolution of the CF ends suddenly close to the
critical moment corresponding to the CET. The tip velocity drops suddenly at tCET and the CF
stops. During the third phase ( t > tCET ) the CF velocity is almost zero contrary to the local
isotherm velocity at the CET interface.

III-34
Submitted to Journal of Physics D: Applied Physics

Figure 23: The isotherm velocity at the CF and the CF velocity with respect to time

In this context one important question is why the tip velocity, during the second phase in
Figure 23 evolves in a quasi-steady state with respect to local conditions, despite the continuous
changing cooling conditions ( Vis , G) at the columnar tips (see for example Figure 18). To give a
pertinent response to this question it may be interesting to analyze the characteristic times scales of
the CF dynamics and of the changes in local cooling conditions at the columnar tips. As pointed
out previously, the time scale characterizing the CF dynamics when perturbed from a quasi-steady
state is no more than trs (cf. Equation (15)). Basically, trs measures the time during which a
columnar tip reach back the quasi-steady state after a small perturbation from an initial quasi-
steady state or a sudden change in local cooling conditions at the tip ( Vis , G). Thus it is interesting
now to compare this time scale with the time scale characterizing the changes in local cooling
conditions at the columnar tips. Let us denote this latter parameter by tth . This time scale could be
approached from a relative simple scale analysis of the energy equation particularized to the given
ingot configuration. Let us denote by xcf the distance of the columnar front from the bottom wall
at time t and with δ s the total steel sheet thickness between the cooper chill and the ingot bottom
wall. Indeed, the authors vary the heat transfer coefficient hext by varying the thickness of the steel
sheet between the water cooled copper block (maintained at 293 K ) and the mold bottom wall. The
time scale tth characterizing the thermal changes at the CF will be no more than the travel time of a
thermal diffusion wave from the cooper chill (maintained at Text ) to the CF position. One should
therefore analyze the following thermal equivalent configuration composed from two inline
thermal resistances (Figure 24).

III-35
Submitted to Journal of Physics D: Applied Physics

Figure 24: The thermal model used to compute the characteristic thermal time tth

Supposing that the characteristic time related to the latent heat release is larger than the
characteristic molecular diffusion time, one could neglect the influence of the latent heat release on
tth . This later approximation should be physically reasonable since we are dealing with high
conductive materials (liquid metals) and relatively low solidification rates. Hence, from the
simplified energy equation (drop of the latent heat release term) one can write approximately

∆T 1 ⎛ ∆T ⎞
ρcp ≈ ⎜ ⎟ (21)
tth xcf + δ s ⎜⎝ 1/ hext + xcf / λt ⎟
⎠

Noting that 1/ hext >> xcf / λt for the experiments reported in Table 2 ( xcf ≈ 0.05 m ,
λt ≈ 30 − 100W .m−1.K −1 and hext = 50 − 200W .m−2 .K −1 ) and that δ s ( < 0.005m [4]) is typically
lower than xcf , one finally obtains the following estimate

ρ c p xcf
tth ≈ (22)
hext

A comparison between tth and trs can now be assessed. In Figure 25 the evolution of tth and trs at
the CF with respect to time is plotted. One can distinguish three important stages in the evolution
of the CF, similar with the previous remarks (Figure 23). The first stage ( t < t1 ) represent a
transition to a quasi-steady state phase for CF. Indeed since trs > tth the changes in local conditions
at the tip are much more rapid than the columnar tips are able to readapt to these new local
conditions. The CF find it difficult to approach the local Vis . In contrast, the second phase
( t1 < t < tCET ) corresponds to trs  tth . Hence, the columnar tip response time to changes in local
cooling conditions are much faster than the thermal conditions at the tip are actually changing.
Basically, the CF is readapting almost instantly to the new thermal conditions and a quasi-
stationary state at the CF is maintained ( Vcf  Vis ). Finally, due to the drop in the local thermal
gradient and the fast enriching of the liquid at the CF by the equiaxed grains ahead, trs increases
rapidly close to the CET and becomes larger than trs . The CF is not readapting anymore to the
constantly changing conditions at the tip and the equiaxed zone ahead is extending further favoring
the equiaxed grains development. The CET is imminent. Thus it seems that when trs becomes of

III-36
Submitted to Journal of Physics D: Applied Physics

the same order of magnitude with tth the conditions triggering the CET are reached. As we will see
next, the relationship between these two time scales is very important since it enables one to
quantitatively approach the critical moment corresponding to the CET. The strong correlation
between tth and trs seems indeed to be a relevant indicator regarding the CET phenomena.

Figure 25: The thermal ( tth ) and tip response time ( trs ) evolution at the CF with respect to time

4.2.2 Comparison with the experiments. In the present paragraph a quantitative comparison with
the experiments in [4, 5] is attempted. The full transient model summarized in Table 1 is used to
simulate the Al-Cu and Sn-Pb unidirectional experiments (Table 2). One expects that the CET
position varies strongly with the initial superheat ∆T0 , the heat transfer coefficient hext , the initial
alloy composition C0 and the equiaxed grain density ne . A first analysis of the influence of all
these parameters on the CET position is very useful for the comprehension of the CET phenomena.

Figure 26: The influence of the heat transfer coefficient and the equiaxed grain density on the CET
position for an Al-3wt%Cu alloy ( ∆Tne = 0 , ∆T0 = 20 K , C0 = 3% ). Comparisons with the
experimental data [5] ( ne  5 × 106 m −3 ) and with the Martorano et al. approach [7].

III-37
Submitted to Journal of Physics D: Applied Physics

a) b)
Figure 27: (a) The influence of the initial solute concentration and the equiaxed grain density (Sn-
Pb alloy, ∆Tne = 0 K , ∆T0 = 10 K , hext = 120 W / m2 ) and (b) the influence of the pouring
superheat and the equiaxed grain density ( Sn-10wt%Pb alloy, ∆Tne = 0 K , hext = 120 W / m2 ) on
the CET position. Comparison with Wang and Beckermann model [13].

In Figure 26 the variation of the CET position with respect to the heat transfer coefficient
hext and the equiaxed grain density ne is presented for an Al-3wt%Cu alloy. As one expected, the
increase of ne will determine a decrease of the CET position. Indeed, since for high equiaxed grain
density the solutal enriching process of the extra-dendritic liquid is more effective, the solutal
blocking effect is stronger. In contrast, the increase of hext favors the columnar zone and delays the
CET. This can be easily explained by considering the fact that high hext means higher heat flux at
the bottom of the ingot which in turn will lead to higher thermal gradients at CF. This narrows the
equiaxed zone ahead of the CF, favoring the columnar zone. In Figure 27a the influence of the
pouring superheat on the CET is analyzed for a Sn-10wt%Pb alloy and for three different ne
values. It can be observed that a higher ∆T0 favors the columnar zone. This was expected since
higher ∆T0 means as well higher thermal gradients ahead of the CF at the beginning of the
solidification, delaying further the CET. Again the strong influence of the value of ne on the CET
must be noted. In Figure 27b the sensitivity of the CET to the variation of the initial solute
concentration is analyzed for the Sn-Pb alloy. Basically, increasing C0 leads to a lower CET
position. A reasonable explanation for this behavior can be found in the influence of C0 on the tip
kinetics. Indeed, let us consider given cooling condition ( Viso , G). If the CF arrives to a quasi-
steady state ( Viso  Vcf ) then the undercooling at the CF will be larger for an increase value of C0 .
For a given thermal gradient G at the tip this will lead to an extension of the equiaxed zone ahead
of the CF favoring thus the equiaxed structure.
A comparison with the Martorano et al. model [7] can be quantitatively assessed if the
present model is modified to mimic the model in [7]. By approximating the mixed zone behind the
CF with a pure columnar one the Martorano et al. model underestimates the CET predictions
(Figure 26). Note however that the differences between the present model and the Martorano et al.
model are not important for the present solidification conditions. As pointed out this is due on one
hand to the absence of the macroscopic diffusion terms in the solute equations and on the other
hand to the larger columnar density nc0 compared to ne .
The cases analyzed in Figure 27 permit a direct comparison with the Wang and
Beckermann model [13]. Significant differences can be observed between the present model and
the results provided in [13] (see Figure 27). Considerable higher CET positions are predicted by
Wang model. This can be mainly explained by the fact that no solutal coupling between the CF and

III-38
Submitted to Journal of Physics D: Applied Physics

the equiaxed zone ahead is taken into account in [13].The CF velocity is computed by considering
that the local extra-dendritic solute concentration is always at C0 , neglecting therefore the solutal
blocking effect of the equiaxed grains. Accordingly an over estimation of the CET can be logically
expected.

Figure 28: The influence of the equiaxed nucleation undercooling on the CET position for an Al-
3wt%Cu alloy: ne = 5 × 106 m −3 , ∆T0 = 20 K .

In the absence of a reliable value for ne in the case of Sn-Pb experiments, a comparison
with the Al-Cu experiments is assessed first, knowing that for these experiments, detailed grain size
measurements were carried out. Basically, the equiaxed grain density varies slightly from one
experiment to another around the value ne  5 × 106 m −3 [5]. Note however that an important
parameter of the model is not known a priori, i.e., the equiaxed nucleation undercooling ∆Tne . In
Figure 26 a comparison with the experiments is assessed in the limit of a vanishing equiaxed
undercooling, ∆Tne = 0 K . As one can see no agreement is obtained with the experimental data.
Using, ∆Tne = 0 K , Martorano et al. [7] arrived at the same conclusion, that is, a vanishing
equiaxed undercooling cannot explain the CET experimental data in [14]. Thus similarly with the
Martorano et al. approach [7], a parametric study is attempted to determine the value of ∆Tne
fitting the experimental results. In Figure 28 the influence of ∆Tne on the CET is analyzed for the
experiments 22 and 25 in Table 2. As expected the increase of ∆Tne favors the columnar zone. The
influence on the CET position is not important for small values of ∆Tne and increases rapidly when
∆Tne approaches a critical value ∆Tnecrit . In fact for ∆Tne > ∆Tnecrit a fully columnar zone is always
obtained ( xCET = 0.1 m , see Figure 28). Basically, the same qualitative results are obtained for the
four other Al-Cu experiments in Table 2. It is important to notice that similar variations of the CET
position with respect to ∆Tne were obtained by Martorano et al. in [7]. This particular result can be
explained by considering that, for a given solidification condition and in the absence of the
equiaxed grains ahead of the CF, the local undercooling at the columnar tips, ∆Tcf remains always
lower than a maximum value ∆Tnecrit which depends mainly on the cooling conditions of the ingot
and C0 . Basically for a more efficient ingot cooling, the maximum CF velocity increases as well
and consequently ∆Tnecrit increases. It is worthwhile to notice that the real CET position
corresponds to a ∆Tne very close to the ∆Tnecrit , slightly lower than ∆Tnecrit (Figure 28). Essentially,
the same result was obtained by Martorano et al. [7] when they validated their model with the
experiments of Gandin [14]. This interesting remark reveals however some very important issues.

III-39
Submitted to Journal of Physics D: Applied Physics

In fact ∆Tne  ∆Tnecrit means that the equiaxed grains nucleate very close to the CF and moreover
that their nucleation corresponds more or less to the moment when ∆Tcf reaches a maximum. As
Martorano et al. [7] pointed out, the most reasonable explanation for this phenomenon is that the
equiaxed grains are in fact fragments detached form the columnar grains and transported in front of
the CF by the inherent natural convection acting close to the columnar macro-front. This result was
also supported by the findings of Gandin [15] which observed that if no equiaxed zone ahead of the
CF is assumed, the real CET position corresponds more or less to a local maximum value of the CF
velocity, that is to a maximum value of the columnar tips undercooling, ∆Tcf . This means as well
that prior to this moment the CF velocity will exhibit a decrease. Since Jackson et al. [16] showed
that a change in the CF velocity may cause dendrite arm remelting and fragmentation it is
reasonable to think that the columnar dendrite fragmentation is at the base of the CET phenomena
and not the equiaxed nucleation ahead the CF. Note that this should be mainly valid for non-refined
alloys. For highly refined alloys, the probability of equiaxed nucleation ahead of the CF should not
be neglected. However important questions remain to be answered: Why the CF velocity (in the
absence of the equiaxed grains ahead) reaches a local maximum? Why the sudden decrease in the
CF velocity is susceptible to produce remelting of columnar dendrite? Is there a way to predict the
moment corresponding to ∆Tnecrit ?

4.2.3 Fragmentation of dendrite columns. A consequence of the CF instability? In the following we

will try to give a pertinent response to the questions enounced above. Note first that ∆Tne  ∆Tnecrit
meaning that the equiaxed grains nucleate close to the columar front when the columnar
undercooling reach a local maximum, would means as well that the equiaxed zone ahead would not
influence the CF evolution prior to the critical moment corresponding to the CET since they don’t
nucleate before the CET moment. In this light by focusing on the CF evolution alone and not
taking in account the equiaxed zone ahead, one should be able to identify the critical CET moment.
The model in Table 1 is used therefore for the modeling of the pure columnar zone, by completely
ignoring the equiaxed structure. All solidification experiments in Table 2 are analyzed in the
following.

Figure 29: Time evolution of (a) the CF velocity and the temperature at the CF (b) the CF position
and the thermal gradient at the CF for a Sn-10wt%Pb alloy pure columnar solidification (case no.
12, Table2: ∆T0 = 36 K , hext = 134 W / m 2 )

In Figure 29a the CF velocity and the corresponding temperature at the CF are plotted
against time for the 12th Sn-Pb experiment in Table 2. The evolution of the CF position and the
thermal gradient at the CF are plotted in Figure 29b. As expected, since no stopping mechanism is
accounted for this simulation, the CF advances undisturbed in the undercooled melt and the
columnar structure will finally occupy the whole ingot. Somewhat similar to the previous base test

III-40
Submitted to Journal of Physics D: Applied Physics

case, the CF velocity exhibits a sudden increase at the beginning of the solidification ( t < tcr1 )
followed by a quasi-plateau corresponding to a quasi-state at the columnar tips and finally, above a
critical time ( tcr 2 ), a somewhat oscillatory evolution around the mean value characterizing the
plateau. Due to the particular cooling conditions at the bottom of the ingot, the gradient at the CF
decreases as the CF distances away from the lower wall. Finally, for t > tcr 2 G exhibits an
oscillatory movement close to the 0 value. To have a more detailed insight on what could
determine such a particular evolution at the CF, especially for t > tcr 2 , it is instructive to analyze
the time scales characterizing the CF evolution and the changes in local conditions at the columnar
tips. In a similar way as in the previous section, the time scales tth and trs are plotted together in
Figure 30a. As one can observe, the first phase in the CF evolution corresponds to trs > tth , that is a
transition phase of the CF from a non-quasi-steady state to a quasi-steady state. Indeed, even if the
CF is constantly trying to reach a quasi-steady state with respect to the local conditions ( G , Vis ),
due to the fact that the CF readapts to the constantly changing conditions at the tip mush more
slowly ( trs ) than the local conditions are actually changing ( tth ), the columnar tip velocity does not
manage to reach a quasi-steady state. However, for t > tcr1 , trs becomes smaller than tth and the
columnar tips readapt much more rapidly to the local cooling conditions at the tip. Therefore, Vcf
will follow closely the local isotherm velocity Vis (Figure 30b) and a quasi-steady state at the tip is
reached. As long as the gradient at the CF remains important, the condition trs < tth is
accomplished. However, due to the constant decrease of G, trs will finally approach tth for
t > tcr 2 . Finally for trs > tth the CF will not rapidly readapt itself to the local conditions at the tip. A
remarkable fact is that the moment tcr 2 , corresponding to a slightly smaller trs than tth ,
corresponds as well with the beginning of the oscillatory evolution of Vcf and G. The same
conclusion has been obtained for all cases analyzed in Table 2. In Figure 31 two other cases are
presented to exemplify the good correlation between the beginning of the oscillatory variation for
Vcf and the equivalence between the two time scales trs than tth . In Table 2 the CF position xcfcr ,
tthcr , trs
cr
and G corresponding to the critical moment t = tcr 2 are summarized for all experiments.
One can observe the good correlation between the beginning of the oscillatory behavior of the CF
and the two time scales tthcr , trs
cr cr
. Indeed, if trs is plotted against tthcr (Figure 32) one can easily
observe that tthcr ≈ 2trs
cr
. To explain this interesting phenomenon first one should remember that
since we have not taken into account a blocking mechanism for the CF the columnar tips will
evolve undisturbed with respect to the constantly changing conditions at the tip ( G , Vis ). Indeed,
as pointed out in [12] the CF evolution is such that at any moment it tries to reach the isotherm Tts
corresponding to a local tip velocity equal to the isotherm velocity Vis , that is to reach the quasi-
steady state with respect to local cooling conditions. However for t > tcr 2 , one has trs > tth ,
meaning that the CF readapts more slowly to the new conditions at the tip than the later ones are
actually changing. In fact in their search of the quasi-steady state, by the time ( trs ) the CF velocity
reaches the local isotherm velocity, the cooling condition at the tip ( G , Vis ) have already changed.
The main consequence of this CF behavior is the particular oscillatory movement of Vcf and G for
t > tcr 2 . Notice as well that due to the constant decrease of G towards zero, the amplitude of the
isotherm velocity at the CF greatly increases (Figure 30b) since Vis = CR / G . Consequently, the
oscillatory behavior of the CF velocity is further amplified as well. It is important to notice
however that the beginning of the oscillatory phase ( t = tcr ) always corresponds to a positive
thermal gradient at the CF (see Table 2), of order of 10-20 K/m. Notice that a critical gradient of
the same order of magnitude was obtained by Martorano et al. [7].

