Acta mater. Vol. 47, No. 3, pp.

971±982, 1999
# 1999 Acta Metallurgica Inc.
Published by Elsevier Science Ltd. All rights reserved
Printed in Great Britain
PII: S1359-6454(98)00380-2 1359-6454/99 $19.00 + 0.00

THREE-DIMENSIONAL MODELING OF EQUIAXED

DENDRITIC GROWTH ON A MESOSCOPIC SCALE
I. STEINBACH1, C. BECKERMANN2, B. KAUERAUF1{, Q. LI2 and J. GUO2
ACCESS e.V., D-52056 Aachen, Germany and 2Department of Mechanical Engineering, The
1

University of Iowa, Iowa City, IA 52242-1527, U.S.A.

(Received 13 May 1998; accepted in revised form 29 October 1998; accepted 1 November 1998)

AbstractÐA novel mesoscopic modeling technique has been developed to simulate the unsteady growth of
multiple equiaxed dendritic grains into a supercooled melt of a pure substance. In the model, the numerical
calculation of the temperature ®eld in the supercooled melt between the grains is coupled with a stagnant-
®lm model for dendrite tip growth, such that without resolving individual dendrite arms the evolution of
the grain envelope and the internal solid fraction can be predicted. The simulations are in good agreement
with experiments for the growth of a single dendritic grain of the model substance succinonitrile. The
model is then applied to simulate the growth of various con®gurations of up to 14 strongly interacting
grains. The results indicate that the use of local analytical solutions in numerical calculations is a viable
technique for simulating large-scale dendritic growth phenomena. # 1999 Acta Metallurgica Inc. Published
by Elsevier Science Ltd. All rights reserved.

1. INTRODUCTION The development of a model for simulating the

Dendrites are the most common growth mor- growth of an assemblage of equiaxed dendritic
phology in solidi®cation of metals, alloys, and other grains on a mesoscopic scale is the objective of the
substances [1]. Dendritic growth is characterized by present study. Of primary interest are prediction of
the propagation of a primary tip and the nonlinear the evolution of the grain shapes, the growth inter-
evolution of secondary and tertiary side-branches actions between multiple grains, and the nature of
on a microscopic scale (from 10ÿ6 to 10ÿ5 m). the thermal ®eld in the melt between the dendrites.
Equiaxed grains result if the growth is from free For example, for strongly overlapping thermal ®elds
in front of the dendrite tips of neighboring grains,
nuclei into a supercooled melt. An assemblage of
the tip velocities can be expected to be reduced, up
many equiaxed dendritic grains constitutes a so-
to a point where a globulitic (or spheroidal) grain
called mushy zone on a macroscopic scale (10ÿ1 m).
structure results. On the other hand, for weak inter-
Modeling of equiaxed dendritic solidi®cation on the
actions the tip velocities will be similar to those for
scale of a typical mushy zone thus requires the sim-
dendrite growth into an in®nite melt [3].
ultaneous consideration of growth and transport
The direct numerical simulation of dendritic
phenomena over length scales ranging over roughly
growth inside a mesoscopic unit cell cannot be
®ve orders of magnitude. The mesoscopic scale is
achieved with presently available computational
de®ned here as being between the microscopic and
power, because the range of length scales that need
macroscopic length scales, and is of the order of
to be resolved is too large. A recent example illus-
from 10ÿ4 to 10ÿ3 m. A mesoscopic unit cell inside
trating the present limits of direct numerical micro-
a mush would typically contain of the order of 10
structure simulation is the fully resolved, three-
equiaxed grains and literally thousands of dendrite
dimensional calculation of dendritic growth by
arms and tips. Such unit cells or representative el-
Karma and Rappel [4] using the phase-®eld
ementary volumes (REVs) form the basic building
method. Their results are limited to a relatively
blocks of macroscopic or volume-averaged solidi®-
small equiaxed grain, without a full side-branch
cation models used in the simulation of casting
structure, and relatively high dimensionless super-
processes [2]. Accurate modeling of the dendritic
coolings. The extension to fully dendritic grains
growth processes inside a mesoscopic unit cell is not
growing at the low supercoolings typical of casting
only important for the prediction of the ®nal grain
processes would require an increase in compu-
structure of a solidi®ed material, but also for feed-
tational power by at least three orders of magnitude
ing local information (e.g. latent heat evolution,
in both speed and memory. On the other end of the
solute rejection) back to the macroscopic model.
spectrum, relatively simple approximate and one-
dimensional unit cell models of equiaxed dendritic
{To whom all correspondence should be addressed. solidi®cation have been proposed [2, 5] that are
971
972 STEINBACH et al.: THREE-DIMENSIONAL MODELLING

