Combustion and Flame 161 (2014) 405–415

Contents lists available at ScienceDirect

Combustion and Flame

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m b u s t fl a m e

Mechanism optimization based on reaction rate rules

Liming Cai ⇑, Heinz Pitsch
Institute for Combustion Technology, RWTH Aachen University, 52056 Aachen, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Accurate chemistry models form the backbone of detailed computational fluid dynamics (CFD) tools used
Received 18 April 2013 for simulating complex combustion devices. Combustion chemistry is often very complex and chemical
Received in revised form 27 August 2013 mechanisms generally involve more than one hundred species and one thousand reactions. In the deri-
Accepted 28 August 2013
vation of these large chemical mechanisms, typically a large number of reactions appears, for which rate
Available online 23 September 2013
data are not available from experiment or theory. Rate data for these reactions are then often assigned
using so-called reaction classes. This method categorizes all possible fuel-specific reactions as classes
Keywords:
of reactions with prescribed rules for the rate constants. This ensures consistency in the chemical mech-
Chemical mechanism
Optimization
anism. In rate parameter optimizations found in the published literature, rate constants of single elemen-
Reaction rate rules tary reactions are usually systematically optimized to achieve good agreement between model
Uncertainty minimization performance and experimental measurements. However, it is not kinetically reasonable to modify the
n-Pentane combustion rate parameters of single reactions, because this will violate consistency of rate parameters of kinetically
similar reactions. In this work, the rate rules, that determine the rates for reaction classes are calibrated
instead of the rates of single elementary reactions leading to a chemically more consistent model optimi-
zation. This is demonstrated by optimizing an n-pentane combustion mechanism. The rate rules are stud-
ied with respect to reaction classes, abstracting species, broken C–H bonds, and ring strain energy
barriers. Furthermore, the uncertainties of the rate rules and model predictions are minimized and the
pressure dependence of reaction classes dominating low temperature oxidation is optimized.
Ó 2013 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

1. Introduction constants of several important reactions could be determined from

quantum chemistry calculations [7–13] or experimental measure-
Chemical kinetic mechanisms play a significant role in CFD ments [14–16]. However, it is computationally and experimentally
modeling of combustion. For simple species (e.g. H2) or engine rel- difficult to determine the rate constants of all involved reactions
evant fuels (e.g. n-heptane), well-validated models can be found in due to the large number of intermediate species and reactions dur-
the literature [1–3]. However, limited kinetic information is found ing oxidation. One strategy to determine the missing reaction rate
for the combustion of emerging alternative fuels. In order to simu- coefficients is to use rate rules that are defined as rate constants for
late the combustion of such fuels, the development of accurate certain reaction classes [3,17,18].
kinetic models is of particular importance. Subsequently, additional experimental or computational data
A typical model development process is composed of several are used to tune these rate parameters [19]. Typically, sensitivity
steps. First, a kinetic mechanism including C0–C4 chemistry is cho- analyses of rate parameters are performed at the experimental
sen as base mechanism. The C0–C4 mechanism has a great influ- conditions of interest to choose the important rate constants which
ence on the combustion of hydrocarbon species and therefore are needed to be tuned. The deviations between model and exper-
many of the reaction rates specified for those reactions have been iment can be minimized by tuning these important reaction rate
intensively investigated from direct experimental measurements coefficients. In principle, that is done by perfecting the rate param-
or quantum chemical calculations [4–6]. Based on the C0–C4 mech- eters iteratively and comparing the results with experimental
anism, the fuel specific chemistry is then derived using prior observations [20]. An adapted set of rate constants is reached after
knowledge of similar molecules and their reaction classes, which several iterations of this procedure and good agreement between
describe the possible oxidation steps of the fuel. All possible reac- model results and experiments can be obtained.
tions occurring in each oxidation step are categorized as a class of Recently, several automatic frameworks were developed and
reactions with the assigned rate constant expressions. The rate implemented successfully to replace this process [21–24]. The idea
is to quantify the relation between model response [25] and uncer-
⇑ Corresponding author. tain parameters and thus apply a minimization algorithm to fit the
E-mail address: lcai@itv.rwth-aachen.de (L. Cai). parameters for a good agreement between model response and