III-41
Submitted to Journal of Physics D: Applied Physics

a) b)
Figure 30: Time evolution of (a) the thermal ( tth ) and tip response time ( trs ) at the CF and (b) the
isotherm velocity at the CF and the CF velocity for a Sn-10wt%Pb alloy pure columnar
solidification (exp13 in Table2: ∆T0 = 36 K , hext = 134 W / m2 )

a) b)
Figure 31: Time evolution of the CF velocity, the thermal ( tth ) and the tip response time ( trs ) for a
pure columnar solidification: (a) case no. 17, Table 2 and (b) case no. 22, Table 2

Figure 32: The thermal ( tth ) and the tip response time ( trs ) at the CF corresponding to the
beginning of the oscillatory evolution of the CF

III-42
Submitted to Journal of Physics D: Applied Physics

An even more remarkable fact is that the position of the CF evolution at t = tcr 2 , xcfcr ,
corresponds reasonably well with the CET position measured in the experiments. Despite the
relative high degree of incertitude on the real cooling conditions for almost all the experiments, the
relative error between xcfcr and xCET is lower than 20% (Figure 33). Therefore, it is reasonable to
think that the existence of a non-quasi-steady state at the tip with respect to the local cooling
conditions ( trs ≥ tth ) is responsible for the CET phenomena. In contrast with Ziv and Weinberg [5],
which correlated the gradient ahead of the CF with the CET phenomena, in the present study it
seems more likely that the CET position is correlated with the two time scales tth and trs .

Figure 33: Comparison of the predicted CET position ( xcfcr ) with all solidification experiments in
Table 2

An important question remains however: Why the CET phenomena seems to correlate so
well with the equivalence between the two time scales trs , tth ( tthcr ≈ 2trs cr
)? To answer to this
question, one must remember that trs quantified as well the optimal primary arm spacing of a
quasi-steady columnar front λ1 , equation (16). Since the value of time scale trs strongly oscillates
for t > tcr 2 (Figure 30a, Figure 31) λ1 will inherit the oscillatory evolution from trs . As pointed out
in [12] the dendrite spacing is always trying to readapt to the new conditions at the tip, that is to the
optimum λ1 , by means of two mechanisms. If the dendrite spacing is larger than λ1 , the dendrite
division mechanism will determine a decrease in the primary arm spacing. In contrast, if the
dendrite spacing is lower than the optimum λ1 the solutal interactions between two adjacent
columnar tips are strong enough in such a way that the dendrite configuration will not be stable to
small perturbation at the tip. Hence, some of the columnar tips will be overgrown by the neighbor
ones and the dendrite spacing is naturally increased. From this point of view, the oscillatory
evolution of the optimal primary arm spacing λ1 will force the columnar structure to adapt its
dendrite spacing very fast. As one can see the typical oscillations frequency of trs (and λ1
consequently) is rather small (~10 s) and much faster than the time during which the CF is actually
able to readapt ( trs  102 − 103 s ). Consequently primary arm spacing will never reach the optimal
spacing λ1 and will therefore be non-adapted to the local cooling conditions at the tip: either the
solutal interactions between adjacent grains will be too strong, either too weak, leaving between
columnar tips a large undercooled liquid melt. Thus the oscillations of λ1 for t > tcr 2 reflects in

III-43
Submitted to Journal of Physics D: Applied Physics

fact an instable CF state with respect to the local cooling conditions at the tip. The rapid change
from a strong solutal interaction state ( λ < λ1 ) to a weak solutal interaction state ( λ > λ1 ) will be
extremely favorable to dendrite remelting and fragmentation. This, coupled with the inherent
natural convection at the macrofront would determine the existence of very favorable CET
conditions at the CF.
The good agreement between the numerical results and the experimental data support the
above findings. It seems that the columnar fragmentation favored by an unstable regime of the CF
with respect to the local cooling conditions represent the main cause for the CET phenomena, at
least for the non-refined alloys. Even more, since the CET seems to be linked to this unsteady
regime of the CF, the equivalence between the two time scales tth and trs would further permit to
quantitatively approach the critical CET moment.

5 Conclusions
In order to validate the new two-phase Eulerian model developed throughout reference [1], a
numerical implementation of the model was developed and applied here to simulate the 1D
columnar-to-equiaxed transition. Indeed, owing to the statistical nature of the model, one is able to
treat rigorously the coexistence of equiaxed and columnar structures and consequently the CET
phenomena. The implemented numerical model is able to rigorously model 1D directional
experiments in which both columnar and equiaxed structures may interact. The columnar front
evolution is modeled by means of a front tracking technique. Similarly with the Martorano et al.
model [7] the local columnar tip velocity is directly linked to the local undercooling conditions.
Hence the solutal blocking effect of the equiaxed grains ahead of the CF is naturally taken into
account. Moreover, owing to the statistical nature of the model, the mechanical blockage effect of
the developed equiaxed zone ahead of the CF on the columnar zone is intrinsically taken into
account by means of the two blockage mechanisms already identified in [1].
First we focus on the numerical simulation of quasi-steady solidification experiments in
order to obtain corresponding CET maps. Observing the physical decoupling between the evolution
of the pure equiaxed zone ahead of the CF and the columnar zone evolution behind the CF we were
able to easily identify three main zones characterizing the CET phenomena. First, two important
0 *
characteristic equiaxed tip velocities have been identified: w ge and w ge . Then, depending on the
choice of the pulling velocity with respect to these two characteristic velocities three main zones
have been identified on the CET map: the pure columnar, the pure equiaxed zone and finally the
mixed columnar+equiaxed zone. The mixed c/e zone was further analyzed to determine if whether
or not a quasi-steady state at the tip could be identified. Indeed, due to the first mechanical
blockage a continuous rarefying process of the columnar zone can be observed. However, a
redensifying mechanism opposing to the first mechanical blockage was identified as well. The
redensifying mechanism was quantified by means of the time scale trs [12]. Finally, by comparing
the time scales characterizing the redensifying mechanism and the first mechanical blockage we
were able to determine if whether or not a mixed quasi-steady state at the CF could be obtained.
Two CET maps characterizing an Al-3wt%Cu alloy were built and the three main CET zones were
identified. Even more the mixed c/e zone was further quantified by means of a columnar fraction
ε c . The latter one quantifies in a rigorous the way the two coexisting structures share locally a
given space. Basically, ε c will be a measure of the amount of space proper to the columnar
structure. Therefore, the present ensemble averaged model was able to reproduce the mixed
columnar/equiaxed zones within the CET maps. Moreover, the new model enabled us to quantify
rigorously both the solutal and the mechanical blockage mechanism producing the CET
phenomena. Basically, since it intrinsically includes the solutal and the mechanical blocking
effects, the new ensemble model unifies the semi-empirical Hunt approach (pure mechanical
blocking mechanism) and the Martorano et al. approach (pure solutal blocking mechanism).
Secondly the present model was used to simulate various unidirectional solidification
experiments in an attempt to validate further the new ensemble averaged model. The full transient
model in Table 1 was used to simulate different Al-Cu and Sn-Pb unidirectional experiments [4, 5].
The complex coupling between the CF and the equiaxed zone ahead as well as the complex

III-44
Submitted to Journal of Physics D: Applied Physics

evolution of the mixed c/e zone which forms behind the CF were analyzed. We have also
compared the present model with the Martorano et al. approach to the mixed zone behind the CF.
A significant underestimation of the CET position was obtained when the Martorano et al.
approach [7] was used. This was mainly due to the fact that in [7] the mixed zone behind the CF is
approximated with an equivalent columnar zone, the interactions between the equiaxed and
columnar grains not being taken into account. The steadiness of the CF with respect to local
cooling conditions was analyzed as well. When the local CF velocity was compared with the local
isotherm velocity it was observed that the two were almost equal until very close to the CET
moment. Hence the columnar front evolved in a quasi-steady state until very close to the critical
CET moment. This was possible because the CF adapts to the constantly changing conditions at the
tip much more rapidly than the latter ones are actually changing. Indeed, when comparing the two
time scales characterizing the dynamics of the columnar dendrite tips ( trs ) on one hand and the
variation in local cooling conditions at the tip ( tth ), on the other hand it was observed that during
the quasi-steady evolution of the CF one had always: trs  tth . Moreover it was observed that the
CET moment corresponds more or less to the equivalence between these two time scales: trs ≈ tth .
Next a comparison with the Wang and Beckermann model [13] was assessed. Since the latter
model neglected the solutal blocking effect of the equiaxed grains the CET position was constantly
overestimated with respect to the present model. A comparison of the present model with the Al-Cu
alloy experiments is assessed too, in the limit of a vanishing equiaxed undercooling ( ∆Tne = 0 K ).
It is found that the equiaxed nucleation undercooling is close to the maximum columnar dendrite
tip undercooling and that the CET is thus virtually independent of the equiaxed zone ahead of the
CF. This is in agreement with the Martorano et al. findings [7] and supports the idea upon which
the CET is mainly caused by the fragmentation of the columnar dendrites, rather than
heterogeneous nucleation ahead of the CF.
Next, the evolution of the pure columnar structure is analyzed in order to understand the
mechanism pointing to the CET. If the equiaxed zone is further ignored it is observed that the
evolution of the columnar front follows a three stage evolution. First the CF velocity exhibits a
sudden increase at the beginning of the solidification followed by a quasi-plateau corresponding to
a quasi-state at the columnar tips and finally, above a critical time ( tcr 2 ), a somewhat oscillatory
evolution around the mean value characterizing the plateau. It is found that the quasi-steady
evolution of the CF corresponds to the condition at the tip trs  tth hence to a CF which adapts
much more rapidly to the non-steady conditions at the tip than the latter ones are actually changing.
Moreover the critical moment ( tcr 2 ) corresponding to the beginning of the oscillatory evolution of
the CF was well correlated with the equivalence between the two time scales trs , tth ( tth  2trs ).An
even more remarkable fact is that the position of the CF corresponding to tcr 2 corresponds
reasonably well with the CET position measured in the experiments. Hence, the oscillatory
evolution of the CF corresponds to a non-quasi-steady state of the columnar tips. It is found as well
that the oscillatory evolution of the CF determines an oscillatory evolution of the optimal primary
arm spacing which in turn creates very favorable conditions for the fragmentation of the columnar
dendrites. We conclude that the unstable regime of the CF with respect to the local cooling
conditions represent the main cause for the CET phenomena, at least for the non-refined alloys.
Even more, since the CET could be linked directly to this unsteady regime of the CF, the
equivalence between the two time scales tth and trs would further permit a quantitative approach
of the critical CET moment.

Appendix A: The validation of the diffusion length scale δ l − d

To validate the new diffusion length the present model is first compared with Martorano et al. [7]
model and more precisely with the numerical results obtained for a uniformly solidified equiaxed
morphology. The studied case involves a local cooling rate of 0.005 K/s and four different grain
densities corresponding each to the following four cell dimensions: Rcell (=R1) = 10mm; 3mm;
1mm, 0.1mm [7]. The alloy considered is an Al 3wt.% Cu metallic alloy. The corresponding

III-45
Submitted to Journal of Physics D: Applied Physics

physical properties can be found in [7]. In Figure A1(a) the evolution of the grain fraction is
presented for both the present and the Martorano et al. [7] model. A very good agreement between
the two approaches can be observed for the cases involving the smallest grain densities (Rcell =
10mm). On the contrary for smaller grain densities (Rcell = 3mm; 1mm) differences between the two
models can be noticed. First one can observe that the enriching of the liquid and the stop in the
grain growth takes place more rapidly with the decrease of Rcell. Moreover a decrease of the
maximum grain fraction can be observed too with the increase of the grain density. This can be
explained by the fact that to a smaller Rcell it corresponds smaller Peg numbers where
Peg = w g ( R1 − ae ) / Dl is the grain Peclet number, already defined in [1]. As pointed out in [1] the
solute boundary layer around the grain is of order of δ = ( R1 − a ) / Peg . Hence for smaller Peg
numbers (for Peg numbers close to unity) the solute boundary layer around the grain becomes of
the same order of magnitude as the length scale of the extra-dendritic liquid, R1 − a . This together
with the fact that the solute flux at the cell boundary is zero will determine a non-negligible
enriching of the extra-dendritic liquid (see Figure A1(b)). Basically, due to a smaller distance
between the grains (Rcell) the solutal interactions between the grains become more important and
the enriching of the extra-dendritic liquid takes place sooner (given the same cooling rate).

Figure A1: Comparison between Martorano et al. [7] model (dashed line) and present model (full
line): a) the grain fraction evolution for four different grain densities and b) the average extra-
dendritic concentration for Rcell=1mm case.

Now, comparing the present and the Martorano et al. [7] approach one can observe that for
the intermediate cell dimensions the Martorano et al. model predicts a smaller maximum grain
fractions (Figure A1(a)) and a more rapid solutal enriching of the extra-dendritic liquid (Figure
A1(b)). This coupled effect can be explained by the differences in the formulation of the diffusion
length scale δ l − d . Indeed, in order to maintain the zero solute flux at the cell boundary, for Peg < 2
the present model limits the solute boundary layer around the grain at the length scale of the extra-
dendritic liquid and more precisely to ( R1 − a ) [1]. In contrast with the present model the
Martorano et al. model [7] does not impose a zero solute flux condition at the cell boundary.
Hence, for cases involving Peg smaller than unity the Martorano et al. approach would predict a
larger solute flux at the grain boundary compared with the present model and a non-negligible
solute flux at the cell boundary. Therefore, the Martorano et al. model [7] will predict a more
effective enriching process of the extra-dendritic liquid compared with the present approach.

III-46
Submitted to Journal of Physics D: Applied Physics

Figure A2: Comparison between Wang and Beckermann [2] model (dashed line) and present model
(full line): a) the temperature evolution and b) the grain fraction evolution for three different grain
densities
Secondly the present model is compared with the numerical results obtained by Wang and
Beckermann [2] for a uniformly solidified system with an equiaxed morphology. Indeed, the
system analyzed is identical with the one presented first by Rappaz and Thévoz in [6] in the case of
a Al-5wt.%Si. The temperature is assumed to be locally uniform and the local heat extraction
corresponds to a local cooling rate of dT / dt = 45 K / s . Finally, results are obtained for three
different grain densities corresponding to the following cell dimensions Rcell (=R1) = 0.01m;
0.001m; 0.0001m [6]. A detailed comparison of the temperature and the grain fraction evolutions is
presented in Figure A2. One can observe that the agreement between the present model and the
Wang and Beckermann [2] approach is excellent for the whole range of grain densities. Small
differences can be observed for the case involving the smallest grain density (Rcell = 0.0001m).
Indeed, for this case, due to a smaller Rcell , the corresponding involved Pe g numbers will be
lower too. As already pointed out this will determine a faster enriching process of the extra-
dendritic liquid and a lower maximum grain fraction compared with the other two cases (see Figure
A2(b)). Here too a faster enriching process of the extra-dendritic liquid can be observed for the
Wang and Beckermann [2] approach compared with the present model. Indeed, due to the
particular formulation of δ l − d in [2] and especially due to the non-zero flux at the cell boundary,
the flux at the grain boundary is higher than for the present model: consequently the extra-dendritic
liquid is more rapidly enriched in solute and the predicted maximum grain fraction will be lower
with respect to the present model.

III-47
Submitted to Journal of Physics D: Applied Physics

Nomenclature

Symbol Meaning Unit

C solute concentration field [wt. %]

T temperature field [K]
N grain density [ m −3 ]
ρ density field [ kg .m −3 ]
εk the “volume” fraction of phase k ---
Ck* the interfacial concentration at the k phase interface [wt. %]

nc0/ e the columnar/equiaxed nucleation grain density [ m −3 ]

λ1 columnar primary arm spacing [m]
Sk the interfacial area density of phase k [ m −1 ]
δk− p the solute diffusion length at the k-p interface [m]
xcf position of the columnar front [m]
xst the position of the quasi-steady columnar tips [m]
Vcf , Vt columnar front (tip) velocity [m/s]
Vp pulling velocity [m/s]
Vis isotherm velocity [m/s]
w ge averaged equiaxed tip velocity (average velocity of the [m/s]
equiaxed grain envelope)
* maximum equiaxed tip velocity [m/s]
w ge
* equiaxed tip velocity corresponding to the equiaxed [m/s]
w ge
nucleation undercooling ∆Tne
G thermal gradient [K/m]
CF columnar front ---
CR cooling rate [K/s]
Tliq liquidus temperature at the initial concentration C0 [K]
Tcf temperature at the columnar front [K]
Tne the equiaxed nucleation temperature [K]
Tf the fusion temperature of the pure solvent [K]
Tts the isotherm corresponding to the quasi-steady state of the [K]
columnar front
∆Tne equiaxed nucleation undercooling [K]
∆Tcf columnar front undercooling [K]
∆T0 initial superheat [K]
trs the return time to the steady state of a perturbed quasi-steady [s]
dendrite array
tdw the time scale characterizing the solutal interactions by [s]
diffusion between two adjacent columnar tips
tth the time scale characterizing the non-steadiness of the cooling [s]
conditions at the tip
tm1 the time scale characterizing the first mechanical blockage of [s]
the columnar tips

III-48
Submitted to Journal of Physics D: Applied Physics

m the slope of the liquidus line [ K .( wt % ) ]