primarily intended for the prediction of the latent been attempted before, but seems to be a plausible
heat evolution during recalescence. Obviously, these idea given the large separation of length scales in a
models are unable to resolve the actual shape of the dendritic mush. While the present study focuses on
growing grains and the internal structure of the the development of the basic model for equiaxed
mush. One bridge over the gap between direct dendritic solidi®cation of a pure substance, an
microscopic simulation and the one-dimensional extension to alloys and other growth structures will
unit cell models is the so-called cellular automaton be straightforward.
modeling of grain growth by Gandin and The model and the numerical solution procedures
Rappaz [6]. In this method, the orientation and are described in the Sections 2 and 3. Then Sections
evolution of the shape of the growing grains is 4 and 5 provide results for a single equiaxed grain
tracked numerically, allowing for a prediction of and a comparison of the predictions with available
the ®nal grain structure in a solidi®ed part. experimental data. Then, in Section 6, numerical
However, in a manner similar to the one-dimen- results are presented for the growth of various con-
sional unit cell models, the dendrite growth vel- ®gurations of up to 14 interacting grains, which is
ocities are calculated based on some local mean followed by the conclusions in Section 7.
supercooling, and the temperature (or concen-
tration) ®eld between the grains is not resolved. 2. MODEL AND GOVERNING EQUATIONS
Consequently, the cellular automaton model does
not allow for a detailed study of the local growth A schematic illustration of the various length
interactions. scales present in equiaxed dendritic growth is
The modeling of equiaxed dendritic solidi®cation shown in Fig. 1(a). The smallest, microscopic length
on a mesoscopic scale in the present study is facili- scale is denoted by d1, and is of the order of the
tated by a novel approach that combines a numeri- dendrite tip radii (from 10ÿ6 to 10ÿ5 m). The den-
cal solution of the relevant transport equation (i.e. drite tips grow into the supercooled melt surround-
the heat di€usion equation) on the mesoscopic scale ing the grain. The di€usion boundary layers around
with a local analytical solution of the dendrite tip the tips have a thickness of the same order as the
growth problem at the smallest length scale [7]. tip radii. The prediction of the tip growth speeds
With this method, the computational power and radii requires the resolution of the thermal ®eld
requirements are reduced by at least three orders of on the scale d1, something we accomplish in the
magnitude compared to direct microstructure simu- present model using a local analytical solution (see
lation on a microscopic scale. In general, the use of below). The mesoscopic scale d2 is of the order of
a local analytical solution or empirical relation in a
numerical simulation rests on the notion that the
system behavior at the smallest length scale is ``uni-
versal'' and independent of the longer range inter-
actions. Such an approach is commonly utilized in
the modeling of turbulent ¯uid ¯ow, where the
e€ect of the smallest eddies on the mean ¯ow is rep-
resented by relatively simple relations that are valid
for a large variety of ¯ows. One of the earliest
examples is the mixing-length theory [8], which is
based on the assumption that the ®nest-scale turbu-
lence structure is similar everywhere throughout the
¯ow ®eld. Another example is the use of the ``uni-
versal'' law-of-the-wall in modeling various turbu-
lent wall shear ¯ows [8]. The law-of-the-wall is
applied in a thin ``inner'' layer along the surface of
a body, corresponding to a few grid points in a nu-
merical simulation, and matched with the mean vel-
ocity ®eld in the ``outer'' region away from the
wall, which is obtained from a solution of the rel-
evant transport equations on the system scale. The
key point is that the same law-of-the-wall can be
used for large variations in the outer ¯ow.
Similarly, in the present model, the same analytical
relation is employed for all (active) dendrite tips to
describe the growth behavior on the smallest length
scale, and this relation is then matched with the Fig. 1. Schematic illustration of equiaxed dendritic growth:
outer temperature ®eld. The use of this fundamental (a) unit cell, where d1 is the microscopic scale and d2 is the
concept in modeling dendritic solidi®cation has not mesoscopic scale; (b) grain envelope and stagnant ®lm.
 STEINBACH et al.: THREE-DIMENSIONAL MODELLING 973