0010-2180/$ - see front matter Ó 2013 The Combustion Institute. Published by Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.combustflame.2013.08.024
406 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415

experimental measurement, which is called solution mapping by numerical tool to fit the rates and optimize the model predictions.
Frenklach [22]. In addition, the so-called uncertainty quantification The modification of the set of the chemical reactions is typically
(UQ) process has been applied to chemical models allowing the not considered in the calibration. This issue will also be explored
possibility to estimate uncertainties of rate parameters and model in this work. The pressure dependence of low temperature chem-
results [26–28]. While some UQ approaches explicitly propagate istry will be studied in parallel to the calibration process in order
the experimental uncertainty into parameter uncertainties and to obtain a better model.
thus reproduce the experimental uncertainty in the modeling pro- The temperature dependence of the set of the chemical reac-
cess [21,29], other approaches such as the full Bayesian technique tions is assumed to be correct in this study. In general, if the tem-
[30,31], estimate the uncertainty based on probability theory. perature dependence in terms of the activation energy is
However, the above mentioned automatic frameworks focus on considered in the optimization process, it is expected that better
the calibration of individual reactions. As described, the reaction agreement can be obtained for a wide range of conditions as the
constants of most derived reactions in a kinetic model are obtained number of uncertain parameters is doubled. However, this increase
using rate rules and a rule can be used for few to dozens of reac- in the number could lead to a much higher computational expense,
tions in a mechanism. Therefore, in the common reaction class which makes it again difficult to calibrate the low temperature
based way of mechanism construction, rate constants of reaction ignition. For each experimental condition considered in the
classes and rate rules should be modified, not individual reactions calibration, (2n + 1)2 numerical simulations are required to gener-
belonging to a reaction class. Thus, kinetically similar reactions are ate the model response surface for n uncertain parameters [25].
treated equally, which ensures model consistency. Ideally, every More importantly, however, the constraint of both frequency fac-
reaction rate in a mechanism should be determined individually tors and activation energies of reactions is not trivial, because
and with high accuracy, which is unfortunately not the case for uncertainties of reactions found in the literature [7,8] are given
the most reaction mechanisms. If no information on a particular for rate constants k and not for pre-exponential factors A and
reaction is available from experimental measurements or ab initio activation energies E separately. A method for dealing with this
calculations, it is kinetically not reasonable to modify a rate of just has to be developed.
a single reaction of a reaction class, only because this reaction The presentation of the study is organized as follows. In the first
shows a higher sensitivity than other reactions of the same class. part, the reaction classes and the rate rules, as well as the optimi-
Instead, the rate for the associated class should be optimized. zation methodology, are introduced. The second part provides
In this study, the goal is to develop and demonstrate amore details of the development of an n-pentane mechanism and the
appropriate method for automatic calibration of chemical kinetic computational information. Following this, the calibration
models, which can be used when specific information of particular approach is first applied for high temperature combustion and
elementary reactions is lacking, by performing optimization of then extended to the complex low temperature chemistry. The
reaction rate rules. The methodology is demonstrated here by opti- mechanism is further improved through the incorporation of
mizing an n-pentane combustion model. A chemical model is gen- pressure dependent rules.
erated based on a well-validated base mechanism for C0–C4
chemistry. The fuel specific chemistry is derived using reaction
classes and rate rules. Accurate rate parameters determined indi- 2. Methodology
vidually from computational chemistry and experimental mea-
surements are incorporated in the reaction mechanism whenever 2.1. Reaction classes and rate rules
available. For a better agreement with the measured data, the rate
rules for reaction classes are then subjected to an automatic cali- Recently, a number of kinetic mechanisms were built in a
bration process. The base mechanism and all well-studied rate systematic way for components of interest for transportation fuel
parameters are exempted from calibration, as their rate uncertain- surrogates [3,17,18,34,19]. Curran et al. [3] proposed a kinetic
ties are relatively small. Only the reactions that have not been mechanism for n-heptane combustion using the methodology of
studied in detail, here especially the important reaction classes reaction class and rate rule. This methodology was further success-
involving fuel-peroxy and -peroxide species in the low tempera- fully applied to build kinetic models for the combustion of iso-oc-
ture regime, are considered in the calibration procedure. tane [17], n-alkane hydrocarbons [18], 2-methylalkanes [19] and
While ignition delays at high temperature are usually sensitive some oxygenated fuels [34,35]. In the currently common approach,
to only a few reactions, many reactions play a significant role for 30 reaction classes listed in Table 1 are used to build up a reaction
low temperature ignition. The rate coefficients of these reactions mechanism. The first ten reaction classes are used to describe high
will generally be considered as uncertain parameters in the model temperature oxidation of the fuel, while classes 11–30 are needed
calibration. This large number of uncertain parameters leads to an to describe the complicated low temperature combustion. As men-
unfeasible optimization process concerning computational cost. tioned earlier, the rate rules for each class are generated based on
This might be one reason why no studies on automatic calibration available chemical kinetic knowledge [3,19].
of low temperature chemistry are found in the published literature. As an example, the rate rules for class 15, the alkyl peroxy rad-
The calibration of rate rules provides the possibility to reduce the ical isomerization (RO2 ? QOOH), from Sarathy et al. [19] are
number of uncertain parameters by using rate rules that categorize shown in Table 2. The alkyl peroxy radical isomerization plays a
the chemically similar and sensitive reactions into a class. In this significant role in the low temperature regime. With the subse-
study, the ignition delay times at low temperatures are incorpo- quent O2 addition, isomerization, and formation of the carbon-
rated in the calibration process. Thereby, previous calibration ylhydroperoxide, two OH radicals are produced, which lead to
approaches [53,29,22,32] are extended to the engine relevant low low temperature chain branching. If inaccurate rates are employed,
temperature chemistry. a large modeling error can be introduced. The reaction rules for
Note that for most of the published calibration methodologies alkyl peroxy radical isomerization, as shown in Table 2, depend
the assumption is made that the chemistry model is perfect on the ring strain energy barrier and on the nature of the broken
[22,32,33,29]. In other words, if one could obtain data with zero er- C–H bond. In this study, one set of Arrhenius coefficients is treated
rors, the assumption states that the parameters could be fitted per- as one rule and thus one potential calibration object of the model
fectly and the model will be able to predict any other experiment optimization. Therefore, twelve rate rules are prescribed for the
exactly. Thus, the calibration framework is employed as a pure alkyl peroxy radical isomerization class. If a rate rule is calibrated,
 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415 407