−1

k the partition coefficient ---

ck heat capacity for the k-phase [ J .kg −1.K −1 ]
λk the phase k thermal conductivity [ W .m −1.K −1 ]
L the latent heat [ J .kg −1 ]
Ω the dimensionless local undercooling ---
Pet the dendrite tip Peclet number ---
Dk the solute diffusivity within the phase k [ m 2 .s −1 ]
ac / e the local columnar/equiaxed grain radius [m]
R1c / e the cell radius corresponding to the single columnar/equiaxed [m]
grain distributions
R1*c / e the cell radius corresponding to the coexisting [m]
columnar/equiaxed grain distribution
z i the position in the ingot of the ith closest grain to x [m]
if (1) t , x; z 1
( ) the unconditional density function that the closest grain to the [ m −3 ]
c/e
point x is centered within d z 1 of z 1 and is a
columnar/equiaxed grain
δ lk − dk diffusion length characterizing the solute transfer at the [m]
columnar (equiaxed) grain interface
Γk the k phase mass transfer rate [ kg .m −3 .s −1 ]
Over lines
ensemble average
 relative to the nearest neighbor distributions
Subscripts
cf columnar front
c columnar phase
e equiaxed phase
c/e columnar or equiaxed phase
s solid phase
l extra-dendritic liquid phase
d inter-dendritic liquid phase
f total fluid phase ( = l + d )
g grain phase ( = d + s )
kc columnar k phase
ke equiaxed k phase

References

1. Ciobanas A I and Fautrelle Y 2006 "Ensemble Averaged Two-Phase Eulerian Model for
Columnar/equiaxed Solidification of a Binary Alloy. Part I. The Mathematical Model" Journal
of Physics D: Applied Physics (Submitted)

2. Wang C Y and Beckermann C 1993 Metallurgical and Materials Transactions A 24A 2787

3. Wang C Y and Beckermann C 1996 Metallurgical and Materials Transactions A 27A 2754

4. Mahapatra R B and Weinberg F 1987 Metallurgical Transactions B 18B 425

III-49
Submitted to Journal of Physics D: Applied Physics

5. Ziv I and Weinberg F 1989 Metallurgical Transactions B 20B 731

6. Rappaz M and Thévoz Ph 1987 Acta Metallurgica 35.7 1487

7. Martorano M A, Beckermann C and Gandin Ch-A 2003 Metallurgical and Materials

Transactions A 34A 1657

8. Patankar S V 1980 Numerical Heat Transfer and Fluid Flow (New York: Hemisphere
Publishing Corporation)

9. Hunt J D 1984 Materials Science and Engineering 65 75

10. Hunt J D 1979 Solidification and casting of metals (London: Metals Society) 3

11. Kurz W and Fisher D J 1981 Acta Metallurgica 29 11

12. Ciobanas A I, Bejan A and Fautrelle Y 2006 "Dendritic solidification morphology viewed from
the perspective of constructal theory" Journal of Physics D: Applied Physics (Submitted)

13. Wang C Y and Beckermann C 1994 Metallurgical and Materials Transactions A 25A 1081

14. Gandin Ch-A 2000 Iron Steel Inst. Jpn. 40 971

15. Gandin Ch-A 2000 Acta Mater. 48 2483

16. Jackson K A, Hunt J, Uhlmann D and Seward T 1966 Transactions of the metallurgical society
of AIME 236 149

III-50
Référence IV
Modeling the mixed columnar/equiaxed solidification of a
binary alloy with convection. Mathematical model.
A.I. Ciobanas, Y. Fautrelle

EPM/CNRS Laboratory, ENSHMG, BP 95, 38402 Saint Martin

d’Hères cedex, France

Abstract
The ensemble averaged model derived in reference [1] has been extended here to account
for the fluid convection. Owing to the statistical nature of the model, one is able to treat
rigorously the coexistence of equiaxed and columnar structures and consequently the CET
phenomena. We have approached the ensemble average through the probability
distributions functions attached to the macroscopic process to be studied. The new model
inherits form the model in [1] which quantified in a rigorous manner the solutal and
mechanical (geometrical) interactions between the two coexisting grain distributions. The
microscale physical assumptions are reevaluated with respect to the ones in [1]. First, the
fluid flow around the columnar/equiaxed (c/e) permeable grains and the grain growth in
the presence of fluid convection is analyzed. Finally, an analysis of the solute transfer
around the c/e grain in the presence of convection is made. It is concluded that the mass
transfer at the grain boundary in the presence of convection is intensified with respect to
the pure diffusion case. We propose correlation for the computation of the overall mass
transfer coefficient in the limit of small and high c/e grain fractions. Using the microscale
physical assumptions the average balance equations are further derived. First the grain
balance equations are derived. The equiaxed nucleation is modeled as an instantaneous
event whereas the columnar nucleation (the sudden arrival of columnar tips in one point) is
modeled by means of a front tracking algorithm. We propose a simple front tracking
technique having as basis the volume of fluid model. Moreover the coexistence model
developed in [1] is here extended to account for the grain convection. The cell dimensions
R1c* and R1e* as well as the space partition between the columnar and equiaxed distribution
( ε c and ε e ) is here determined form a new general model describing the coexistence state
between two different grain populations. The mass, momentum and energy balance
equations are derived next. Notice that for simplicity reasons a total fluid approach is
considered for the momentum balance. For the energy equation a single mixture balance
equation is derived. Finally the solute balance equations are obtained. Here, a total fluid
approach is considered too along with the two solute balance equations in the c/e solid. In
addition, the two solute balance equations in the c/e inter-dendritic liquid are used to
compute the c/e solidification rates. We have also proposed an algorithm for the implicit
computation of the extra-dendritic liquid concentrations. This is done by coupling at each
time step the solute balance equations within the c/e extra-dendritic liquids with the solute
balance equation in the total fluid (f phase). Finally a complete ensemble averaged model
is proposed for the modeling of the mixed columnar + equiaxed structures in the presence
of convection.

IV-1
1 Introduction
In reference [1] an ensemble averaged model which can take into account the coexistence
between two grain structures (i.e. equiaxed and columnar) was developed. The use of an
ensemble averaging technique in deriving a micro-macro scale model has some important
advantages compared with the volume averaging technique which has been widely used
during the last decade. Firstly, the statistical approach does not need the definition of a
representative elementary volume. Secondly, it is consistent with the random aspect of the
equiaxed grain germination and motion as well as the possible turbulent behavior of the
fluid flow. Finally, and the most important, owing to the statistical nature of the model, one
is able to treat rigorously the coexistence of equiaxed and columnar structures and
consequently the CET phenomena. Knowing that the CET phenomena are a direct
consequence of complex mechanical and solutal interactions between an advancing
columnar front and an equiaxed zone ahead of the columnar dendrite tips, a model that
could integrate mathematically these interactions would be of great interest. However the
final mathematical model proposed in reference [1] is valid for a pure diffusion
solidification case only. Hereafter we propose an extension of the model in [1] which takes
into account the effect of the fluid convection in the ingot.

2 The ensemble average

As pointed out in [1], in order to obtain an ensemble averaged model describing the
solidification process of a binary alloy one has to write first the local conservation
equations describing the balance of mass, momentum, energy and solute. These local
balance equations are valid in each point of the two-phase system and describe the exact
evolution of the two-phase system to the smallest time and length scale. Unfortunately due
to the extremely complex solid-liquid interface the numerical resolution of local balance
equations will be a virtual impossible task for real solidification problems usually
involving ingots larger than 10-1 m. On the other hand we are not always interested on
variations of parameters at the smallest scale but instead on average parameters like
velocity, solute concentration characterizing large vortexes or mesosegregation, that is
valid at larger scale (meso or macro scale). Moreover, the local balance equations state for
the conservation of ψ inside the whole multiphase system. As one would desire to track
the average evolution for each phase separately, the effect of each phase on the total
balance of ψ has to be identified first. This is done by multiplying the local conservation
equations with the characteristic phase function:

⎧⎪1 if ( t , x ) is within the k phase

X k (t, x ) = ⎨ (1)
⎪⎩0 if ( t , x ) is not within the k phase

Then, in order to have an insight on average quantities, one has to further apply the
ensemble average operator defined in [1]. Using then the topological relations
characterizing X k (Drew [2]), one can easily obtain the average conservation equations for
ψ in each phase k as well as the corresponding jump equations [1]. However as to be
useful these averaged equations has to be structured by defining adequate average
variables. These average variables can be of different types following the choice of the
weighted function: average variables weighted with the characteristic function X k , mass
weighted averages (Favré) weighted with X k ρ and finally, interfacial averages weighted
with ∇X k , having a meaning at the interface only. Definition of all these variables is

IV-2
synthesized in [1]. One is able now to write down the ensemble averaged conservation
equation for the general physical field ψ and the corresponding jump conditions:

(
∂ ε k ρ kψ k ) + div ε
∂t
( k ) ( )⎦
ρ k v kψ k = div ⎡⎢ε k jψ k + j′ψ k ⎤⎥ + ε k ρ k bψ k
⎣ (2)
+ Φψ k + Γψ k ; k = {s, l}

∑ ( Φψ
k
k )
+ Γψ k = σ ; k = {s, l} (3)

The final system of averaged equations describing the solidification process will
consist of mass, momentum, energy and solute balance written for each phase in the
multiphase system. This is done by particularizing the general field ψ , the corresponding
molecular flux jψ and the volumetric source bψ with the appropriated expressions for each
balance equation [1].
The terms encircled in the equation (2) are very important for the correct modeling
of the multi-phase system since they express transfer terms (of mass, momentum, etc.)
between the phases. We have:
ƒ Φψ k , the term responsible of the convecto-diffusive transfer at the phase interface;
ƒ Γψ k , the term responsible of the transfer at the phase interface due to the mass
transfer between phases;
ƒ j′ψ k , the dispersive flux due to the ψ fluctuations (also known as the Reynolds
stress).
Notice that all these terms are unknown tensors which do not depend on known averaged
values. Hence the equations (2) and (3) remains unclosed and one need to find for these
terms closure expressions with respect to known average values. This is no more than the
model closure procedure and it is here that the modeling effort is concentrated.

2.1 The cell model approximation to the ensemble average

The concept of ensemble averaging as defined in [1] is rather abstract. Consequently the
unknown terms arising from the averaging procedure will be difficult to model. In order to
better understand the nature of this average technique one should approach the ensemble
average through the distributions functions attached to the macroscopic process to be
studied.
Given two coexisting grain populations (i.e. columnar and an equiaxed)
characterized by different grain densities, nc / e , their influence at a certain point x would
be different as well. The main advantage of the ensemble averaging is that, with the help of
( )
the nearest neighbor distributions if c / e t , x; z 1 [1], one can quantify in a rigorous way the
c/e influence at x. Indeed, if c / e ( t, x; z )
1 measures the density of probability that one c/e
grain centered at z 1 is the closest grain to x among all the grains existing in the ingot.
These two distributions play a vital role in determining the influence of c/e grains at x
since one expects that what happens at x would be mainly influenced by the closest grain
to x. Using the cell approximation to if c / e , two spherical cells centered at x and of radius
R1c* and R1e* can be identified [1], each of them characterizing in a simple way the statistics
of the c/e grain distributions (see Figure 1).

IV-3
Figure 1: The cell model approximation to the columnar (a) and the equiaxed (b) ensemble
average

Indeed, in the light of the cell model the ensemble θ c / e of realizations such as the
closest particle to x be a columnar/equiaxed one is equivalent with the ensemble of
realizations such as the closest particle to x is c/e and is centered within the cell of radius
R1*c / e centered at x. Using the two spherical cells it is possible to separate the influence at x
of the c/e grains by defining corresponding c/e ensemble averages which quantify the
average effects of the two structures at x. Moreover, any ensemble average (i.e. e/c
fractions, solid/liquid fractions, average solute fields, interfacial averages, etc.) can be
easily computed by means of relative simple integrations over all possible positions of the
closest c/e grain to x, within the respective c/e cell.
Thus, following the case where the closest grain to x is columnar or equiaxed, the
model leads us to distinguish two main types of ensemble averages, columnar and
equiaxed ensemble averages respectively (i.e. c/e fractions, c/e average concentrations,
etc). In this respect, two important fractions are identified, the columnar and equiaxed
fraction, ε c and ε e respectively. These two averages quantify the way the two structures
share the space at x. Since the columnar and equiaxed grains share a same space at x their
corresponding fractions will obviously sum to one, ε c + ε e = 1 .
As an example, using the cell approximation, the ensemble average of a generic
field ψ , ψ c ,e becomes [1]:

1
ψ c ,e ( x, t ) =
ε c ,e
( ) ( )
if (1) t , x; z 1 ψi (1) t , x; z 1 d z 1
∫ c ,e
(4)
nc ,e
=
ε c ,e ∫
(1)
( )
ψi t , x; z 1 d z 1
x − z1 ≤ R1*c ,e

(1)
( )
where ψi t , x; z 1 is the field ψ ( t , x ) valid at the micro-scale, given that the closest grain
to x is centered at z 1 . This, in turn, can be computed analytically given the micro-scale
assumptions and considering that the field ψ ( t , x ) is uniquely influenced by the closest
grain to x, therefore neglecting the influence of more distant grains.

IV-4
 (
Moreover, the use of the probability distributions functions if c / e t , x; z 1 ) in
approaching the ensemble average reveals as well another important aspect of the
coexistence between two grain structures. Indeed, the coexistence between columnar and
equiaxed grains implies a sort of geometrical (mechanical) interaction between the two
( )
coexisting structures. Using the probability distribution function if c / e t , x; z 1 one can
rigorously characterize and quantify these interactions. As detailed in [1], the columnar and
the equiaxed grains interact “mechanically” following a two fold mechanism: the first and
the second mechanical blockage effect.

2.1.1 First mechanical blockage

The first mechanical blockage effect of the coexistence will determine a rarefaction of the
columnar zone when entering (“nucleating”) into a developed equiaxed zone. This will
decrease the initial columnar grain density nc0 with a factor of ( 1 − ε ge ). Indeed due to the
finite equiaxed grain fraction the columnar tips have only a limited space (1 − ε ge ) within
which they can “nucleate”. Hence only a fraction of the total number of columns will be
able to enter the equiaxed zone.

Figure 2: The first mechanical blockage

2.1.2 Second mechanical blockage effect

As pointed out in [1], the second mechanical blockage effect of the coexistence reflects the
fact that for a mixed zone (i.e. columnar + equiaxed), the two coexisting structures are
forced to share a same space (Figure 3). Consequently, the space available for each
structure will be different and smaller too from the corresponding space available just
before the coexistence state (the columnar nucleation in this case).

IV-5
Figure 3: The second mechanical blockage: the full line circles represent the two cells
characterizing the mixed c/e zone and the dashed line circles are the cells characterizing
the two structures (c/e) separately, that is just before the coexistence state

As discussed already, the space characterizing the c/e structure is quantified by the
ensemble model by means of two corresponding spherical cells of radius R1*c / e (Figure 3).
Basically the second mechanical blockage states that the length scales R1*c / e are smaller
than the corresponding length scales characterizing the c/e state separately just before their
coexistence, namely R1c / e . Notice again that R1*c / e together with their corresponding grain
densities nc / e are of primary importance for the modeling of the mixed c/e zone. Indeed
these parameters characterize completely the statistics of the coexistence state, enabling
one to define average variables for each one of the two structures. Moreover, since R1*c / e
can be very different form the original length scales R1c / e the evolution of the mixed zone
can be significantly different from that of an equivalent pure columnar or equiaxed zone.

3 Model closure
In view of the complexity of the processes involved in a solidification problem one should
first propose physical models at the micro-scale for all these processes and only then try to
compute closure expressions for the unknown terms in equation (2). Writing down the
physical assumptions valid at the micro-scale, would permit to quantify as exactly as
possible the velocity, temperature, solute fields around each c/e grain. Then with the use of
the cell model approximation to the ensemble average, equation (4), the unknown terms in
(2) can be computed rigorously with respect to average variables. The macroscopic system
can be in this way closed.

3.1 Micro-scale physical assumptions

Almost all physical assumptions valid at the micro-scale made in [1] for the pure diffusion
case remain exactly the same for the case implying convection. The hypothesis linked to
the: phase diagrams constraints, the solute diffusion controlled solidification, the local
thermal equilibrium, the equiaxed and columnar nucleation remain unchanged. Obviously,
the first hypothesis made in [1] concerning the pure diffusion case is no more valid since in
the following we would like to account for the fluid flow effects. In fact, it is precisely the
fluid convection phenomena that will determine a reevaluation of the final hypothesis
made in [1] on the grain growth.

IV-6
3.1.1 The fluid flow around a c/e dendritic grain
To quantify correctly the fluid flow around an e/c dendritc grain one should quantify the
fluid flow across a swarm of permeable particles. Many mathematical models have been
used [3-5] to study the fluid flow relative to a swarm of impermeable particles spheres.
However the model that seems to give correct results within the entire range of porosity is
the well known “free cell model” of Happel [5]. This model was later extended to
accommodate the case of permeable spheres by Neal [6] and Davis [7] and their results
agree well with experimental data obtained in real solidification problems [8]. In this
respect in order to quantify the fluid flow across a mixed c/e zone we will try to adapt the
Neal model to the case of a mixed e/c zone.
Basically the free cell model introduced by Happel [5] tries to model the
interactions between adjacent particles by resolving a limited fluid flow problem around a
single particle of the swarm. Indeed, the velocity field around the permeable particle of
radius a is accounted to a maximum distance R1 from the center of the particle such as

( a / R1 )
3
=εp (5)

where ε p is the volume fraction of the particles. Note the resemblance between the cell of
radius R1 around the grain and the columnar or equiaxed cells of radius R1*c / e (Figure 1)
which quantify rigorously the statistics of the columnar/equiaxed coexistence at x. As we
have pointed out in [1], the two cells enveloping each c/e grain enclose the space uniquely
perturbed by the respective c/e grain. Outside this cell one enters into the ray of influence
of a neighbor particle. Remember as well that R1*c / e quantifies as well the average distance
between the c/e particles. In this respect in order to describe the fluid flow around a
particle it is physically sound to focus mainly on the flow inside the characteristic cell;
outside this cell the flow would be characterizing a neighbor cell. This ensemble average
interpretation can only enforce the choice made by Happel [5] concerning the dimension of
the cell, equation (5).