the spacing between the grains, which is typically where G is the Gibbs±Thomson coecient and s*
two orders of magnitude greater than d1. The pre- is a selection constant. Combining equations (1)
diction of the overall shape of a grain and the and (2) to provide an explicit relation of the form
growth interactions between the grains inside the v = f(DT) can be accomplished by numerical inver-
unit cell requires the resolution of the thermal ®eld sion and curve ®tting. It is important to realize
on the scale d2, which is done here by solving nu- that this local analytical solution is only used on
merically the transient, three-dimensional heat the microscopic scale, d1. Hence, the ®lm thick-
equation in the supercooled melt. ness, df, is chosen to be of the same order of mag-
Central to the coupling of the microscopic and nitude as, but greater than the tip radius R. For
mesoscopic transport processes in the present model df>>R, equation (1) would reduce to the Ivantsov
is the de®nition of a grain envelope, in a manner solution.
similar to previous models [2, 5, 6]. As shown in For a given supercooling across the ®lm, the
Fig. 1(b), the grain envelope is a smooth surface above local analytical solution can be used to calcu-
connecting the tips of all ``active'' or ``surviving'' late the speeds of every active (primary and second-
dendrite branches, where a branch is judged as ary) dendrite tip. Because the exact locations of the
active when it is longer than the next active branch dendrite side-branches are not known and the grain
closer to the primary tip. Therefore, the growth vel- envelope is a continuous surface, the tip speeds are
ocities of the grain envelope can be obtained from calculated for every point on the envelope. For this
dendrite tip speeds, as described below. Note that purpose, the same ®lm thickness df is chosen for all
the volume inside the grain envelope contains ®nely tips. Then, as shown in Fig. 1(b), a confocal grain
``dispersed'' solid and melt. The temperature of the envelope can be de®ned that is located the distance
solid/liquid mixture inside this very form-®tting df away from the regular grain envelope. As
envelope can safely be assumed to be at the melting opposed to the regular grain envelope, the tempera-
point, Tm. Consequently, all growth takes place at ture of the confocal envelope, Tce, is not uniform,
the envelope surface and the latent heat is con- because the tip speeds or, equivalently, the super-
ducted into the supercooled melt between the coolings across the ®lm, DT = TmÿTce, are di€erent
grains. for every active dendrite tip. The distribution of Tce
The local analytical solution used to resolve the is obtained from the numerical solution of the heat
temperature ®eld on the microscopic scale d1 is the equation on the mesoscopic scale, as explained
so-called ``stagnant-®lm'' modi®cation by Cantor below. Note that the choice of a uniform df leading
and Vogel [9] of the well-known Ivantsov solution to the de®nition of the confocal grain envelope is
of the heat ¯ow problem around a growing, iso- simply a matter of convenience in the numerical al-
thermal paraboloid of revolution representing a gorithm, and one could have chosen a variable df
dendrite tip. When coupled with a selection corresponding to an isothermal confocal envelope.
criterion [2], this solution provides both the den- For later use, the tip speeds are converted to nor-
drite tip speed, v, and radius, R, as a function of mal envelope velocities, vn, using the following re-
the supercooling DT. The regular Ivantsov solution lation
cannot be used because it is only valid when the vn ˆ vn
cosy 3†
melt is in®nite in extent. In the stagnant-®lm sol-
ution, the supercooling is instead applied at a con- where n is the exterior normal to the envelope and
focal isothermal paraboloid located a ®nite y is the angle between the normal and the growth
distance df away from the dendrite tip and moving axis of the closest dendrite arm. The calculation of
with the same speed as the tip. The stagnant-®lm the normal is described in the next section. The
solution is given by [9] angle is obtained by specifying the crystallographic
orientation of the grain and assuming cubic aniso-
L tropy.
DT ˆ Pe exp Pe†
c The temperature distribution on the mesoscopic
     scale, d2, is obtained by solving numerically the fol-
df
 E1 Pe† ÿ E1 Pe 1 ‡ 2 1† lowing transient, three-dimensional heat di€usion
R
equation in the supercooled melt between the grain
where DT is the supercooling across the stagnant envelopes in the unit cell:
®lm of thickness df, Pe = vR/(2a) is the tip Peclet @T
number, a is the thermal di€usivity, L is the latent ˆ ar2 T 4†
@t
heat, c is the speci®c heat, and E1 is the exponen-
tial integral function. The selection criterion The boundary conditions are the ®xed melting tem-
needed in addition to equation (1) to obtain a perature, Tm, at (and inside) the moving grain
unique value of the tip speed can be written as [2] envelope and an adiabatic unit cell. Initially, the
melt is taken to be at a uniform temperature T1
2aG corresponding to an initial supercooling of
R2 v ˆ 2†
s* L=c DTi=TmÿT1.
974 STEINBACH et al.: THREE-DIMENSIONAL MODELLING