Table 1 2.2. MUM–PCE

Reaction classes.

Number Class The method of uncertainty minimization using polynomial

1 Fuel decomposition chaos expansion MUM–PCE [21] will be used to calibrate the mod-
2 H-atom abstraction from the fuel el and provide uncertainty estimates for the kinetic parameters. In
3 Alkyl radical (R) decomposition general, experimental observations are used to validate model per-
4 Alkyl radical (R) isomerization formance. Uncertainties of rate constants incorporated in a mech-
5 H-atom abstraction from alkenes
6 Addition of O and OH to alkenes
anism and their propagations into the uncertainties of model
7 Addition of HO2, CH3O2, and C2H5O2 to alkenyl radicals predictions are typically not considered in the development pro-
8 Alkene decomposition cess. The MUM–PCE approach by Sheen and Wang [21] used in this
9 Alkenyl radical decomposition study provides the possibility to utilize the experimental results
10 Retroene decomposition
and their uncertainties for the model calibration and uncertainty
11 Addition of O2 to alkyl radicals (R)
12 R + RO2 ? RO + RO quantification. The solution mapping method developed originally
13 R + HO2 ? RO + OH by Frenklach [22] is employed in the MUM–PCE approach to opti-
14 R + CH3O2 ? RO + CH3O mize the rate parameters and hence to improve model predictions.
15 Alkyl peroxy radical isomerization By inversely propagating the experimental deviations into the
16 Concerted eliminations of RO2
uncertainties of the rate coefficients, the model uncertainties can
17 RO2 + HO2 ? ROOH + OH
18 RO2 + H2O2 ? ROOH + HO2 be minimized.
19 RO2 + CH3O2 ? RO + CH3O + O2 For this study, the Arrhenius pre-exponential factors are chosen
20 RO2 + RO2 ? RO + RO + O2 as the kinetic rate parameters that will be calibrated. These are
21 ROOH ? RO + OH
normalized as
22 RO decomposition
23 QOOH ? cyclic ether + OH
 