Figure 4: The c/e cell and the corresponding fluid flow

Therefore one should quantify the fluid flow around a dendrite grain to a distance
R1*c / e from the center of the c/e particle (Figure 4). However an important issue must be
solved: what are the proper boundary conditions at the cell boundary representative for the

IV-7
interactions between the adjacent grains? Following Neal [6] it is assumed that the particle
is fixed and the cell boundary is traversed by a uniform fluid field of velocity U c / e
corresponding to the flow rate per unit area across the c/e structure. For the case of
stationary solid, U c / e is no more than the superficial c/e velocity characterizing c/e fluid
(f=d+l phase):

U c / e = v fc / e .ε fc / e (6)

If the solid has a non-zero velocity, then U c / e is no more than the relative flow rate per unit
rate across the c/e structure:

U c / e = v fc / e .ε fc / e − v sc / e (7)

The boundary conditions at the cell boundary are [6, 7]:

• The normal velocity component at the cell boundary matches that of the imposed
flow:

( n ⋅ U = n ⋅ u )r = R *
1
(8)
where u denotes the velocity field within the characteristic cell of radius R1* .
• The tangential stress at the cell boundary is zero:

( n ⋅ T ⋅ t )r = R *
1
=0 (9)

where t is the unit tangent vector to the cell surface. Note that these boundary conditions
are not randomly chosen but they are representative of the interactions between neighbor
grains. In fact they reflect the hypothesis that the fluid flow to a distance R1* from the
center of the particle is uniquely influenced by that particle. In this respect, each sphere is
isolated in the sense that it only interacts with its neighbors through the average fields.
Indeed, the zero tangential stress at the cell boundary means that locally there is no
momentum exchange between neighbor particles. Instead the particles interact with their
neighbors through the average velocity field U c / e , equation (8).
Assuming low Reynolds numbers, the flow around a particle can be fairly well
described with a Stokes approximation. Knowing the internal permeability of the dendrite
grain K cd/ e , the flow within the permeable particle can be described by the Brinkman
equation [6, 7]. Finally by imposing the usual cinematic boundary conditions at the grain
boundary [6, 7] one can analytically resolve the fluid flow within the cell of radius R1*c / e
(Figure 4). Consequently one can compute the drag force on the solid particle produced by
the fluid flow around it. One obtains a net drag force of [6, 7]:

Fdk = 6πµ f ak Ω k U k , k = {c, e} (10)

where µ f is the dynamic viscosity of the fluid, ak is the c/e grain diameter and Ω k is a
coefficient depending on the c/e grain diameter ak , its internal permeability K kd and the
ratio η = ak / R1*k [6]:

IV-8
 tanh β
2 β 2 + 4 β 2η 5 / 3 + 20η 5 −
β
( 2β 2
+ 8β 2η 5 + 20η 5 )
Ωk = (11)
⎛ 2 β − 3β η + 3β η − 2 β η + 90 β −2η 6 + 42η 5 − 30η 6 + 3 −
2 2 2 5 2 6
⎞
⎜ ⎟
⎜ − tanh β ( −3β 2η + 15β 2η 5 − 12 β 2η 6 + 90 β −2η 5 + 72η 5 − 30η 6 + 3) ⎟
⎜ β ⎟
⎝ ⎠

where β = ak / K kd . This coefficient ( Ωk ) accounts for both the influence of the internal
permeability as well as for the interaction between neighbor particles. In fact, for K kd → 0
equation (10) becomes identical with the free cell model proposed by Happel [5] for solid
particles. Note as well that for K kd → 0 and η k = 0 , Ω k = 1 . Hence the well known Stokes
drag force on a single particle is obtained. It is also important to notice that for the case of
a coexistence state between equiaxed and columnar structures, the ratio η k is no more
than:

ηk = ( ε gk / ε k )
1/ 3
, k = {c, e} (12)

Similarly with the Wang approach [8], one can use a Karman-Cozeny model [3] to
estimate K kd :

d
(1 − ε sk / ε gk )3
K = (13)
5 ( As / Vg )
k 2

where As is the c/e solid/liquid interface area [ m 2 ] and Vg is the grain volume (solid
+inter-dendritic liquid). Note that using a simple plat-like model for the dendrite shape
Wang and Beckermann [9] quantified the ratio As / Vg as:

As / Vg = 2 / λ2 (14)

where λ2 is the secondary arm spacing.

3.1.2 The grain growth. The envelope model.

In a similar way as in reference [1], an envelope model [9] is used to parameterize the
smallest scale of the dendrite, that is the inter-dendritic spacing. The dendrite envelope is
defined as the virtual surface surrounding the c/e dendrite grain and connecting all c/e
dendrite tips. This leads us to distinguish two types of liquid, the inter- and the extra-
dendritic one, and this for each one of the two grain structures. Therefore one divide the
total liquid into four sub-phases: the equiaxed inter- (de) and extra- (le) dendritic liquid and
the columnar inter- (dc) and extra- (lc) dendritic liquid. The inter-dendritic liquid is
supposed to be in a state of perfect solutal mixing, its solute concentration, Cl* , being
uniquely determined from the local temperature and the phase diagram. On the other hand,
due mainly to its relative large diffusion length scale, the extra-dendritic liquid may be in a
non-equilibrium state with respect to the local solid-liquid interface, meaning that its

IV-9
average concentration Clc / e can be different from Cl* . The consequent local undercooling
Cl* − Clc / e will drive the c/e grain growth which in turn strongly influences the solute
transfer in the extra-dendritic liquid. Since the solidification process is solutally controlled,
special care must be paid to the modeling of the solute transfer at the grain envelope. As
pointed out in [1], due to the equivalence between the length scale of the columnar and
equiaxed extra-dendritic liquids ( R1*c  R1*e − ae ) and the pure diffusion hypothesis, the
solute field in the e/c extra-dendritic liquid will be a function of r only and the e/c cell
boundary will play the role of a symmetry surface between two adjacent grains. Given this,
the solute field in the extra-dendritic liquid was finally approached analytically with a
relative simple expression [1].
However, if fluid flow around the grain is considered, the solute field around the
grain and the grain growth rate can be significantly changed compared with the pure
diffusion case. In the following we will suppose that the fluid flow across the permeable
bed is characterized with low Re numbers, hypothesis generally valid within the c/e mushy
zone. One has to note however that for relative high fluid velocities of the bulk liquid (i.e.
high electromagnetic forces) the low Re hypothesis will not be valid for dispersed c/e
grains (i.e. zone close to the columnar dendrite tips). Hence, the results provided in the
following must be seen as a first order approximation to the grain growth under the
influence of the fluid flow.
It was proven [10] that the fluid flow around a dendrite tip can significantly change
the operating point of the dendrite. Both the stability criterion and the solute field around
the tip are changing. Several studies have been conducted in order to understand the heat
transfer around the advancing tip. A summary has been given by Ananth and Gill in [11].
For a Stokes fluid flow regime an analytical solution of the solute field around a parabolic
tip has been found by Cantor [12]. However the dimensionless undercooling
Ω = (Cl* − C0 ) / Cl* / (1 − k ) and the local tip Peclet number, Pet = Vt rt / Dl are found to be
linked by a very complex implicit function. Wang and Beckermann [13] proposed a fitting
formula to the complex function in [11] that works reasonably well for Sc numbers grater
than the 20 (which is the case of metallic alloys):

b
⎛ Ω ⎞ a = 0.4567 + 0.173Peu0.55
Pet = a ⎜ ⎟ , (15)
⎝1− Ω ⎠ b = 1.195 − 0.145 Peu0.16

where

Peu = v l − v s rt / Dl (16)

is the ambient Pe number computed with the relative fluid-tip velocity. However with the
use of equation (15) the tip velocity cannot be computed explicitly since the tip radius is
not known. Equation (15) must be completed with a stability criterion linking rt and Vt .
For fluid convection cases the marginal stability criterion valid for a pure diffusion case
would not be anymore valid. Experimental evidence exists to confirm this statement [10].
However for low velocity fields one can assume that the influence of the fluid flow on the
stability condition at the tip is not important with respect to the influence on the solute field
around the tip. In this respect supposing that the marginal stability criterion remains valid
one can write for the tip velocity [14]:

IV-10
 Dl m(k − 1)Cl*
Vt = Pet 2 (17)
π Γ
2

More recently, Gandin et al [15] found a more reliable correlation between Ω and Pet .
The approach is less empirical than the one used by Wang and Beckermann [13], equation
(15) in the sense that the kinetic boundary layer around the tip, δ (the stagnant layer) is
modified accordingly with respect to the local Re and Sc numbers. A correlation of the
form

δ = 2rt / ( Sh − 2 )
Sh = 2 + A ⋅ Re B ⋅ Sc C ⋅ f (θ ) (18)
Re = 2rt v l /ν

is proposed by the authors. Moreover, by introducing a correction function f (θ ) their

model can be easily modified to account for the influence of the relative angle θ between
the dendrite growth direction and the fluid flow direction. By coupling the above equation
with the solution of the diffusion field within the stagnant film, Gandin [15] obtained a
better agreement with the exact solution in Ananth and Gill [11] than the model of Wang
[13] did. The main inconvenient of the Gandin correlation is that no explicit expression of
Pet is provided with respect to the tip undercooling Ω . Despite the explicit formulation of
the stagnant film δ , the inverse function Pet = f −1 ( Ω, δ ) has not an explicit form. This
would require a numerical solution for f −1 which could be highly time consuming when
coupled with a macroscopical model. Hence, for reasons of simplicity, the Wang and
Beckermann [13] correlation, equation (15), will be used in the following.

3.1.3 The mass transfer around a dendrite grain

In [1] it was proved that for a pure diffusion case and a local thermal equilibrium at the
scale of the grain, the solute field in the extradendritic liquid will be a function of r only. A
simple analytical form for the solute field in the extra-dendritic is proposed in [1]. If fluid
flow around the grain is accounted, the solute field will not be anymore depending on r
only. For low Re numbers since no hydrodynamic instabilities form behind the spherical
grain, the solute field can be at best approximated with an axisymetric field, Cl ( r , θ ) (see
Figure 5). In the following a scale analysis of the mass transfer around the grain is
conducted.

Figure 5: The solute boundary layer around a dendrite grain

IV-11
The equation controlling the solute transfer within the extra-dendritic liquid is

∂Cl
+ div ( vCl ) = div ( Dl ∇Cl ) (19)
∂t

If one integrates the above equation on the extra-dendritic liquid one obtains the solute
balance in the extra-dendritic liquid:

∂ ∂
∫ Cl dV + Cl* Ag w ge + ∫ ( vCl ) n d Σ =
∂t Vl −∫Vsb
Cl dV +
∂t Vsb Σcell
(20)
= ∫
Σcell
Dl ∇Cl n d Σ + ∫
Σg
Dl ∇Cl n d Σ

where Vsb and Vl are the volume of the solutal boundary layer and of the extra-dendritic
liquid respectively (Figure 5), Σ cell and Σ g are the surface boundaries of the cell and the
grain respectively, Ag is the area of the grain boundary ( = 4π a 2 if the grain is supposed to
be spherical) and finally w ge is the interfacial average velocity of the grain envelope. One
can admit as a first approximation that w ge can be computed with the help of equation (17)
in which the role for the far concentration C0 is played by the average extra-dendritic
liquid concentration C l .
The first two terms on the LHS of the above equation represent the variation of the
solute within the boundary layer around the grain and outside the boundary layer
respectively. The third term on the LHS is no more than the solute loss of the extra-
dendritic liquid due to the grain growth ( w ge ≠ 0 ). Finally the fourth term is no more than
the solute flux at the cell boundary due to the liquid advection. This last term is usually
positive (a solute loss) because the average concentration characterizing the outflow liquid
( C out ) is larger than the mean concentration characterizing the inflow liquid C in . This is
obvious due to the enriching process of the liquid within the wake behind the particle
(Figure 5). Hence, one can rewrite the last term on the LHS of as:

∫ ( vC ) n d Σ = U π R ( C )
2
l 1 out − C in (21)
Σcell

where U is no more than the flow rate per unit area across the permeable bed, equation (7).
Now the first term on the RHS of (20) is the molecular solute flux at the cell boundary.
This term can be however neglected since one expects however that the gradients at the
cell boundary be small compared with the ones at the grain boundary. Finally the last term
on the RHS is no more than the total solute flux at the grain interface. Due to convection
around the spherical grain, the solute flux at the grain interface will be dependent on the
angle θ . Since we are interested for average quantities that could integrate the local effects
of the fluid flow one could rewrite the total flux at the grain boundary as:

Cl* − C0
∫
Σg
Dl ∇Cl n d Σ = Dl Ag
δ cv
(22)

IV-12
where δ cv represent the average diffusion length at the grain boundary and remains to be
explicited from the solute balance equation (20). Similar to the analysis made in [1] one
can suppose the existence of a quasi-steady solute field around the grain growth. In this
respect the parameters characterizing the solute boundary layer remains unchanged with
respect to time. Therefore if a quasi-steady state is considered, the time derivates of the
mean concentration and of the dimensions characterizing the boundary layer are negligible.
Therefore, the solute mass balance in (20) becomes:

Cl* − C0
(
Ag w ge ( Cl* − C0 ) + U π R12 C out − C in = Dl Ag ) δ cv
(23)

Hence the mean diffusion length δ cv becomes

δ cv =
Dl
, c=
(
U π R12 C out − C in ) >0 (24)
w ge + c Ag ( C − C0 )
*
l

It is important to notice that since C out > C in one has always δ cv < Dl / w ge , that is smaller
than the diffusion length correspondent to the pure diffusion case [1]. This was expected
since the effect of the fluid flow can only intensify the solute flux at the grain boundary.
The equation (24) can be rewritten as follows:

(
⎛ R − a ⎞ w ge + c ( R1 − a )
Sh ⎜ = 1
)
⎟=
⎝ δ cv ⎠ Dl
(25)
= Peg +
1R 1
2
( C out − C in ) Pe
4 a 2
(C *
l − C0 )
u

where Peg = w ge ( R1 − a ) / Dl is the Peclet number characterizing the grain growth (see [1])
and Peu = U ( R1 − a ) / Dl is the Peclet number relative to the fluid flow around the grain.
Note as well that in the above equation one has as well

R12 / a 2 = ( ε gk / ε k )
−2 / 3
, k = {c, e} (26)

Equation (25) emphasizes the influence of the fluid flow and the grain growth on the
global solute transfer around the dendritic grain. It is interesting to note as well that, for a
pure diffusion case, one has Shdiff = Peg , an identical result with the one obtained in [1].
To detail further the result in (25) one should be able to quantify the difference C out − C in .
One can expect that C in be close to C0 . In contrast, the mean concentration of the outflow
C out is more difficult to quantify since it directly depends on the flow characterizing the
wake behind the grain. This in turn depends strongly on the local Reynolds number
( Re = URg /ν ) and can be extremely complex for average-high values of Re. An
interesting limit case can be however treated: for ratios a 2 / R12 close to unity (c/e grain

IV-13
fractions close to the maximum c/e fractions) one expects that, due to the effective solute
flux at the grain boundary, the outflow liquid to be in a state close to the perfect solutal
mixing. Hence one has C out  Cl* and the overall solute transfer can be characterized as
follows:

1
( ε gk / ε k ) Peu ,
−2 / 3
Sh = Peg + k = {c, e} (27)
4

Note again that the above correlation would be valid uniquely for ε gk / ε k close to unity. In
fact the above equation has a meaning only for non-zero values of the c/e grain
fraction ε gk .
In contrast for small values of ε gk / ε k , one expects that the solutal interactions between the
grains and the extra-dendritic liquid to be small. Hence the overall fluid flow effect on the
solute transfer at the grain interface could be approximated as being similar to the case of
the mass transfer around a sphere placed in an infinite medium. Several correlations
accounting for this case exists in the literature [16]. A typical scaling law describing the
mass transfer around a spherical object can be written as follows:

Shcv = A Re n Sc n , Re = Ua /ν ; Sc = ν / Dl (28)

Since the diffusion equation describing the solute transfer in the extra-denditic liquid is
linear one can expect that the overall mass transfer around the grain in the limit ε gk → 0 to
be a sum of the effects produced by the grain growth and the fluid flow around the grain
respectively. Therefore one has:

Sh = Peg + Shcv (29)

We have thus tried to approach the analytically the overall mass transfer coefficient at the
grain interface in the limit ε gk → 0 (equation (29)) and ε gk → ε k respectively (equation
(27)). These two scaling laws must be seen only as a first attempt in describing the solute
transfer around a dendrite grain in the presence of convection and a future detailed study
must be conducted in order to better quantify the solute flux at the grain boundary. Despite
the simple approach used before an important conclusion can be drawn: the solute transfer
at the grain interface is intensified by the fluid flow compared with the pure diffusion case
( Sh > Peg ), meaning that:

δ cv < Dl / w ge (30)

One should also note that in an average model C0 is not known a priori. Using however a
similar procedure as in reference [1], by assuming a known concentration profile within the
solutal boundary layer δ cv , one can express C0 as a function of C lk , R1k* and ak
( k = {c, e} ). In this respect the average solute flux at the grain boundary becomes finally
[1]:

IV-14
 Cl* − C lk
ϕ gk = Dl
δ lk − dk
(31)
⎡ 10a 2δ + 10a δ 2 + 4δ 3 ⎤
δ lk − dk = δ ⎢1 − k k
⎥ k = {c, e}
⎢⎣ 5 ( 1k ak 3 ) ⎥⎦
R *3
−

where:

⎧δ cv
⎪ if 2δ cv < ( R1*k − ak )
δ =⎨ * (32)
⎪⎩( R1k − ak ) / 2 if 2δ cv ≥ ( R1*k − ak )

It is important to notice that, the average diffusion length δ lk − dk verifies always:

δ lk − dk < Dl / w ge (33)

a necessary condition for δ lk − dk to be physically meaningful [9].