As mentioned above, the numerical solution of well approximated by the numerical solution on the
the heat equation provides the confocal grain envel- mesoscopic scale. If, on the other hand, the ®lm
ope temperature distribution, which can formally be thickness is too large, the local analytical solution
written as can become invalid due to thermal ®elds from
neighboring dendrite branches and other grains
Tce ˆ Tjconfocal envelope 5†
penetrating into the ®lm. Nonetheless, the relatively
Furthermore, the knowledge of the temperature large separation of the length scales d1 and d2
gradient at the envelope from the numerical sol- enables one to ®nd a ®lm thickness that works for a
ution allows for the calculation of the fraction of reasonably wide range of supercoolings, which is
solid, fs, formed at the envelope using the following demonstrated below.
energy balance:
@T L
ÿa ˆr vn fs 6†
@n envelope c
3. NUMERICAL SOLUTION PROCEDURES
Because the envelope continually propagates in an The numerical method used to solve the model
outward direction and the temperature gradient var- equations consists of two parts: the envelope propa-
ies with both time and location on the envelope, the gation algorithm and the solution of the heat di€u-
solution of equation (6) will establish a certain solid sion equation. The coupling of both establishes the
fraction distribution within the grain. solution of the mesoscopic model.
The overall structure of the mesoscopic model as Equations (1)±(3) provide the normal grain envel-
stated above is summarized in Fig. 2, which shows ope velocity, vn, as a function of the supercooling
a schematic of the temperature pro®le normal to a across the ®lm. In order to propagate the envelope
grain envelope. The local analytical solution is used across the regular and ®xed numerical grid, we use
to resolve the temperature pro®le in the stagnant a phase-®eld-like algorithm [10]. A ®eld variable, f,
®lm for the calculation of the dendrite tip/grain is introduced that varies across the ®lm from 1 in
envelope velocity, while the numerical solution of the grain near the envelope to 0 in the supercooled
the heat equation provides the temperature distri- melt away from the confocal envelope. It should be
bution on the mesoscopic scale outside of the grain emphasized that this ®eld variable is simply used as
envelope and also the confocal envelope tempera- an indicator function for purely numerical reasons,
ture. The two solutions are matched both at the and has no physical meaning. The grain envelope is
regular and the confocal grain envelope. The nu- assigned a value of f = 0.95, while the confocal
merical solution should only be viewed as an ap- envelope is placed at f = 0.5. The transition region
proximation in the ®lm region, because it does not for f, de®ned as the distance df over which f var-
resolve the microscopic temperature ®eld around in- ies from 0.05 to 0.95, is chosen to be twice the ®lm
dividual dendrite tips. Proper matching of the ana- thickness df. The ®eld variable f allows for the cal-
lytical and numerical solutions depends primarily culation of the normal to the envelope according to
on the choice of the ®lm thickness, df. If the ®lm
thickness is too small, the steep microscopic tem- rf
nˆÿ 7†
perature gradients near the dendrite tips are not jrfj

Fig. 2. One-dimensional temperature distribution normal to a grain envelope illustrating the matching
of the numerical solution and the local analytical solution at the stagnant ®lm.
 STEINBACH et al.: THREE-DIMENSIONAL MODELLING 975

The propagation of f is given by terial and the selection constant, s*, is reasonably
36 well known [11±13].
f_ ˆ f2 1 ÿ f†2 vn DT † ‡ stab f† 8† As a ®rst test, the mesoscopic model is used to
df
predict the tip velocity of a single dendrite arm. In
where the dot above f denotes the time derivative this limiting case, the Ivantsov theory, coupled with
and stab(f) is a stabilization operator. This oper- the selection criterion, provides an analytical predic-
ator acts as an anti-di€usion ¯ux to keep the f-pro- tion of the tip speed. A comparison between the
®le compact and is given by theoretical prediction and the calculations is shown
" 2  # in Fig. 3(a). The open squares correspond to meso-
df 1 scopic model calculations with a ®lm thickness of
stab f† ˆ C n
rjrfj ÿ f 1 ÿ f† ÿ f 9†
72 2 200 mm and a grid spacing of 100 mm. It can be
seen that the model results are in reasonable agree-
where C is a stabilization constant. Taking C1 0.1/Dt ment with the theory for supercoolings ranging
results in magnitudes of the anti-di€usion ¯uxes from 0.2 to 1.0 K. It is important to note that the
that work for any time increment Dt. The f-pro®le tip radius varies signi®cantlyÐfrom 112 mm (at
is compact when stab(f) = 0, which leads to
the following stable pro®le of f in the normal direc-
tion n
 
1 3n
f n† ˆ 1 ÿ tanh 10†
2 d
Again, the basic structure of the above equations is
essentially borrowed from the phase-®eld
method [10] and found to work well as a propa-
gation algorithm for the grain envelope. The formu-
lation ensures that the envelope is a smooth surface
and is propagated at the correct velocity regardless
of the orientation of the f-surfaces (i.e. the grain)
with respect to the grid. Equation (8) is solved
using standard ®nite di€erences. The grid is uniform
and chosen such that at least two grid points lie
within the ®lm of thickness df.
The heat di€usion equation, equation (4), is
solved using a standard control volume discretiza-
tion technique. The temperatures of all cells inside
the grain envelope, i.e. all cells with fr 0.95, are
set to the melting temperature, Tm. The confocal
envelope temperatures, Tce, are obtained from the
calculated temperature ®eld by interpolation to
f = 0.5. Finally, equation (6) is discretized and
solved at each time step to update the solid fraction
distribution in the grains.