Ak
24 QOOH ? alkene + HO2 xk ¼ ln ln f k ; ð1Þ
25 QOOH ? alkene + carbonyl + OH
Ak;0
26 Addition of O2 to QOOH
27 Isomerization of O2QOOH
where Ak is the Arrhenius pre-factor of the kth reaction and Ak,0 is its
28 Decomposition of carbonylhydroperoxide nominal value in the mechanism. The normalized parameters are
29 Cyclic ether reactions with OH and HO2 treated as normally distributed and bounded by two standard devi-
30 Decomposition of carbonyl species and carbonyl radicals ations 2r to the interval {1, 1}. fk is the uncertainty factor of the
rate coefficient, and thus the uncertainty limits of the rate parame-
ters are defined as
Table 2
Rate constants for class 15: The alkyl peroxy radical isomerization (RO2 = QOOH), per
Ak =fk 6 Ak 6 Ak fk : ð2Þ
H-atom basis.
In the solution mapping approach [22,23], the normalized pre-
Ring size Site (°) A (1/s) n E (kJ/mol) factors are fitted by minimization of the least-squares difference
5 1 1.00E+11 0.00 121.336 between the model predictions gr(x) and the experimental mea-
2 1.00E+11 0.00 110.667 surements gobs
r . This is expressed as the minimization of the follow-
3 1.00E+11 0.00 99.161 ing objective function
6 1 1.25E+10 0.00 100.416 !
n 
X 2 X
M
2 1.25E+10 0.00 85.563 gr ðxÞ  gobs
r 2
U ðx0 Þ ¼ min þ 4xk ; ð3Þ
3 1.25E+10 0.00 78.241 x
r¼1
robs
r k¼1
7 1 1.56E+09 0.00 91.839
2 1.56E+09 0.00 78.032 where x0 is the set of the optimized normalized pre-factors, n is the
3 1.56E+09 0.00 69.664 number of experimental targets, and M is the number of reactions
8 1 1.95E+08 0.00 105.228 considered in the optimization of n experiments. For the rth exper-
2 1.95E+08 0.00 90.584 iment, the nominal value of the target gobs and its measured uncer-
r
3 1.95E+08 0.00 82.216
tainty robs
r have to be provided. In the optimization function, the
model prediction gr(x) is calculated through the so-called response
it means that the rate constants of all reactions using this rule are surface technique, which relates the kinetic rate parameters to the
optimized. prediction targets, such as ignition delays and laminar flame
The uncertainties of the rate rules are assumed to be rather speeds, in form of a second order polynomial. For each target, a
large, since there is insufficient information from literature. These sensitivity analysis of reactions at the corresponding condition is
large uncertainties of the rules are also not unexpected, as two performed to determine the important reactions which should be
sources of uncertainty are to be considered. First, uncertainties considered in the response surface. The model optimization is per-
can come from the estimation of the rate constants of individual formed here by minimizing Uðx0 Þ in Eq. (3). The second term in Eq.
reactions. The variability of experimental measurements and the (3) is a constraint on changes in the rate parameters x. As the rate
precision of the numerical calculations result in uncertainties in parameters are normally distributed in this study, a minimal
these rate constants. Moreover, uncertainties can also be intro- change of the rate parameters x is expected during the optimiza-
duced by taking such estimated rates as rate rule for similar reac- tion. The second term provides a penalty in the objective function
tions. Therefore, the rate rules can be modified within a large range for any change that is made to the rate constants and thereby en-
in the ‘‘tuning’’ process presented in the literature [3,17,18,34,19] sures that the optimization is performed with the smallest possible
in order to obtain accurate model predictions. Both types of uncer- changes of the rate parameters. An alternative here would be to
tainty are considered as the uncertainty of rate rules in this study. treat each rate parameter as a single experiment with a standard
The chemical mechanism for n-pentane used in this study is deviation r of 1/2.
developed based on the reaction classes introduced by Curran For the subsequent uncertainty quantification, the calibration
et al. [3] and Sarathy et al. [19]. The mechanism is described in rate parameters x are expressed using a first order polynomial
Section 3.1 and attached in Supplementary material. chaos expansion
408 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415