4 The final averaged equations

Using now the cell model approximation to the ensemble average together with the
physical hypothesis enounced above one is able to compute the unknown terms and tensors
in the averaged equations of mass, momentum, energy and solute, equations.
As pointed out, through the use of the appropriate probability distributions, the new
ensemble model enables one to rigorously model the coexistence of equiaxed and
columnar grains. However one should not confound the c/e phase with the c/e solid. In fact
the c/e phase quantifies the space (liquid + solid) which is under the direct influence of the
c/e structure. Based on the hypothesis that at x a physical field ψ (velocity, solute conc.,
etc.) is mainly influenced by the closest c/e particle to x one can quantify locally the
amount of space corresponding to each of the two structure. By introducing further the
grain envelope model, the c/e phase is further divided into three sub-phases: solid (s) ,
inter-(d) and extra-(l) dendritic liquid. Even if, in reality one deals with a two phase system
(solid + liquid), due to the highly complex morphology of the mixed c/e zone (very
different c/e grain densities, nc ≠ ne ; scale separation between the inter- and the extra-
dendritic zone) one needs to consider separately the evolution of the c/e solid, inter- and
extra- dendritic liquid, that is a six phase multiphase system. Fortunately, the six phase
approach is only needed in describing the solute diffusion problem since this later process
is strongly influenced by the length scales present in the system. Moreover, the solute
diffusion process needs special attention since it controls directly the solidification process
(solid and grain growth).

4.1 The grain balance

As pointed out by Drew [2], the grain balance equation for a given grain distribution can
be written as follows:

∂n
+ div ( nv s ) = nn + nc + nb (34)
∂t

IV-15
where v s is the grain mean velocity, nn the rate of change of the number of particles per
unit volume due to nucleation and nc and nb the rate of change of the number of particles
due to aggregation or breakup phenomena. In reality experimental evidence exists [17] to
support strong equiaxed grain aggregation, especially at high grain fractions. Moreover, the
breakup (fragmentation) of columnar tips has been frequently cited [18, 19] as one of the
most important phenomena causing the columnar-to-equiaxed transition. For reasons of
simplicity we will only account in the following for the nucleation phenomena. Note
however that a future extension of the model in order to account the breakup and
aggregation phenomena will be straightforward. One should only propose appropriate
models for nc and nb . In this respect the c/e grain balance equations becomes:

∂nk
+ div ( nk v sk ) = nk , k = {c, e} (35)
∂t

where the c/e nucleation rate, nk , remains to be modeled.

For the equiaxed grains, similar with the approach used in [1], we will adopt in the
following the instantaneous nucleation model. This model states that the equiaxed
nucleation triggers instantaneously when the local undercooling reaches a certain value,
∆Tne , and is characterized with an homogenous grain density ne0 . Both the nucleation
undercooling ∆Tne and the nucleation grain density ne0 represent input parameters which
have usually a relative high degree of incertitude. In the frame of the new ensemble model
one is able to rigorously account for the nucleation of one structure into another (i.e.
equiaxed nucleation within the inter-granular columnar spacing or the columnar
“penetration” within an existing equiaxed zone). In this respect, besides the evident
equiaxed nucleation within an undercooled pure melt, one is able to account as well for an
eventual equiaxed nucleation within an undercooled columnar zone and more precisely
within the undercooled columnar extra-dendritic liquid. As pointed out in [1], when a grain
structure nucleates within an already existing structure the first one will be partially
blocked by the later one. Indeed the newly germinating structure has only a limited space
for nucleation, that is, the extra-dendritic liquid of the existing structure. This blocking
effect represents no more than the first mechanical blockage (Figure 2) as detailed in [1].
Applied to equiaxed nucleation within the columnar zone, the first mechanical blockage
states that the equiaxed grain density prior to the nucleation phenomenon, ne , will be
smaller than its nucleation density ne0 with a factor of (1 − ε gc ):

ne = ne0 (1 − ε gc ) (36)

where ε gc is the columnar grain fraction at the moment of equiaxed nucleation. In this
respect the equiaxed nucleation rate becomes

ne = ne0 (1 − ε gc ) δ (T − Tliq + ∆Tne ) (37)

where Tliq is the local liquidus temperature.

As pointed out in [1], in contrast with the equiaxed nucleation, the columnar
“nucleation” (the sudden arrival of the columnar dendrite tips at x) cannot be treated as a

IV-16
local event depending solely on local parameters. Indeed, the presence at x of the columnar
tips will be a function of its evolution history and therefore a front tracking technique must
be used in order to determine the precise position of the columnar front (c.f.) at x and at
time t. The model has to be completed therefore with a front tracking equation, describing
the precise evolution of the c.f position with respect to time, xcf (t ) . In this light, the
columnar nucleation can be approached with a local instantaneous germination

nc = nc' δ ⎡⎣ x − xcf (t ) ⎤⎦ (38)

where nc' is the columnar grain density valid after the columnar nucleation within the
equiaxed zone and δ ⎡⎣ x − xcf (t ) ⎤⎦ is the Dirac function pointing to the columnar front
surface. Taking into account the first mechanical blockage effect on the columnar tips,
produced by an eventual equiaxed zone ahead the columnar front (Figure 2) one has as
well:

nc' = nc0 (1 − ε ge ) (39)

where ε ge is the local equiaxed grain fraction at the moment of nucleation and nc0 is the
columnar nucleation density which can be directly related to the primary arm spacing λ1
and therefore to the local cooling conditions: local gradient G and local isotherm velocity
Vis . One could use for λ1 the well known laws of Hunt [20] or Kurtz [21] or the new
developed scaling law in [22] which relates the primary arm spacing to the local velocity
tip gradient at the columnar front,

−1
⎛ ∂V ⎞
λ1 = (16.Dl .trs )
1/ 2
, trs = − ⎜ tip ⎟ (40)
⎝ ∂n ⎠cf

The above scaling law has been derived form the perspective of the constructal theory [23]:
the primary arm spacing characterizing the columnar structure is viewed as the optimum
arm spacing with respect to the local cooling conditions (G and Vis ). A detailed
comparison with various directional experiments shows that, despite its simplicity,
equation (40) provides results which concord remarkable well with various experimental
data.

4.1.1 Columnar front tracking algorithm

The most difficult issue to be resolved in the case of the columnar structure is however the
need for a columnar front tracking. Indeed the implementation of a front tracking technique
would be extremely time consuming. If for one-dimensional case this can be done in a
simple way [19, 24], the extension of a direct front tracking technique to 2D and 3D would
be extremely tedious. Moreover the discrete approach of the front tracking procedure
would not fit in the frame of an averaged multiphase model who’s main advantage is
precisely to avoid the full modeling of the solid-liquid interface. Attempts in coupling an
average model and a discrete columnar front technique were made by Brown [25]. This
later model is limited however to bi-dimensional solidifications and to pure diffusion cases
(no convection). The extension of this model for convection and three-dimensional cases is

IV-17
not straightforward. Moreover a discrete approach for the columnar front would be
difficult to imagine for the modeling of the complex shapes encountered in the case of
solidifications in the presence of convection: freckles and channels characterized by a very
anisotropic shape (large length and very small thicknesses).
An alternative to this approach would be a VOF (volume of fluid) model. The VOF
model attach to each cell (control volume) of the mesh grid a fraction of fluid, in this case
a fraction of the columnar structure, quantifying the fraction of the total control volume
occupied by the columnar phase (Figure 6). Thus each control volume in the mesh grid can
be characterized with a columnar fraction, f c such as if

⎧
⎪1 then the cell is empty (of the col. structure)
⎪⎪
f c (celli ) = ⎨0 then the cell is full (of the col. structure) (41)
⎪ the cell contains the interface between the col. str.
⎪∈ (0,1) then
⎪⎩ and the bulk liquid (or equiaxed zone) ahead

Figure 6: The VOF method

Since the columnar structure is stationary the equation controlling the evolution of f c
becomes:

∂f c
= ScVtip (42)
∂t

where Vtip is the local columnar tip velocity and S s the columnar interfacial area relative to
the control volume ( = Ac / Vcell , the ratio between the interface area within the cell and
volume of the cell). Since the mixed columnar and equiaxed grains share the same
undercooling, similar with the approach in [19, 24], the local columnar tip Vtip can be
directly linked to the local tip velocity:

Vtip = w g (43)

IV-18
As observed (Figure 6) the interfacial area Sc depends both on the columnar structure
fraction f c and the growth direction of the columns. Knowing that columnar dendrites
grow opposite to the local heat flux direction, one has n c = ∇T / ∇T . Since the shape of
the control volume is a priori known the interfacial area can be now computed

Sc = funct ( n c , shape of the cell, f c ) (44)

Due to the non-regular shape of the cell the above function can be quite complex, therefore
a simpler model would be of great interest. The simplest model would be to consider that
there is no preferred growth direction for the columnar grains and to assume a simple
isotropic shape for the cell: a cube of border b aligned with the columnar growth direction.
A measure for b can be easily obtained from the volume of cell

b  Vcell1/ 3 (3D case)

(45)
 Vcell1/ 2 (2 D case)

Obviously the above scaling law would have a physical meaning only if the real shape of
the cell is close to an isotropic one. In this respect and assuming no preferred growth
direction for the columnar grains the interfacial area Sc can be easily expressed as:

Sc = 1/ b (46)

Equation (42) and (46) form a complete model for the evolution of the columnar structure
fraction f c . However, equation (42) would predict the evolution of f c within a control
volume containing already a columnar front interface. One needs as well a model for the
further expansion of the columnar front from one control volume to another. There are
various models to account this issue (the geometric reconstruction scheme, the donor-
acceptor scheme). Due to their high complexity we will instead propose here a very simple
model describing the progress of the columnar front from one cell to another: when a cell
partially filled 0 < f c < 1 finally becomes completely filled with the columnar structure ( f c
reach 1), only then, the columnar tips are supposed to grow outside the cell and the
columnar front progress to the neighbor cells. Notice that Wu and Andreas [26] used a
similar columnar front tracking model in their volume averaged model.
The above front tracking model must be seen only as a first order approximation to
the real columnar front position. However its main advantage is the fact that one can easily
integrate such a front tracking technique within an averaged multiphase model. Indeed,
since one has an insight on the moment when the columnar front passes from one cell to
another, the columnar nucleation can be easily synchronized with this moment: one can
assume that the columnar front nucleation within a cell triggers when the columnar front
reaches the respective cell (at time tcf ). Hence one has:

nc = nc0 (1 − ε ge ) δ ( t − tcf ) (47)

The use of such an approach avoids the need of a discrete modeling of the c.f. interface.
However, the above VOF approach to the columnar front must be used with care since the
method can be extremely dependent on the mesh grid, especially if the controls volume are

IV-19
highly anisotropic, which can be the case for complex unstructured mesh. Moreover, the
accuracy of the prediction of the columnar front position is strongly dependent on the size
of the mesh. Knowing that the above VOF approach will converge to the exact discrete
front tracking approach for:

Vcell → 0 (48)

one should use a very fine mesh in order to minimize the errors in the c.f. prediction. In
turn this would greatly increase the computation time and would limit the use of such a
method to relative small domains. In this respect, the use of an adaptive mesh technique
close to the columnar front would be of significant interest.

4.1.2 The second mechanical blockage

Subsequent to the nucleation phenomena of one structure into another one, the two
coexisting structure are forced to share a same space and consequently the space available
for each structure will be smaller with respect to the space characterizing the two structures
separately (i.e. before their coexistence). As already pointed out, this is no more than the
second mechanical blockage effect of the coexistence: the length scales R1*c / e ,
characterizing the coexisting state between the columnar and the equiaxed grains, are
smaller than the corresponding length scales characterizing the c/e state separately just
before their coexistence, namely R1c / e (see Figure 3). In fact the length scales R1*c / e
together with the two corresponding grain densities nc / e enable one to completely describe
the statistics of the coexistence state and consequently to rigorously compute any average
field relative to the columnar or the equiaxed structure by means of relative simple
integrations within the two corresponding c/e cells (Figure 1). In [1] we have demonstrated
that the two length scales characterizing the c/e coexistence state can be written as follows:
−1/ 3
⎛4 ⎞
R1*c = ⎜ π nc ⎟ εc
⎝3 ⎠
−1/ 3
(49)
⎛4 ⎞
R = ⎜ π ne ⎟
*
1e εe
⎝3 ⎠

where nc and ne represent the columnar and equiaxed grain densities respectively
characterizing the two coexisting structures and ε c and ε e are the columnar and equiaxed
fractions respectively. From a statistical point of view these two fractions quantifies the
way the two coexisting structures shares the space at x. One has obviously:

εc + εe = 1 (50)

In [1] a model for ε c / e was proposed for the case of the columnar nucleation within an
already developed equiaxed zone. As detailed in [1], the two fractions depends solely on
columnar nucleation grain density nc0 , equiaxed grain density, ne and the equiaxed grain
fraction valid at the moment of columnar nucleation, ε ge . One has:

IV-20
 ⎧ ⎡10n + 36 ( n ) ( n 0ε ) 2
3
1
3 ⎤
⎪1 − ε − 1 ⎢ e e c ge
⎥ if R1c ≤ R1e (1 − ε ge1/ 3 )
⎪ ge 0
⎢ ⎥
20nc +45 ( n ) n 0ε
( c ge )
1 2

εc ⎣⎢
3

⎪ ⎦⎥
3
e
=⎨ (51)
(1 − ε e ) ⎪ 0 ⎡( ε )2 − 20ε + 45 ( ε ) ⎤ 2
3

n
⎪ c ⎢ ⎥ R1c > R1e (1 − ε ge1/ 3 )
ge ge ge
if
⎪ 20ne ⎢ −36 ( ε ) + 10 ⎥
1
3

⎩ ⎣ ge ⎦
where:
R1c = ( 4 / 3π nc0 )
−1/ 3
R1e = ( 4 / 3π ne )
−1/ 3
, (52)

It is important to notice that, since the columnar/equiaxed nucleation phenomena within a

developed equiaxed/columnar zone are statistically speaking the same phenomena, the
above model can be straightforward extended to account for the equiaxed nucleation within
a developed columnar zone as well. One has only to replace in equations (51), (52) ne with
nc , nc0 with ne0 and ε ge with ε gc .
Note that besides the constraint in equation(50), the two fractions ε c / e in (51)
verify as well

ε e / c > ε ge / c and therefore ε c / e < 1 − ε ge / c (53)

These two inequalities express one another important constraint of the coexistence state:
the newly germinated c/e phase cannot occupy a space larger than the available space for
its germination existing prior to c/e “nucleation”, namely the e/c extra-dendritic liquid
1 − ε ge / c . Indeed, due to the finite dimension of the e/c grains, the columnar tips can
“nucleate” only within the e/c extra-dendritic liquid and the subsequent columnar fraction
will be therefore always smaller that 1 − ε ge / c .
Moreover, one more important characteristic of the two fractions ε c / e was
identified in [1]. The statistical analysis of the coexistence between the two grain structures
provided us with one important property of the length scales R1*c / e : the two length scales of
the c/e structures, R1*c / e , are such that the length scales of the columnar and equiaxed extra-
dendritic liquids are of the same order of magnitude:

R1*e / c − ae / c  R1*c / e (54)

Therefore, the initial e/c extra-dendritic liquid, 1 − ε ge / c , is shared by the two structures
such as the length scales characterizing the two coexisting extra-dendritic liquids have the
same order of magnitude (Figure 3). In reference [1] the property in equation (54) has been
entirely demonstrated analytically but one can provide as well a physical interpretation. In
fact, since one assumes a homogenous distribution in space of the two coexisting
structures, the columnar and equiaxed solid will also be distributed in space such as the
void spaces between the c/e particles are equal. The two grain distributions rearrange in
space such as the average distance between two neighbor particle, regardless of their nature
(columnar or equiaxed) be equal. There is no reason, given the hypothesis of local
homogeneity, to find some particles closer to their neighbors than other particles are. This
is one key aspect of the coexistence state between two different grains population because

IV-21
the length scales characterizing the two extra-dendritic liquids (le and lc phases) are
significantly influencing the solutal evolution of the two liquids and consequently the
columnar and equiaxed grain growth ( w gc / e ) and solidification rates ( Γ sc / e ). Evidence to
this statement can be found in evolution of the c/e extra-dendritic concentration C lc / e
within the mixed columnar + equiaxed zone behind the columnar front as modeled in [24].
Indeed due to the equivalence between the length scales of the lc and le phases, equation
(54), we have obtained that C lc  C le prior to the “columnar nucleation” into the equiaxed
zone.
Note however that the equation (51) expressing the two fractions ε c / e can be
successfully used in the modeling of the coexistence state existing prior to a columnar or
equaixed nucleation within an already developed equiaxed or columnar zone. In fact, the
two ε c / e fractions computed by means of equation (51) can be successfully used for the
case of pure diffusion case only. Indeed, subsequent to the nucleation of one phase into
another the two fractions ε c / e does not change anymore since the columnar and the
equiaxed grains are stationary. However if convection is accounted, the way the two
structures are sharing the same space at x can be subject to changes. Indeed due to fluid
flow the equiaxed grains are free to move through the ingot and therefore the equiaxed
grain density in one point will be subject to constant changing. In the case of a mixed c/e
zone the change of ne will determine a redistribution of the space between the columnar
and the equiaxed structures and therefore a local change in ε c / e . Since the columnar zone
just behind the columnar zone is very permeable (small columnar grain fractions ε gc ) one
can expect that the fluid flow within the bulk liquid to modify the equiaxed grain density of
the mixed c/e zone behind the columnar front. Obviously, since nc remains constant due to
the static columnar grains, a modified ne would also determine a redistribution of space
( ε c / e ) between the columnar and equiaxed grains. For example if ne increases one expects
that ε e increases and ε c decreases since more equiaxed grains and the same number of
columnar grains are forced to share a same volume ( ε c + ε e = 1 ).