4. RESULTS FOR A SINGLE GRAIN

Before validating the mesoscopic model using ex-

perimental data and studying the growth inter-
actions of multiple grains inside a unit cell, it is
useful to examine the predictions from the model
for the limiting case of a single grain growing into
an essentially in®nite melt (i.e. uniform far-®eld Fig. 3. Predicted results for a single dendrite (SCN) grow-
ing into a uniform supercooling: (a) comparison of pre-
supercooling). This discussion is intended to clarify
dicted dendrite tip speeds for various supercoolings with
the coupling between the local analytical solution the Ivantsov theory; q show the results for a ®lm thick-
and the numerical solution of the heat equation, ness of 200 mm and a grid spacing of 100 mm; the error
and shed light on the proper choice of the ®lm bars correspond to variation in the grid spacing from
thickness df. All thermophysical properties are 100 mm to 50 mm (w) and to variation in the ®lm thickness
from 200 mm to 400 mm (q); (b) partitioning of the total
taken to be those of the commonly used metal- supercooling between the stagnant ®lm and the melt out-
model material succinonitrile (SCN), mainly side the confocal envelope as a function of the dendrite
because experimental data are available for this ma- tip speed.
976 STEINBACH et al.: THREE-DIMENSIONAL MODELLING

0.2 K) to 15 mm (at 1.0 K)Ðin this supercooling lier, the ®lm thickness should generally be of the
range. Nonetheless, the ®lm thickness of 200 mm same order of magnitude as, but larger than the tip
gives good results over the entire range. The agree- radius, for the use of the local analytical solution to
ment somewhat deteriorates above 0.8 K, which can be meaningful. A series of calculations was per-
be attributed to the ®lm thickness becoming too formed in order to investigate in more detail the
large relative to the tip radius. As mentioned ear- sensitivity of the results to the numerical grid

Fig. 4. Predicted envelope shape for a single equiaxed grain growing into a uniform supercooling: (a)
octahedric shape that results when the heat di€usion equation is not solved and all dendrite tips grow
at the same velocity; (b) and (c) two views of a fully coupled prediction; (d) cut through a primary
branch of the grain depicted in (b) and (c).
 STEINBACH et al.: THREE-DIMENSIONAL MODELLING 977

spacing and the ®lm thickness. The error bars in remainder is outside the confocal envelope where it
Fig. 3(a) were obtained by varying the grid spacing is resolved by the numerical solution. For a total
by a factor of two, i.e. from 100 mm to 50 mm, while supercooling of 0.2 K, the portion contained within
keeping the ®lm thickness constant at 200 mm and the ®lm is less than 20%. Despite these variations
by varying the ®lm thickness by a factor of two, i.e. in the partitioning of the supercooling, the tip vel-
from 200 mm to 400 mm, while keeping the grid spa- ocities are predicted well, as discussed above.
cing constant at 100 mm. The resulting tip speeds While the above comparison establishes the mini-
vary up to 225%. Note that a ®lm thickness of mum ®lm thickness needed to predict the correct tip
400 mm in one variation is more than 25 times speeds, it does not address the ability of the meso-
greater than the tip radius at 1.0 K. Such a large
scopic model to predict the thermal interactions
ratio is, of course, not recommended.
between adjacent dendrite branches and produce a
The partitioning of the total supercooling is
realistic grain envelope shape. The establishment of
further illustrated in Fig. 3(b) for the same con-
the grain envelope shape is illustrated in Fig. 4. The
ditions as in Fig. 3(a). A ®lm thickness of 200 mm
implies that up to a distance of 200 mm from the calculations are started with a spherical grain envel-
dendrite tip, the temperature ®eld is resolved by the ope that has a radius approximately equal to the
local analytical solution, while beyond 200 mm it is ®lm thickness. Figure 4(a) shows the octahedral
resolved by the numerical solution of the heat envelope that is predicted by the envelope propa-
equation. It can be seen from Fig. 3(b) that for the gation algorithm when the same tip speed (50 mm/s)
present ®lm thickness of 200 mm, the portion of the is used everywhere. In other words, the octahedral
total supercooling contained within the ®lm shape results when the envelope propagation algor-
increases with tip speed. For example, for a tip ithm is decoupled from the numerical solution of
speed of 85 mm/s about 70% of the total supercool- the heat equation and the confocal envelope tem-
ing of 0.8 K is contained within the ®lm, while the perature (0.4 K here) is uniform around the grain.