Figure 7: The equiaxed and columnar grains distribute in space such as the average
distance between two neighbor grains, regardless of their nature, be equal

IV-22
 In this respect the question to be answered is: how to compute the new c/e fractions
ε c / e and therefore the corresponding length scales R1*c / e characterizing the statistics of the
coexistence, all this with respect to the new conditions at ne , ε ge , nc and ε gc ? A simple
model for ε c / e is proposed in the flowing.
Note first that a physical meaningful model for R1*c / e and ε c / e would have to verify the
three constrains expressed in equation (50), (53) and (54) respectively, that is:
1. The columnar and equiaxed grains are sharing at x the same space:

εc + εe = 1 (55)
2. The columnar/equiaxed fractions are always larger than their respective grain
fractions:

ε c / e ≥ ε gc / e (56)

3. The two grain distributions rearrange in space such as the average distance between
two neighbor grains, regardless of their nature be equal. Consequently the length
scales of the c/e extra-dendritic liquids are equal (see Figure 7):

R1*c − ac = R1*e − ae = δ l (57)

In the frame of the cell model approximation to the ensemble average, equation (4), the
first two constraints, equations (55) and (56) can be rewritten as follows:

4 4
nc π R1*c 3 + ne π R1*e3 = 1 (58)
3 3
and
4 4
nc / e π R1*c / e3 ≥ nc / e π ac / e 3 ⇔ R1*c / e ≥ ac / e (59)
3 3

respectively. Now from (57) and (58) one can easily compute the two characteristic lengths
R1*c / e and therefore ε c / e . Indeed equation (58) can be rewritten as the sum of the c/e grain
and c/e extra-dendritic liquid fractions,

4 4
nc π ( R1*c 3 − ac 3 ) + ne π ( R1*e3 − ae3 ) = 1 − ε gc − ε ge (60)
3 3

Furthermore, using (57), the above equation can be rewritten as a third degree function of
δl :

4 ⎡ n ⎤
π ncδ l ⎢ (δ l 2 + 3acδ l + 3ac 2 ) + e (δ l 2 + 3aeδ l + 3ae 2 )⎥ = 1 − ε gc − ε ge (61)
3 ⎣ nc ⎦

−1/ 3
⎛4 ⎞
where ac / e = ⎜ π nc / e ⎟ ε gc / e1/ 3 .
⎝3 ⎠

IV-23
Solving the above third order equation and retaining the positive real root only, one finally
obtains:

R1*c / e = ac / e + δ l
4 (62)
ε c / e = nc / e π R1*c / e3
3

Equations (62) where δ l is computed from (61) form a complete model describing the
coexistence state between two different grain populations, in occurrence here between the
columnar and equiaxed structures. Obviously this model verifies implicitly the three
conditions in (55), (56) and (57) respectively.

a) b)
Figure 8: a) the columnar fraction for three different equiaxed fractions: full line, the new
model in equation (62); dashed line, the model in equation (51). b) the columnar fraction
for three different columnar fractions.

A comparison with the model developed in reference [1], equation (51), is

presented in Figure 8a where the columnar fraction, ε c is plotted against n c for a given
n e = 1e10 m −3 and three different equiaxed grain fractions ( ε ge = 0; 0.3; 0.5 ). One can
observe the good agreement between the two models. This was somewhat expected since
the model in equation (51) verified too the condition (57). In Figure 8b the columnar
fraction ε c computed with the new model in equation (62) is plotted against a variable n c
for a given ne = 1010 m3 , ε ge = 0.1 and three different columnar fractions
ε gc = 0.1, 0.3, 0.5 .
It is important to notice that in the frame of this new model, no special condition
need to be imposed in order to commute from a pure e/c zone to a mixed c+e one. Indeed,
the nucleation of one structure into another is intrinsically taken into account by the model
by means of equations (37) and (38). Hence, by knowing at x the c/e grain densites nc / e as
well as their grain fractions ε gc / e one is able to separate the columnar and equiaxed average
effects at x by means of the two c/e cells of radius R1*c / e (Figure 1). In turn R1*c / e can be

IV-24
locally computed at any moment with respect to nc / e and ε gc / e respectively with the help
of equations (61) and (62).

4.2 The mass balance

Similar with the approach in reference [1], by particularizing the generic average equation
(2) with ψ = 1 and X k with X se and X sc respectively one can easily obtain the c/e solid
mass balance:

∂ ( ρ sε sk )
+ div ( ρ sε sk v sk ) = Γ sk , k = {c, e} (63)
∂t

where v sk is the average c/e solid velocity. Since the columnar phase is supposed to be
stationary one has v sc = 0 . Γ sk are the mass exchange rates between the c/e solid and the
liquid ( Γ sk > 0 and Γ sk < 0 means respectively solidification and fusion at x ). Γ sk
represents an interfacial averaged variable and can be expressed analytically using the cell
model approximation similarly with the approach used in [1]. Note again that in view of
the ensemble average the two interfacial averages Γ sk can be computed as relative simple
integrations over the position of the c/e grain in the c/e cell such as the point x is on the c/e
solid interface. As detailed in reference [1] one can rewrite Γ sk as follows:

Γ sk = ρ s S sk wsk , k = {c, e} (64)

where S sk and wsk are the c/e solid-liquid interfacial area and the mean solid-liquid
interface velocity respectively. Both S sk and wsk are parameters extremely difficult to
model due to the complex solid-liquid interface. As we will see, using the separation
between the inter- and the extra-dendritic liquid, one is able to approach analytically the
two solidification rates Γ sk form the solute balance in the inter-dendritic liquid.
It is important to notice that form the point of view of the fluid momentum balance,
the use of the six phase approach requiring four momentum balance equations would be
extremely complex and not practical since the numerical solving of four non-linear
equations would also be extremely time consuming. In this respect a simple approach
would be of great interest. In the following we will use a single fluid approach meaning
that a single fluid momentum balance will be resolved and this for the total fluid phase (f
phase = de + dc + le + lc ). The arguments supporting this choice are further detailed in the
next two subsections. Hence using now ψ = 1 and X k = X f in (2) one easily obtains the
mass balance of the total fluid phase:

∂(ρ fε f ) + div
∂t
(ρ ε
f f v f ) = −Γ sc − Γ se (65)

where v f is the averaged fluid velocity and Γ sc / e the c/e solidification rate identified
above. By solving the three mass balances in (63) and (65) together with the corresponding
momentum balances one obtains the exact evolution of the ε sc , ε se , ε f fractions. However

IV-25
as pointed out, one also need to quantify the inter-dendritic spacing as well. Hence a mass
balance for the c/e inter-dendritic liquid must be considered too. Using this time, ψ = 1 and
X k = X dc / e in (2) one easily obtains

∂ ( ρ f ε dk )
+ div ( ρ f ε dk v dk ) = Γ dk , k = {c, e} (66)
∂t

where v dk is the averaged inter-dendritic velocity and Γ dk the mass exchange rate between
the inter-dendritc liquid, the solid and the extra-dendritic liquid. Indeed, the inter-dendritic
liquid interacts with two phases in the same time, the solid and the extra-dendritic liquid.
This comes from the particular choice of the grain boundary which envelops the dendrite
tips and has no finite interface with the solid (Wang and Beckermann [9]). As detailed in
reference [1] one can rewrite Γ dk as follows:

Γ dk = −Γ sk + Γ gk ; Γ gk = ρ f S gk w gk , k = {c, e} (67)

where Γ gk is the mass exchange between dk and lk phase due to grain growth,
S gk = nk ( 4π ak 2 ) represents the grain interfacial area and w gk is the mean velocity of the
grain envelope, computed from (17). An important question remains however: how to
model the c/e inter-dendritic velocities, v dk . One would be tented to use for v dk the mean
velocity characterizing the inter-dendritic flow. This would not be physically sound since
the inter-dendritic void is not advected by the inter-dendritic flow but on the contrary by
the solid “flow” (movement). Indeed, the inter-dendritic spacing is intrinsically linked to
the solid matrix since it quantifies the amount of fluid captured between the secondary and
tertiary solid arms. In this respect it is evident that the inter-dendritic phase is advected by
the solid flow and not the inter-dendritic fluid flow. One has therefore:

v dk = v sk (68)

This approach is opposite to the one used by Wang and Beckermann model [13]. Indeed,
the later consider that v dk is equal to the average inter-dendritic liquid velocity. We
believe that, for the reasons enounced above, this approach is physically unsound and do
not reflect the fact that the dk phase is moving with the same velocity as the solid (sk
phase). The most evident counterexample to the approach used in [13] is the case of a
columnar mushy zone within which the fluid velocity is non zero. This case is frequently
encountered during directional solidifications process where the fluid flow in the bulk
liquid determine, through the pressure distribution at the columnar macro-front, the
existence of a non-negligible creeping flow within the mushy zone. Following the Wang
and Beckermann approach [13], since v dk will be non-zero the inter-dendritic spacing
( ε dc ) will be advected with the non zero fluid flow within the mushy zone. This is
obviously untrue since the columnar solid and consequently the inter-dendritic voids are
stationary.

IV-26
4.3 The momentum balance
As pointed out before in order to simplify the overall formulation of the model we will
consider in the following only the solid and the total fluid momentum balance equations.
Using in equation (2) ψ = v and X k = X sc , X se , X f one finally obtains after some tedious
manipulations [27] the general averaged momentum balance equation for the columnar
solid, equiaxed solid and the total fluid respectively:

∂ (ε k ρk v k )
+ ∇ ( ε k ρ k v k v k ) = −ε k ∇pk + ∇ ⋅ ε k (τ k + τ k' ) + ε k ρ k b v k +
∂t (69)
+ ( pk ,i − pk ) ∇ε k + Γ k v k ,i + M k , k = {sc, se, f }

Neglecting the surface tension effects the jump equation becomes:

∑ (Γ v
k =s, f
k k ,i + pk ,i ∇ε k + M k ) = 0 (70)

where v k is the average velocity, τ k the average stress, τ k' the fluctuating average stress,
Γ k v k ,i is no more than the momentum exchange rate due to the mass transfer between the
two phases ( Γ v k in equation (2)), v k ,i is the interfacial velocity of the k th phase [27], pk ,i
is the interfacial pressure of the phase k and finally M k is no more than the interfacial
force density Φ v k in equation (2) from which the effect of the average interfacial pressure
was substracted.
As one can easily observe the balance equations in (69) are not closed. One would have to
propose closure expression for: τ k , τ k' , v k ,i , pk ,i and M k with respect to average variables
like v k , pk , interfacial areas, etc. It is important to notice that since the columnar solid
phase is stationary ( vsc = 0 ) one does not need to further consider a momentum equation
for the columnar solid phase. In fact, solving the columnar solid momentum equation
would only reflect the pressure modification inside the columnar solid with respect to the
stress imposed by the interfacial drag at the sc-f interface.

4.3.1 Interfacial pressure

It is usually assumed [2, 27] that the kth phase pressure pk , does not differ much from its
value on the interface, pk ,i , thus having pk = pk ,i . As Drew pointed out [27] this supposes
that there is instantaneous pressure equilibration within the kth phase, which is the case
since the speed of sound in each phase is large compared with the local velocities.
Moreover neglecting the effect of surface tensions one has as well ps ,i = p f ,i . Using the
pressure equilibrium hypothesis above, one finally obtains that:

ps = p f = p (71)

meaning that locally the fluid and the solid share the same pressure.

IV-27
4.3.2 The interfacial velocity v k ,i
The momentum exchange due to phase change can be modeled by introducing an
interfacial velocity v k ,i such as Γ v k (equation (2)) equals Γ k v k ,i [27]. Furthermore Ishii
[28] showed that the difference v s ,i − v f ,i is proportional with the density difference
between the solid and the liquid phase, ∆ρ , and the mass transfer rate Γ k . For
solidification problems, the density difference ∆ρ is relative small compared, for example,
to a liquid-gas two phase flow. Hence, excepting the very rapid solidification processes,
the interfacial velocities in the liquid and the solid become equal. Furthermore, for rigid
and non rotating particles one has as well v s ,i = v s . Therefore one finally obtains:

Γ v sc = 0 ; Γ v se = Γ se v se ; Γ v f = −Γ se v se (72)

4.3.3 The Reynolds stress τ k'

In a two phase flow there are inherent velocity fluctuations about the mean value v . These
are due on one hand to the turbulent regime of the liquid flow and on the other hand due to
the complex solid-liquid interface morphology. Indeed, the flow around a complex solid-
liquid interface will generate spatial fluctuation of v [2]. These fluctuations coupled with
the non-linear character of the advection term in (2) generate a double correlation term
which has a similar form with the molecular flux div ⎡ε k jψ k ⎤ . This term is usually called
⎣ ⎦
the Reynolds stress and it is customary to add it to the real molecular flux. Traditionally,
the Reynolds stress is modeled by means of a turbulent diffusivity which adds to the real
molecular diffusivity (viscosity, thermal or mass diffusivity). The inherent velocity
fluctuations produced by the complex solid morphology transports momentum in a similar
way as the turbulent “eddies” does in turbulent single phase flow. However, since we
supposed a laminar flow, one should focus in the following on the velocity fluctuations
produced by the complex solid morphology only. Using the cell model approximation to
the ensemble average [1] the fluid Reynolds stress can be approached analytically as
follows:

X f ρv 'v '
τ 'f = −
εf
(73)
1 ⎡⎢ ⎤
= nc ∫ X f ρ v ' v ' d z 1 + ne ∫ X f ρ v ' v ' d z 1 ⎥
ε f ⎢ x − z1 ≤ R* x − z1 ≤ R1*e
⎥
⎣ 1c ⎦

Using now the exact fluid field around the permeable c/e dendrite grain as explained in
paragraph 3.1.1, the Reynolds stress can be analytically resolved using the above equation.
The result would be however extremely complex. The fact that given a fluid flow model at
the grain level one can approach analytically the Reynolds stress using the cell model is
however of great interest. For example, by assuming an inviscid flow over a swarm of
solid spheres Drew [2] proposed, using the cell model, an expression for the fluid Reynolds
stress τ 'f :

IV-28
 1
τ 'f = − ε s ρ f ⎡⎣( v f − v s )( v f − v s ) + 3 ( v f − v s )( v f − v s ) I ⎤⎦ (74)
20ε f

Note however that the models for τ k' are quite complex. This was somewhat expected
since the main role for τ k' is to recover the relevant information regarding the fluid flow at
the micro-scale, lost with the averaging process. The Reynolds stress should be therefore
representative for the complex solid morphology of the two phase flow.
In the following we will however neglect the dispersive flux. The arguments for this
hypothesis are:
ƒ at small grain fractions (i.e. close to the columnar dendrite tips) the Reynolds stresses is
negligible with respect to the advection terms since the influence of the small particles
on the fluid flow remains negligible;
ƒ on the other hand, for grain fractions larger than 0.4 − 0.5 (i.e. developed equiaxed
zone or profound columnar zone) the fluid flow becomes heavily perturbed by the solid
particles and the velocity fluctuations becomes important too. Moreover since we have
chosen a single fluid formulation (inter- + extra-dendritic liquid) the velocity
fluctuations will further be amplified. Indeed, due to the low inter-dendritic
permeability to fluid, the inter-dendritic velocity will be usually much lower that the
velocity characterizing the extra-dendritic liquid. Hence, the velocity fluctuations
around the mean v f will be important and will in the end reflect a non-negligible
Reynolds stress with respect to the non-linear advection term. In fact this is one of the
main disadvantages of the unique fluid formulation with respect to a separate
description of inter- and extra-dendritic liquid. The Reynolds stress / advection term
ratio corresponding to the separate momentum balance in the inter- and extra-dendritic
liquid will be much smaller than the same ratio corresponding to the single fluid flow
formulation. However, one should also take into account the fact that for relative high
grain fractions ε g > 0.4 − 0.5 the bed permeability is sufficiently low as the fluid flow
to be close to a creeping flow state characterized by very small velocities. Hence, the
non-linear advection term in equation (69) and the Reynolds stress become negligible
in front of the interfacial stress M k and the pressure gradient. In this respect, excepting
a small range in the grain fraction ε g = 0.1 − 0.4 , the Reynolds stress can be negliged
since either it is too small in front of the advection term (for small ε g ), either it is
negligible compared with the preponderant terms: M k and the pressure gradient (for
average-high ε g ).
ƒ for typical grain densities nc / e = 106 − 109 m −3 and cooling rates the grain growth can be
very rapid from small values ( < 0.1 ) to relative high ones ( ε g > 0.4 − 0.5 ) due to small
solutal interactions between the grains [24]. In this respect it is expected that the
Reynolds stress to have a negligible influence even at average grain fractions.

All these findings entitles us to neglect the term τ k' in the momentum balance equation
(69). Notice that Wang and Beckermann [13] proposed a model which takes into account
the scale separation between the inter- and the extra-dendritic liquid. By summing the
separate momentum balances in the inter- and extra-dendritic liquid Wang and
Beckermann [13] obtained the total fluid balance equation. Supposing further a creeping
flow across the dendrite bed, the authors managed to obtain a model for the dispersion flux

IV-29
associated to the differences between the inter- and the extra-dendritic liquid velocities.
However, it is important to notice that by assuming creeping flow across the bed, the
convective terms and the Reynolds stresses are negligible small with respect to the
pressure gradient and interfacial stress. Therefore the development of a Reynolds stress
model having at the base the hypothesis of a creeping flow would not be entirely justified
physically.