Fig. 5. Superimposed SCN dendrite images taken at four di€erent times during the IDGE microgravity
experiment of Glicksman and coworkers [11±13] ¯own on the USMP-2 mission (DT = 0.370 K); the
time interval between the images is 83.25 s.
978 STEINBACH et al.: THREE-DIMENSIONAL MODELLING

It can be seen that the algorithm produces the experimental validation of the predicted envelope
expected shape. shape is provided in the Section 5.
The corresponding solution, where the full meso-
scopic model is solved and the envelope propa-
gation algorithm is coupled to the numerical 5. EXPERIMENTAL VALIDATION FOR A SINGLE
GRAIN
solution of the heat equation, is shown in Fig. 4(b)±
(d). Due to reduced heat ¯uxes at the faces of the The mesoscopic model is validated for a single
octahedron, compared to the corners, the tempera- equiaxed dendritic grain by comparing predicted
tures of the confocal grain envelope there are closer grain envelope shapes and solid fractions to
to the melting temperature, resulting in lower super- measurements from an experiment involving di€u-
coolings and growth speeds at the faces. sion-controlled growth of a pure substance in a uni-
Conversely, the heat ¯uxes from the corners are formly supercooled melt. The experiment is the
greater, resulting in lower confocal envelope tem- Isothermal Dendritic Growth Experiment (IDGE)
peratures and higher tip speeds. Consequently, a of Glicksman and coworkers [11±13], launched on
realistic looking equiaxed grain envelope evolves the space shuttle by NASA in March 1994 as part
that has six primary branches at right angles. of the United States Microgravity Payload (USMP-
Figure 4(d) shows a cut through one of the primary 2). The IDGE uses pure SCN which is contained
branches. Due to di€erent heat ¯uxes at the edges inside a temperature-controlled growth chamber.
and sides and, hence, di€erent speeds of the second- After melting the SCN and establishing the desired
ary and tertiary dendrite arm tips, a realistic look- supercooling level, the growth of a dendrite is in-
ing cross-sectional shape evolves. A detailed itiated on a stinger by activating a thermoelectric

Fig. 6. Dendrite arm A reconstructed from the superimposed images in Fig. 5 and rotated to lie in a
side-branch plane; the interrupted lines are the parabolas ®tted to the tip.
 STEINBACH et al.: THREE-DIMENSIONAL MODELLING 979

cooler. During growth, photographs are taken at branch is judged active when it is longer than the
regular time intervals along two perpendicular opti- next active branch closer to the primary tip.
cal paths. The image analysis techniques used to Figure 7 shows measured active secondary tip
extract the desired data from the photographs are positions obtained from experimental runs at four
described in detail in Ref. [14]. di€erent supercoolings (DT = 0.287 K, 0.370 K,
Figure 5 shows four superimposed images taken 0.470 K, and 0.609 K). By normalizing the lengths
at di€erent times during one of the experimental by the primary dendrite tip radius, R, all data
runs. The orientation of the dendrite is determined obtained at various times during growth and the
and the solid/liquid interface is tracked by marking four supercooling levels collapse along a single line,
it with a sucient number of points. Then, the indicating self-similar growth behavior for the
coordinates of the points are transformed by ro- envelope shape. Ignoring the data in the primary
tation and translation, such that the side-branch tip region (Z/R < 10), the branch tip positions and,
plane of interest is placed in the X±Z plane of refer- hence, the envelope shape are correlated by
ence and the symmetry axis of the dendrite arm is  0:859
Xtip Z
oriented along the Z-axis. An example is shown in ˆ 0:668 11†
R R
Fig. 6. The envelope shape is obtained by measur-
ing the distance, Xtip, of the active secondary side- This experimental correlation is valid in the self-
branch tips from the primary axis as a function of similar regime given by 1<<Z/R<<1/Pe, where Pe is
the distance, Z, from the primary tip. Again, a the tip Peclet number as de®ned above [14].