4.3.4 The average viscous stress τ k

Having as guiding principle “the principle of frame indifference” Drew [2] proposed a
general constitutive equation for the average stress τ k . The principle of frame indifference
states that [2] the constitutive equations are invariant under a change of the reference frame
since the way in which a two-phase flow distributes forces must be independent of the
coordinate system used to express it. Drew [2] identified a list of tensors that are objective
regarding the change of coordinate system:

v ki = v k − v i

Dk =
1
2
(
∇v k + ( ∇v k )
T
) (75)

⎛ ∂v ⎞ ⎛ ∂v ⎞
a ki = ⎜ k + v i ⋅∇v k ⎟ − ⎜ i + v k ⋅∇v i ⎟
⎝ ∂t ⎠ ⎝ ∂t ⎠

In this respect one should look for constitutive relations for τ k on the form [2]:

τ k = µkk Dk + µki Di + Ek11 v ki v ki + Ek12 ( v ki ∇ε k + ∇ε k v ki )

(76)
+ Ek 13 ( v ki a ki + a ki v ki )

where µkk , µkk , Ek11 , Ek12 , Ek13 are coefficients to be determined. As one can see a precise
closure equation for the average stress would be extremely complex. Drew [2] assumed
that in the low velocity regime, when the inertia term are negligible small, one should
retain in equation (76) only the linear terms in the velocity obtaining thus:

τ k = µkk Dk + µki Di + Ek12 ( v ki ∇ε k + ∇ε k v ki ) (77)

It is important to notice that the above constitutive equation has a similar formulation with
the result of Ishi [28] or Ni and Beckermann [29]. In the following we will neglect the
higher order terms in (77) by assuming that µki and Ek12 are negligible small in front of
µkk . We obtain therefore:
1
τ k = µkk ∇v k + ( ∇v k )
2
(T
) (78)

For the fluid phase it is customary to consider µ ff = 2 µ f . In contrast for the solid phase, if
the collision between particles is not important then µ ss = βµ f where β > 1 since the solid
particles are expected to concentrate the fluid shear stress [2]. In fact their presence results
in a larger effective viscosity of the solid-liquid mixture. Indeed, for dilute suspension of

IV-30
solid particles ( ε g  1 ) one has the well known Einstein law where β = 7 / 2 . In contrast,
for compacted bed of particles ( ε g > 0.65 ) the solid phase become a rigid structure and
µ ss → ∞ . For the present model we will use the same approach for µ ss as Wang and
Beckermann did [13], that is:

⎡⎛ ε ⎞−2.5ε g ⎤
0

µ
µ ss = 2 f ⎢⎜ 1 − g ⎟ − (1 − ε g ) ⎥ (79)
εg ⎢⎜⎝ ε g0 ⎟⎠ ⎥
⎣ ⎦

where ε g0 = 0.637 and ε g is the grain fraction of the analyzed solid phase (equiaxed in our
case). The above law was proposed initially by Krieger [30] and has the advantage that for
ε g → 0 , µ ss = 3.5µ f and for ε g → ε g0 , µ ss → ∞ . Note however that, applied to a case
where columnar and equiaxed particles coexists, the equiaxed grain fraction in equation
(79) should be scaled to the maximum available space for the equiaxed phase, that is ε e .
Indeed, since the equiaxed phase occupy only a fraction of the space at x ( ε e ) one should
consider for the ε g in (79) not the absolute equiaxed grain fraction ε ge but the internal
equiaxed grain fraction:

ε g ( = ε gi ) = ε ge / ε e (80)

since the available space for the equiaxed grains is ε e < 1 and not 1.
If the transport of particle momentum is dominated by the particle collision the
solid viscosity can be related to the particle fluctuating kinetic energy usRe [2]. This is
somewhat similar with the kinetic theory of gases where the molecular viscosity is related
to the mean kinetic energy. Note however that for the case of equiaxed solidification the
collisional effect on the momentum transport is not of great importance since the collisions
between dendritic particles is far from being elastic (or near elastic). In turn one should
expect that phenomena like dendrite breakup or particle aggregation to have a significant
effect on the momentum transport.

4.3.5 The interfacial momentum transfer M k

As pointed in [2] the interfacial momentum transfer M k contains the forces on the solid
phase due to viscous drag, wake and boundary layer formation, and unbalanced pressure
distributions leading to lift or virtual mass effects. For simplicity reasons we will retain
hereafter only the contribution of the drag force. In this respect the interfacial force
momentum transfer M f will a sum of the drag force of the fluid flow on the columnar and
equiaxed particles respectively. Note that the fluid flow around a c/e particle and the drag
force on the c/e particle are known from the micro-scale assumptions (paragraph 3.1.1).
Hence one obtains for M f :

M f = M fe + M fc
(81)
= ne 6πµ f ae Ω e ( v se − ε fe v fe ) + nc 6πµ f ac Ω c ( −ε fc v fc )

IV-31
where Ωc / e were defined in (11), v fc and v fe are the columnar and equiaxed fluid
velocities respectively and finally ε fc / e are the c/e fluid fractions ( = ε dc / e + ε lc / e ). Note
however that the above equation is defined with respect to the c/e fluid velocities. These
two velocities are not known a priori since we have chosen a single fluid approach and
only the total fluid velocity, v f is known. One could however quantify the c/e fluid
velocities if one assumes creeping flow across the mixed c/e zone. Indeed, in the low
velocity range the average momentum balance equations within the c/e liquid become:

−ε fc ∇p + nc 6πµ f ac Ωc ( −ε fc v fc ) = 0
and (82)
−ε fe∇p + ne 6πµ f ae Ωe ( v se − ε fe v fe ) = 0

The above equations represent no more than the Darcy approximation to the momentum
balance. Using equations (82) and knowing that ε f v f = ε fc v fc + ε fe v fe one can easily
obtain:

ε f ε fe v f K c + ε fc v se K e
v fc = (83)
ε fe ( ε fe K c + ε fc K e )
and
K e ( ε f v f − v se )
v fc = (84)
ε fe K c + ε fc K e

where K c = nc 6πµ f ac Ω c and K e = ne 6πµ f ae Ω e . Finally the interfacial momentum transfer

M f becomes:

K c K eε f ( v se − ε f v f )
Mf = (85)
ε fe K c + ε fc K e

As well, the interfacial momentum transfer M se corresponding to the equiaxed solid

becomes:

M se = −M fe
K c K eε fe ( v se − ε f v f ) (86)
=−
ε fe K c + ε fc K e

4.4 The energy equation

Similar to the approach used in reference [1] and given the local thermal equilibrium, the
energy balance can be approached with a unique mixture equation. Since the density and
the thermal conductivities are common to both equiaxed/columnar liquid and solid phases,
the energy balance equation is identical with the one obtained in [1] in which the advection
term has to be added. One can easily obtain:

IV-32
 ∂ ⎡⎣( ε s ρ s cs + ε f ρ f c f ) T ⎤⎦
+ div ⎣⎡ε se ρ s cs v se + ε f ρ f c f v f ⎦⎤ =
∂t (87)
= div ⎡⎣( ε s λs + ε f λ f ) ∇T ⎤⎦ + Γ s L

where ε s = ε se + ε sc and Γ s = Γ se + Γ sc .

4.5 The solute balance equations

Using ψ = C , jCk = ρ D∇C and X k = X sc , X se , X f , X dc , X de in (2) one obtains:

∂ ( ε k ρ k Ck )
∂t ⎣ (
+ div ( ε k ρ k v k Ck ) = div ⎡⎢ε k jC k + j′C k ⎤⎥
⎦ ) (88)
+Φ C
k +Γ C
k ; k = {sc, se, f , dc, de}

that is a total of five equations. Notice that the solute balances in the c/e inter-dendritic
liquid are explicit ones since the solute concentration C dc / e is fixed at the equilibrium
concentration Cl* computed with respect to the local temperature. Notice as well that one
can correctly compute the solute balance within a certain phase only if the corresponding
velocity verifies the mass balance within that phase. Since we have chosen a single fluid
approach for the momentum balance we are somewhat constrained to consider the solute
balance equation within the total fluid and not the two solute balance in the c/e fluid phase
( fc, fe ). In fact, even if we have approximated the c/e fluid velocities, equation (83) and
(84), these does not verify the mass balance within the c/e fluid phase.

4.5.1 The mean flux jC k

Since the solute concentration fluctuations within the phase k is

Ck' = C − C k (89)

one can rewrite the average stress as follows:

ε k jC k = X k ρ D∇C
= ε k ρ D∇C + ρ D∇ ( X k Ck' ) − ρ DCk' ∇X k (90)
 
=0 ≈0

It is easy to see that the second term on the RHS of the above equation is zero. Noting that
∇X k = −n k δ (x − xi ) where δ (x − xi ) is the Dirac delta function pointing at the interface,
the last term in (90) is no more than the interfacial average of the solute concentration
fluctuations, averaged on the phase k interface. In turn the solute concentration at the solid-
liquid interface is uniquely determined by the temperature, Cl* = (T − T f ) / m . Due to the
high thermal diffusivity of the liquid metals one expects to have small thermal
perturbations at the grain scale. In this respect this last term can be neglected in front of the
first term on the RHS of (90). The average flux becomes:

IV-33
 ε k jC k = ε k ρ D∇C k (91)

Notice that the above average flux is similar with results obtained by various authors [13,
29].

4.5.2 Dispersion flux j′C k

Since the solute concentration inside the inter-dendritic liquid is equal to Cl* and the
temperature fluctuations at the scale of the grain are negligible one can neglect as well the
two inter-dendritic dispersive flux: j′C dc and j′C dc . As well, due to the extremely low solute
diffusivity in the solid the two dispersive fluxes characterizing the c/e solid phase can be
neglected too.
In contrast, due to the difference between the concentrations in the inter- and extra-
dendritic liquid, the dispersion flux j′C f characterizing the total fluid phase may not be
negligible. In fact this is the main disadvantage of choosing the total fluid approach. In the
flowing however we will neglect this term and the arguments supporting this hypothesis
are:
ƒ for high grain densities or low cooling rates due to the strong solutal interactions
between grains the extra-dendritic liquid reach a state of perfect solutal mixing
( Clc / e  Cl* ) very rapidly after the beginning of the solidification (see reference [24]).
Hence the total fluid concentration C f will be equal to Cl* and consequently the solute
fluctuations will be negligible small similar to the case of the inter-dendritic liquid.
ƒ in contrast, for low grain densities or average-high cooling rates the solutal interactions
between adjacent grains are negligible small and the extra-dendritic liquid remains at
its initial concentration C0 until the grain fraction ε g becomes close to unity (in the
case of c/e coexistence until ε gc / e becomes close to ε c / e ). Only then the solutal
interactions between grains become non-negligible and the extra-dendritic liquid
enriches in solute to finally approach Cl* . In this respect the dispersion solute flux j′C k
cannot be neglected during this transition stage since the differences between the inter-
dendritic concentration ( Cl* ) and the extra-dendritic one ( C0 ) are significant. However
one should note that the time scale of this transition phase is small due to the rapid
grain growth (see reference [24]). Usually this time scale is smaller than the time scale
characterizing the convection within the mushy zone. In this respect neglecting the
dispersive flux j′C k would not change significantly the solute transport at the scale of
the ingot.

4.5.3 The interfacial sources Φ C k and ΓC k

The interfacial sources Φ C k and ΓC k remains practically unchanged with the ones already
computed in reference [1]. Note however that in [1] the hypothesis of no convection was
assumed. Here the fluid flow is accounted. As already pointed out the fluid flow across a
permeable bed is influencing the solute mass transfer at the grain boundary. Basically the
overall mass transfer coefficient is increased with respect to the pure diffusion case. It was
proven in equation (29) that Sh > Peg and that:

IV-34
 ⎛ Dl ⎞
δ cv < δ diff ⎜ = ⎟ (92)
⎝ w ge ⎠

Hence in the terms Φ C k and Γ C k obtained in [1] one should only replace the diffusion
length δ lk − dk (k=c,e) with the diffusion length valid for the convection case, equation (31).
One obtains:

Φ Csk = 0 ΓCsk = Cs∗Γ sk

ρ f S gk Dl ∗ , k = {e, c} (93)
Φ Clk =
δ lk − dk
( Cl − Clk ) ΓClk = −Cl∗Γ gk

and finally,
ρ f S gk Dl ∗
Φ Cdk + ΓCdk = Cl∗Γ gk − Cs∗Γ sk −
δ lk − dk
( Cl − Clk ) , k = {e, c}
(94)
Φ + Γ = −C ( Γ sc + Γ se )
C
f
C
f
∗
s

Notice that since Cdc / e = Cl* the solute balance in the c/e inter-dendritic liquid
transform into equations for the solidification rate Γ se and Γ sc . Indeed, one has:

∂ ( ε dk ρ f Cl* )
+ div ( ε dk ρ f v dk Cl* ) = div ( ε dk ρ f D f ∇Cl* ) + Cl∗Γ gk
∂t (95)
ρ f S gk Dl ∗
∗
− C Γ sk −
s
δ lk − dk
( Cl − Clk ) ; k = {c, e}

Using the mass balance within the inter-dendritic liquid the above equation become
further:

∂Cl* ρ f S gk Dl ∗
Γ sk ( Cl* − Cs∗ ) = ε dk ρ f + ( Cl − Clk ) +
∂t δ lk − dk (96)
+ε dk ρ f v dk div ( C ) − div ( ε dk ρ f D f ∇C
*
l
*
l ); k = {c, e}

Usually the macroscopic diffusion flux (the last term on the RHS of the above equation)
can be neglected in front of the solute flux at grain boundary. Evidence to this hypothesis
can be found in [9]. Notice that the inter-dendritic velocity v dk is not a priori known. Since
the inter-dendritic fluid permeability is known, equation (13), one could approach these
velocities from the Darcy approximation of the momentum equation written for the inter-
dendritic liquid. For simplicity reasons we will approach the inter-dendritic liquid velocity
from the c/e fluid velocities v fc / e . We will suppose in the following that the inter- and
extra-dendritic liquids share a same mean velocity:

v dk ( = v lk ) = v fk (97)

where v fk is computed with the help of equations (83) and (84).

IV-35
 ρ f S gk Dl ∗
As one can see the solute flux at the c/e grain boundary,
δ lk − dk
( Cl − Clk ) , is
expressed with respect to the mean concentration Clc and Cle . The later ones are not
explicitly known. However, as pointed out in [24], due to the equivalence between the
length scales characterizing the c/e extra-dendritic liquid ( R1*e − ae  R1*c − ac ) the two
liquids will follow a similar solutal evolution during the c/e coexistence. Therefore one
expects to have Clc  Cle . This is a physical sound hypothesis since the two structures
finally share the same liquid around the grains. In this respect, one can now easily compute
the two concentrations Clc and Cle as follows:

ε f C f − (ε dc + ε de )Cl*
Clc ( = Cle ) = (98)
ε f − (ε dc + ε de )

At this moment the whole system formed by the balance equations is closed.

4.6 Summary of the model

In the following the summary of the model equations is presented.

4.6.1 The grain balance equations

The equiaxed grain balance is:

∂ne
+ div ( ne v se ) = ne0 (1 − ε gc ) δ (T − Tliq + ∆Tne ) (99)
∂t

where ne0 is the equiaxed nucleation density [m3], ε gc is the columnar grain fraction at the
moment of nucleation, Tliq is the local liquidus temperature, ∆Tne the equiaxed nucleation
undercooling and δ (T − Tliq + ∆Tne ) the Dirac function pointing to the equiaxed nucleation
temperature.
The columnar grain balance is:

∂nc
= nc0 (1 − ε ge ) δ ⎡⎣ x − xcf (t ) ⎤⎦ (100)
∂t

where nc0 is the columnar nucleation density which can be directly related to the primary
arm spacing λ1 by means of equation (40), ε ge is the local equiaxed grain fraction at the
moment of nucleation and δ ⎡⎣ x − xcf (t ) ⎤⎦ is the Dirac function pointing to the columnar
front surface, xcf (t ) . This in turn can be computed by means of a front tracking algorithm
as detailed in paragraph 4.1.1.

4.6.2 The mass balance

The columnar/equiaxed solid mass balance is:

IV-36
 ∂ ( ρ sε se ) ∂ ( ρ sε sc )
+ div ( ρ sε se v se ) = Γ se ; = Γ sc (101)
∂t ∂t

where v se is the average velocity of the equiaxed solid and Γ se and Γ sc the equiaxed an
columnar mass transfer rates respectively. The latter ones will be computed from the solute
balance in the e/c inter-dendritic liquid.
The fluid mass balance is:

∂(ρ fε f ) + div
∂t
(ρ εf f v f ) = −Γ se − Γ sc (102)

where v f is the average fluid velocity.

The inter-dendritic mass balance is:

∂ ( ρ f ε de ) ∂ ( ρ f ε dc )
+ div ( ρ f ε de v se ) = −Γ se + Γ ge ; = −Γ sc + Γ gc (103)
∂t ∂t

where v se is the average velocity of the equiaxed solid, and Γ gk is the mass exchange
between dk and lk phase due to grain growth:

Γ gk = ρ f S gk w gk ; k = {c, e} (104)

where S gk = nk ( 4π ak 2 ) represents the grain interfacial area and w gk is the mean velocity
of the grain envelope, computed from (17).