Fig. 7. Experimental scaling relation between the normalized envelope width X/R and the normalized
distance away from the primary tip, Z/R; the symbols represent all data obtained at di€erent times for
both sides of a dendrite at four di€erent supercooling levels (DT = 0.287, 0.370, 0.470 and 0.609 K);
only the squares were used in ®tting the scaling relation.
980 STEINBACH et al.: THREE-DIMENSIONAL MODELLING

The image analysis procedures used to measure

the solid volume of the dendrites from the IDGE
photographs are also explained in Ref. [14]. The
solid volumes are converted to mean solid fractions
for an entire equiaxed grain by multiplying the
volume of a primary branch by six and dividing it
by the volume of a certain grain envelope. In order
to avoid any ambiguities related to the envelope
shape, the mean solid fraction, fs , is de®ned here
for a spherical envelope having a radius equal to
the primary arm length Z. The resulting experimen-
tal correlation is given by [14]
 ÿ0:90
fs ˆ 3:86 Z 12†
R
which again is only valid in the self-similar regime Fig. 9. Comparison of predicted mean solid fractions in a
1<<Z/R<<1/Pe. For Z/R>1/Pe, the mean solid spherical envelope for di€erent supercoolings (symbols)
fraction approaches the dimensionless supercooling with the experimental scaling relation, equation (11).
DT/(L/c). Note that the axial coordinate Z/R can
also be interpreted as a dimensionless time t = t
v/R overprediction of the tip velocity for the higher
= Z/R, where t is the time. supercoolings in Fig. 3(a) and (b): the stagnant-®lm
The above two experimental correlations are now thickness (of 200 mm) is becoming too large relative
used to validate the mesoscopic model predictions. to the tip radius. Figure 8 also shows that the envel-
Figure 8 compares the envelope shapes for a pri- ope width for the lower supercoolings is underpre-
mary branch predicted for three supercoolings dicted in the dendrite root region, away from the
(symbols) to the experimental correlation given by tip. In the root region, the local supercooling in
equation (10) (line). The agreement is reasonably front of the secondary dendrite tips is very small
good, keeping in mind that the envelope shape in (below 0.1 K) resulting in the ®lm thickness becom-
the simulations is not prescribed but is a result of ing too small relative to the tip radii. In summary,
the coupling of the microscopic growth model to the above comparison indicates that it is generally
the mesoscopic temperature ®eld around the envel- important to ensure that everywhere on the envel-
ope. In the simulations, the dendrite arms were ope the ®lm thickness is of the same order of mag-
grown to the same length (in dimensional coordi- nitude as, but larger than, the tip radii. However,
nates) for all supercoolings. Then, the scaling of the reasonably accurate results can be obtained with
lengths with the tip radius in Fig. 8 results in the the same ®lm thickness for a range of growth con-
predicted envelopes appearing shorter with decreas- ditions.
ing supercooling. It can be seen from Fig. 8 that Figure 9 shows that reasonable agreement also
the mesoscopic model somewhat overpredicts the exists between the measured and predicted mean
envelope width for the highest supercooling. This solid fraction evolution, further validating the
disagreement is of the same origin as the slight mesoscopic model. The solid fraction predictions
for the lower two supercoolings are somewhat

Fig. 8. Comparison of the predicted envelope shape with Fig. 10. Predicted evolution of the temperature distri-
the experimental scaling relation, equation (10), for three bution between two dendrite branches approaching each
di€erent supercoolings. other (DTi=0.48S K, SCN).
 STEINBACH et al.: THREE-DIMENSIONAL MODELLING 981

Fig. 11. Predicted evolution of the tip speeds for a den-

drite branch approaching another grain; results are shown
for the aligned case corresponding to Fig. 10 for the ®rst
100 s (®lled symbols) and another case where the reference
grain is rotated by 458 (open symbols) (DTi=0.48 K,
SCN).

lower than the experimental correlation, which can

be explained by limitations associated with the nu-
merical grid used. In the calculations, the grain
evolves from a spherical seed that must have a
minimum radius of about two grid spacings (i.e.
200 mm). This seed size is too large compared to the
grain sizes expected in the self-similar regime.
Either a ®ner grid or an improved seeding pro-
cedure will be implemented in the near future.

6. RESULTS FOR MULTIPLE GRAINS

In this section, we present several examples of

simulations involving multiple equiaxed grains. As
a ®rst step, the growth of two interacting grains is
examined. The initial supercooling is 0.48 K and the
center-to-center distance between the grains is
10 mm. Figure 10 plots the temperature distribution
at di€erent times between the tips of two grains
with aligned growth directions (i.e. the tips grow Fig. 12. Predicted envelope shapes of four equiaxed den-
directly toward each other). Until 40 s, the tips dritic grains centered at the corners of a tetrahedron
(equal distances of about 4 mm) at 30 s and 100 s after
grow uncoupled. Thereafter, the temperature ®elds
seeding (DTi=0.48 K, SCN).
signi®cantly overlap. Figure 11 plots the evolution
of the tip speed for the aligned case as well as for a
case, where the second grain was rotated by 458 favorable conditions to grow. The inward pointing
with respect to the ®rst grain. Due to the inter- branches grow more slowly because of thermal in-
actions, the speed reduces to zero at about 100 s. In teractions between the grains.
the case where the second grain is rotated such that Finally the growth of 14 grains is simulated for
the reference tip points into the region between pri- two di€erent seeding con®gurations. The results for
mary branches, the ®nal growth stage is slightly the grain envelopes are shown in Fig. 13. The
prolonged. ordered con®guration shows an overall faster
Another simulation is performed with four growth than the random con®guration due to better
grains, located at the corners of a tetrahedron. The ®lling of the space in the ordered case. While the
nuclei are oriented randomly at a distance of about above results for multiple grains need to be ana-
4 mm from each other. Figure 12 shows dendrite lyzed in more detail and also require experimental
envelopes for the four grains at two di€erent times. validation, they illustrate the capabilities of the pre-
The outpointing branches are much more elongated sent model. The number of numerical grid points
than the branches pointing into the center of the used in these calculations is 120  120  120, and
box, because the outpointing branches have more the grid size is dx = dy = dz = 100 mm. The
982 STEINBACH et al.: THREE-DIMENSIONAL MODELLING

tion of the evolution of the grain envelope shape

and the internal solid fraction. The model is suc-
cessfully validated for the case of a single grain
through a comparison of the predictions with exper-
imental measurements from the IDGE.
Several simulation results for the growth of mul-
tiple grains are presented in order to illustrate the
model capabilities. They reveal the transient vari-
ations in the growth speeds and the development of
asymmetric grain shapes due to thermal interactions
between the grains. Future work will include a
more detailed analysis and experimental veri®cation
of these and other results. We envision that the
basic modeling concept presented here can be
extended to simulate columnar dendritic growth,
grain structure transitions, solutal transport, and
even growth in the presence of melt ¯ow.

AcknowledgementsÐThe authors I. Steinbach and B.

Kauerauf were supported by DARA under grant WS-
50WM9565. The authors C. Beckermann, Q. Li and J.
Guo were supported by NASA under grant NCC8-94.

REFERENCES
1. Flemings, M. C., Solidi®cation Processing. McGraw-
Hill, New York, 1974.
2. Beckermann, C. and Wang, C. Y., in Annual Review
of Heat Transfer, Vol. 6, ed. C. L. Tien. Begell House,
New York, 1995, pp. 115±198.
3. Caroli, B. and MuÈller-Krumbhaar, H., Iron Steel Inst.
Japan Int., 1995, 35, 1541.
4. Karma, A. and Rappel, W., Phys. Rev. Lett., 1996,
77, 4050.
5. Rappaz, M. and Thevoz, Ph., Acta metall., 1987, 35,
1487.
6. Gandin, Ch. A. and Rappaz, M., Acta metall., 1994,
42, 2233.
7. Steinbach, I., Kauerauf, B., Beckermann, C. and Guo,
Fig. 13. Examples of predicted envelope shapes of 14 J., in Solidi®cation 1998, ed. S. P. Marsh et al. TMS,
equiaxed dendritic grains; the upper panel corresponds to Warrendale, P A, 1998, pp. 5±14.
an ordered arrangement of seeds, while the lower panel 8. Hinze, J. O., Turbulence. McGraw-Hill, New York,
shows a random con®guration. 1975.
9. Cantor, B. and Vogel, A., J. Cryst. Growth, 1977, 14,
109.
calculations require of the order of 4.5 CPU hours
10. Diepers, H. J., Beckermann, C. and Steinbach, I., in
on a 100 MFlops workstation. Solidi®cation Processing 1997, ed. J. Beech and H.
Jones. SRP Ltd., Exeter, U.K., 1997, pp. 426±430.
7. CONCLUSIONS 11. Glicksman, M. E., Winsa, E. A., Hahn, R. C.,
LoGrasso, T. A., Tirmizi, S. H. and Selleck, M. E.,
The growth of multiple equiaxed dendrites of a Metall. Trans., 1988, 19A, 1945.
12. Glicksman, M. E., Koss, M. B., Bushnell, L. T.,
pure substance growing in a supercooled melt is LaCombe, J. C., Smith, R. N. and Winsa, E. A., in
modeled using a novel mesoscopic simulation tech- Heat Transfer in Microgravity Systems, Vol. 290, ed.
nique. The model couples the solution of the heat S. S. Sadhal and A. Gopinath. ASME, New York,
di€usion equation in the supercooled melt around 1994, p. 1.
13. LaCombe, J.C., Koss, M.B., Fradhov, V.E. and
the equiaxed grain envelopes with a local analytical Glicksman, M.E., Phys. Rev. E, 1995, 52, 2778.
solution (i.e. the stagnant-®lm model) for calculat- 14. Li, Q. and Beckermann, C., Phys. Rev. E, 1998, 57,
ing the dendrite tip speeds. It allows for the predic- 56.