4.6.3 Momentum balance

The fluid momentum balance is:

∂ (ε f ρ f v f )
∂t ⎢⎣ (
+ ∇ ( ε f ρ f v f v f ) = −ε f ∇p + ∇ ⎡ε f µ f ∇v f + ( ∇v f )
T
)⎤⎥⎦ (105)
+ε f ρ f g − Γ se v se + K f ( v se − ε f v f )
where
K c K eε f
Kf = (106)
ε fe K c + ε fc K e

is the momentum transfer coefficient at the phase f interface and K c / e = nc / e 6πµ f ac / e Ωc / e

where Ωc / e is computed by means of equation (11). The coefficient K f is no more than a
measure of the drag force acting at the c/e solid interface and can be expressed as well as
an inverse of an overall permeability to fluid.
The equiaxed solid momentum balance is:

IV-37
 ∂ ( ε s ρ s v se )
∂t ⎣ (
+ ∇ ( ε s ρ s v se v se ) = −ε s ∇p + ∇ ⎡ε s µ s ∇v se + ( ∇v se ) ⎤
T

⎦ ) (107)
+ε s ρ s g + Γ se v se − K se ( v se − ε f v f )
where

K c K eε fe
K se = (108)
ε fe K c + ε fc K e

and µ s is the equivalent solid viscosity:

⎡⎛ ε i ⎞ −2.5ε g ⎤
0

µ ε ge
µs = if ⎢⎜1 − ge ⎟ − (1 − ε ge ) ⎥ ;
i
ε gei = (109)
ε ge ⎢⎜⎝ ε g0 ⎟⎠ ⎥ εe
⎣ ⎦

4.6.4 The energy balance

The energy balance equation is:

∂ ⎡⎣( ε s ρ s cs + ε f ρ f c f ) T ⎤⎦
+ div ⎣⎡ε se ρ s cs v se + ε f ρ f c f v f ⎦⎤ =
∂t (110)
= div ⎡⎣( ε s λs + ε f λ f ) ∇T ⎤⎦ + Γ s L

where cs and c f are the solid and liquid mass specific heats, L the latent heat, and
Γ s = Γ se + Γ sc .

4.6.5 The solute balance

The solute balance in the columnar/equiaxed solid is:

∂ ( ε se ρ s Cse )
+ div ( ε se ρ s v seCse ) = div ( ε se ρ s Ds ∇Cse ) + Cs∗Γ se
∂t
and (111)
∂ ( ε sc ρ s Csc )
= div ( ε sc ρ s Ds ∇Csc ) + Cs∗Γ sc
∂t

The solute balance in the total fluid is:

∂ (ε f ρ f C f )
+ div ( ε f ρ f v f C f ) = div ( ε f ρ f D f ∇C f ) − Cs∗ ( Γ se + Γ sc ) (112)
∂t

4.6.6 Supplementary equations

The e/c solid mass transfer rates Γ se and Γ sc are computed from the solute balance
equations in the inter-dendritic liquid:

IV-38
 ∂Cl* ρ f S gk Dl ∗
Γ sk ( C − C ) = ε dk ρ f
*
l
∗
s + ( Cl − Clk ) +
∂t δ lk − dk (113)
+ε dk ρ f v fk ∇Cl* ; k = {c, e}

where v fk are computed from (83) and (84), S gk = nk ( 4π ak 2 ) , δ lk − dk is computed with the
help of equation (31) and Clk are computed from (98).

4.6.7 Auxiliary expressions

Auxiliary expressions (k={c,e})
Constraints: ε le + ε de + ε se = ε e ε de + ε se = ε ge
, εc + εe = 1,
ε lc + ε dc + ε sc = ε c ε dc + ε sc = ε gc
Interfacial ⎧kC ∗ if Γs ≥ 0 Cl∗ = (T − T f ) m
concentrations: Cs∗ = ⎨ l
⎩Cs if Γs < 0
Grain radius ak = ( 4 / 3π nk )
−1/ 3
ε gk 1/ 3
Equiaxed/colum ε e , ε c = f ( ne , nc , ε ge , ε gc ) , see equation (62)
nar fractions
Cell radius R1*k = ( 4 / 3π nk )
−1/ 3
ε k1/ 3
Grain envelope S gk = 4π ak 2 nk
interfacial area
Diffusion length δ lk − dk , see equation (31)
Grain growth Dl m ( k − 1) Cl∗ 2
rate Γ gk = ρ f S gk w gk , w gk = ⎡⎣ Pet ( Ω k ) ⎤⎦
π Γ2

Notice that the cell dimensions R1k* and the equiaxed and columnar fractions ( ε e
and ε c ) are computed locally at each moment with respect to the model presented in
paragraph 4.1.2, equation (62).

Observations
As one can see, for reasons of simplicity the total fluid formulation is chosen for both the
momentum and the solute balance. Note however that the main disadvantage of this
approach is the explicit formulation of the interfacial sources in the fluid solute balance
equation as well as the explicit formulation for Clc and Cle , equation (98). As already
discussed in [24] the solute enriching of the extra-dendritic liquid can be extremely rapid
due to the highly stiff character of the solute diffusion problem within the extra-dendritic
liquid for high cooling rates and low grain densities. The explicit formulation of the Clc
and Cle in the inter-dendritic solute balance equations may determine convergence
problems especially when the mass transfer rates Γ sc and Γ se are computed, equation
(113). One should therefore look for an implicite formulation for Clc and Cle . The easiest
way to do this is to further consider the solute balance equations in the c/e extra-dendritic
liquid. One has:

IV-39
 ∂Clk
ε lk ρ f + ε lk ρ f v fk ∇Clk = ∇ ⋅ ( ε lk ρ f D f ∇Clk ) +
∂t
(114)
⎛ρ S D ⎞
+ ⎜ f gk l − Γ gk ⎟ ( Cl∗ − Clk ) , k = c, e
⎝ δ lk − dk ⎠

As one can notice the above formulation allows one to easily implicite the solute rejection
in the extra-dendritic liquid, the last term on the RHS of (114). Moreover since
ρ S D
δ lk − dk < Dl / w ge one has as well f gk l > Γ gk . Hence, the continuous enriching of the
δ lk − dk
extra-dendritic liquid is ensured. However solving the above equations and the solute
balance in the total fluid would be redundant. In fact, one should not try to solve them
separately but to couple at each time step equations (114) and the solute balance in the total
fluid, equation (112). This can be done quite easily if analyzing the discretized form of
(114) with respect to a finite volume method:

( Clk )i − ( Clk )
n n −1

Vcell ( ε lk ρ f ) + ( ε lk ρ f v fk ) ∑
n n
( C )σ nσ Sσ =
n
lk
i ∆t i
σ
n
(115)
⎛ ρ f S gk Dl ⎞
= ∑ ( ε lk ρ f D f ∇Clk ) − Γ gk ⎟ ⎡⎢( Cl∗ ) − ( Clk )i ⎤⎥ , k = c, e
n n n
⋅ nσ Sσ + Vcell ⎜
⎝ δ lk − dk ⎠i ⎣ ⎦
σ i
σ

where Vcell is the volume of the mesh cell, σ the number of the cell face, nσ the exterior
normal to the σ th cell face, Sσ the area of the σ th cell face, n the nth time iteration and i
the ith loop iteration. As one can notice, the main advantage of the above formulation is the
implicite approach of the last RHS term of (115). Indeed, for high cooling rates or low
grain densities, during the rapid growth of the grain, δ lk − dk can become very small. Hence
an implicit approach for the solute rejection term in (115) would greatly help to the
robustness of the algorithm. Now, to couple at each time step equation (115) with the
solute balance equation in the total fluid, equation (112), one has only to express ( Clk )
n −1

with respect to the last converged value of C f , that is ( C f )

n −1
. One obtains:

n −1
⎡ ε f C f − (ε dc + ε de )Cl* ⎤
( Clk )
n −1
=⎢ ⎥ (116)
⎢⎣ ε f − (ε dc + ε de ) ⎥⎦

This coupled algorithm would enable the implicit computation of Clc / e and consequently
an implicit formulation for Γ sc / e in (113).

5 Conclusions
The ensemble averaged model derived in reference [1] has been extended to account for
the fluid convection. We have approached the ensemble average through the probability
distributions functions attached to the macroscopic process to be studied. Indeed, the main
advantage of the ensemble averaging is that, with the help of the nearest neighbor
( )
distributions if c / e t , x; z 1 (see [1]), one can quantify in a rigorous way the c/e influence at

IV-40
x. Using the cell approximation to if c / e , two spherical cells centered at x and of radius R1c*
and R1e* were identified, each of them characterizing in a simple way the statistics of the
c/e grain distributions. The new model inherits form the model in [1] which quantified in a
rigorous manner the solutal and mechanical (geometrical) interactions between the two
coexisting grain distributions.
The microscale physical assumptions are reevaluated with respect to the ones in
[1]. First, the fluid flow around the c/e permeable grains is analyzed. Using the Neal [6]
model we managed to quantify the drag force acting on the c/e grain. This in turn is used to
compute the interfacial momentum transfer at the solid-liquid interface. Secondly the grain
growth in the presence of fluid convection is analyzed. Despite the accuracy of the Gandin
model [15], for reasons of simplicity, we use in this study the Wang and Beckermann [13]
correlation for the tip kinetics. Finally, the solute transfer around the c/e grain is analyzed.
Here a scale analysis of the mass transfer around the grain is conducted. It is concluded
that the mass transfer at the grain boundary in the presence of convection is intensified
with respect to the pure diffusion case. We propose correlation for the computation of the
overall mass transfer coefficient in the limit of small and high c/e grain fractions.
Using the microscale physical assumptions the average balance equations are
further derived. First the grain balance equations are derived. If the equiaxed nucleation is
modeled as an instantaneous event, the columnar nucleation (the sudden arrival of
columnar tips in one point) cannot be approached with a local event. Hence, a columnar
front tracking algorithm needs to be considered. We therefore propose a simple front
tracking technique having as basis the volume of fluid model. Moreover the coexistence
model developed in [1] is here extended to account for the grain convection. Indeed, the
cell dimensions R1c* and R1e* as well as the space partition between the columnar and
equiaxed distribution ( ε c and ε e ) is here determined form a new general model describing
the coexistence state between two different grain populations, in occurrence here between
the columnar and equiaxed structures. The mass, momentum and energy balance equations
are derived next. Notice that for simplicity reasons a total fluid approach is considered for
the momentum balance. For the energy equation a single mixture balance equation is
derived. Finally the solute balance equations are obtained. Here, a total fluid approach is
considered too along with the two solute balance equations in the c/e solid. Moreover the
two solute balance equations in the c/e inter-dendritic liquid are used to compute the c/e
solidification rates. We have also proposed an algorithm for the implicit computation of the
extra-dendritic liquid concentrations. This is done by coupling at each time step the solute
balance equations in the c/e extra-dendritic liquids with the solute balance equation in the
total fluid (f phase). Finally a complete ensemble averaged model is proposed for the
modeling of the mixed columnar + equiaxed structures in the presence of convection.

6 Nomenclature
Symbol Meaning Unit

v velocity field [m/s]

C solute concentration field [wt. %]
T temperature field [K]
n grain density [ m −3 ]
ρ density field [ kg .m −3 ]
ψ the generic physical field ---

IV-41
εk the “volume” fraction of phase k ---
fc the columnar fraction attached to the mesh cell ---
w gk the interface velocity of the k phase interface [m/s]
Ck* the interfacial concentration at the k phase interface [wt. %]
Xk the k phase characteristic phase function ---
nc0/ e the columnar/equiaxed nucleation grain density [ m −3 ]
λ1 columnar primary arm spacing [m]
Sk the interfacial area density of phase k [ m −1 ]
δk− p the solute diffusion length at the k-p interface [m]
xcf position of the columnar front [m]
Tne the equiaxed nucleation temperature [K]
Tf the fusion temperature of the pure solvent [K]
m the slope of the liquidus line [ K . ( wt % ) ]
−1

k the partition coefficient ---

ck heat capacity for the k-phase [ J .kg −1.K −1 ]
λk the phase k thermal conductivity [ W .m −1.K −1 ]
L the latent heat [ J .kg −1 ]
Ω the dimensionless local undercooling ---
Pet the dendrite tip Peclet number ---
Dk the solute diffusivity within the phase k [ m 2 .s −1 ]
ac / e the local columnar/equiaxed grain radius [m]
R1*c / e the cell radius corresponding to the coexisting [m]
columnar/equiaxed grain distribution
z i the position in the ingot of the ith closest grain to x [m]

c/e (
if (1) t , x; z 1 ) the unconditional density function that the closest grain
to the point x is centered within d z 1 of z 1 and is a
[ m −3 ]

columnar/equiaxed grain
Γk the k phase mass transfer rate [ kg.m −3 .s −1 ]
Over lines
ensemble average
 relative to the nearest neighbor distributions
Subscripts
c columnar phase (lc+dc+sc)
e equiaxed phase (le+de+se)
s solid phase
l extra-dendritic liquid phase
d inter-dendritic liquid phase
f total fluid phase ( = l + d )
g grain phase ( = d + s )
kc columnar k phase
ke equiaxed k phase

IV-42
7 References
1. Ciobanas, A.I., and Y. Fautrelle: "Ensemble Averaged Two-Phase Eulerian Model for
Columnar/equiaxed Solidification of a Binary Alloy. Part I. The Mathematical
Model.", Materials Science and Engineering A, (Submitted 2006).

2. Drew, D.A., and S.L. Passman. Theory of Multicomponent Fluids. . New York:
Springer-Verlag, 1999.

3. Carman, P.C. Flow of Gases Through Porous Media. . London: Butterworth Scientific,
1956.

4. Happel, J., and H. Brenner. Low Reynolds Number Hydrodynamics. . Leyden, The
Netherlands: Noordhoff International Publishing, 1973.

5. Happel, J.: AiChE Journal, 1958, vol. 13, pp. 122-125.

6. Neale, G., N. Epstein, and W. Nader: "Creeping Flow Relative to Permeable Spheres.",
Chemical Engineering Science, 1973, vol. 28, pp. 1865-1874.

7. Davis, R.H., and H.A. Stone: "Flow Through Beds of Porous Particles.", Chemical
Engineering Science, 1993, vol. 48.23, pp. 3993-4005.

8. Wang, C.Y, et al.: "Multiparticle Interfacial Drag in Equiaxed Solidification.",

Metallurgical and Materials Transactions B, 1995, vol. 26B, pp. 111-119.

9. Wang, C.Y, and C. Beckermann: "A Multiphase Solute Diffusion Model for Dendritic
Alloy Solidification.", Metallurgical and Materials Transactions A, 1993, vol. 24A, pp.
2787-2802.

10. Huang, S.C., and M.E. Glicksman: Acta Metallurgica, 1981, vol. 29, pp. 701-715.

11. Ananth, R., and W.N. Gill: Journal of Crystal Growth, 1991, vol. 108, pp. 173-189.

12. Cantor, B., and A. Vogel: "Dendritic Solidification and Fluid Flow.", Journal of
Crystal Growth, 1977, vol. 41, pp. 109-123.

13. Wang, C.Y, and C. Beckermann: "Equiaxed Dendritic Solidification With Convection:
Part I. Multiscale/multiphase Modeling.", Metallurgical and Materials Transactions A,
1996, vol. 27A, pp. 2754-2764.

14. Lipton, J., M.E. Glicksman, and W. Kurz: "Dendritic Growth Into Undercooled Alloy
Melts.", Materials Science and Engineering, 1984, vol. 65, pp. 57-63.

15. Gandin, Ch.-A., et al.: "Boundary Layer Correlation for Dendrite Tip Growth With
Fluid Flow.", Materials Science and Engineering, 2003, vol. A342, pp. 44-50.

16. Burmeister, L.C. Convective Heat Transfer. . 2nd ed. New York: Wiley, 1993.

IV-43
17. Beckermann, C., and C.Y Wang: "Equiaxed Dendritic Solidification With Convection:
Part Iii. Comparison With Nh4cl- H2o Experiments.", Metallurgical and Materials
Transactions A, 1996, vol. 27A, pp. 2784-2795.

18. Jackson, K.A., et al.: "On the Origin of the Equiaxed Zone in Castings.", Transactions
of the metallurgical society of AIME, 1966, vol. 236, pp. 149-158.

19. Martorano, M. A., C. Beckermann, and Ch.-A. Gandin: "A Solutal Interaction
Mechanism for the Columnar-To-Equiaxed Transition in Alloy Solidification.",
Metallurgical and Materials Transactions A, 2003, vol. 34A, pp. 1657-1674.

20. Hunt, J.D.: "Cellular and Primary Dendrite Spacings.", Solidification and casting of
metals - Metals Society, London, 1979, pp. 3-9.

21. Kurz, W., and D.J. Fisher: "Dendrite Growth at the Limit of Stability: Tip Radius and
Spacing.", Acta Metallurgica, 1981, vol. 29, pp. 11-20.

22. Ciobanas, A.I., A. Bejan, and Y. Fautrelle: "Dendritic solidification morphology

viewed from the perspective of constructal theory.", Physical Review E, (Submitted
2006).

23. Bejan, A. Shape and Structure, From Engineering to Nature. . New York: Cambridge
University Press, 2000.

24. Ciobanas, A.I., and Y. Fautrelle: "Ensemble Averaged Two-Phase Eulerian Model for
Columnar/equiaxed Solidification of a Binary Alloy. Part II. Simulation of the
Columnar-To-Equiaxed Transition.", Materials Science and Engineering A, (Submitted
2006).

25. Browne, D.J., and J.D. Hunt: "An Interface Tracking Model of Moving Boundaries in
Multi-Phase Systems: Application to Solidification.", Archives of thermodynamics,
2003, vol. 24.1, pp. 25-35.

26. Wu, M., and A. Ludwig: "Volumne Averaging Concept for Mixed Columnar-Equiaxed
Solidification.", Proceedings of McWASP XI, 2006.

27. Drew, D.A.: "Mathematical Modeling of Two-Phase Flow.", Ann. Rev. Fluid Mech.,
1983, vol. 15, pp. 261-291.

28. Ishii, M. Thermo-Fluid Dynamic Theory of Two-Phase Flow. . Paris: Eyrolles, 1975.

29. Ni, J., and C. Beckermann: "A Volume-Averaged Two-Phase Model for Transport
Phenomena During Solidification.", Metallurgical Transactions B, 1991, vol. 22B, pp.
349-361.

30. Krieger, I.M.: Adv. Colloid Interface Sci., 1972, vol. 3, pp. 111-136.

IV-44