X
M et al. [47] was reduced to a skeletal level using a multi-stage
x ¼ x0 þ ai ni ; ð4Þ reduction strategy developed by Pepiot-Desjardins and Pitsch
i¼1
[50]. The reduction procedure involved the elimination of species
with mean x0 and covariance matrix R = aTa. Here, n is a vector of M and reactions. The lumping of chemical species presented in Ref.
normally distributed random variables with a mean value of 0 and a [50] was excluded, as the isomerizations of intermediate species
variance of 1 and ai are the rate parameter uncertainties. By treating are the important reaction steps and thus the rate coefficients play
fk as 2r uncertainties, the unoptimized uncertainties ai,0 are equal a significant role in the model prediction. The database used to
to 1/2. Inserting the polynomial chaos expansion of x into the perform the reduction includes the conditions at which experi-
response surface, a stochastic response surface of the model predic- mental measurements of ignition delay times at high temperatures
tion can be obtained. [40–42] are reported. This reduction process eliminated the species
The second order response surface model is linearized in the and reactions involved in the combustion chemistry of higher
neighborhood of the optimal value x0 . By inserting the linearized hydrocarbon species, e.g. n-heptane, n-octane, or which have neg-
response surface into the joint probability density function for x, ligible effect on reaction kinetics at the relevant temperature and
the following relation for the covariance matrix R can be obtained, pressure range. The reduced high-temperature mechanism con-
" #1 sists of 45 species and 228 reactions (forward and backward
XN
Wr ðx Þ counted separately). The simulated results using the mechanism
T  0
R¼a a ¼ þ 4I ; ð5Þ
ðr obs Þ2 from Wang et al. [47] and the reduced mechanism are compared
r¼1 r
and shown in the Supplementary material. Both mechanisms yield
with identical results. The reduction procedure is performed to reduce
T T T the computational cost of the optimization process. The accuracy
Wr ðx0 Þ ¼ ðbr þ br Þx0 xT T
0 ðbr þ br Þ þ ar x0 ðbr þ br Þ of model predictions is not compromised by this reduction and re-
T
þ ðbr þ br Þx0 aTr þ ar aTr ; ð6Þ mains unchanged.
In a similar way, the low temperature chemistry of n-pentane
and where I is the identity matrix. In contrast to the expression by was taken from the mechanism published by Curran et al. [17].
Sheen and Wang [21], the matrix br is not assumed to be symmetric, The mechanism from Curran et al. [17] is built up to describe
which leads to an improved estimate. Through the Cholesky factor- combustion of primary reference fuels (PRF) for gasoline, which
ization of R, the parameter uncertainties ai can be minimized. consists of iso-octane and n-heptane. Although this mechanism
Based on these optimized parameter uncertainties, the uncertainty has not been validated for n-pentane combustion, it includes the
of the model prediction can be calculated by taking the expectation low-temperature pathways of n-pentane built upon reaction clas-
of the stochastic response surface. ses. Thus, this mechanism is reduced to a skeletal level to describe
only the combustion of n-pentane in the low temperature regime.
3. Chemistry modeling The experimental data reported by Minetti et al. [36] and Ribacour
et al. [37] are considered as the reduction target here and the
3.1. Mechanism development resulting mechanism is composed of 77 chemical species among
340 reactions.
Premixed n-pentane combustion is used here to demonstrate The inclusion of the two reduced mechanisms of n-pentane, for
the methodology described above. Recently, a variety of kinetic high temperature and low temperature respectively, into the base
mechanisms including n-pentane oxidation chemistry have been mechanism was accomplished using an interactive tool [51] that
developed. Dealy et al. [43] proposed a chemical mechanism to automatically identifies common species and reactions from the
describe the oxidation of C1–C5 alkane and natural gas blends. different mechanisms. Rate conflicts detected during the merging
Detailed kinetic models for the autoignition of alkanes including were always resolved in favor of the validated base mechanism,
n-pentane were developed by Buda et al. [44] and Ranzi et al. therefore leaving that part of the mechanism virtually unchanged.
[45]. A comprehensive review of such detailed kinetic models Overall, 107 species and 1135 reactions are included in the com-
can be found in Ref. [46]. From the standpoint of experimental bined mechanism.The incremental part consists mainly of low-
investigation, ignition delay times of n-pentane have been temperature chemistry of n-pentane and some reactions for small
obtained in rapid compression machines at low to intermediate molecules not included in the original base mechanism. A compar-
temperatures [36–39], at high temperatures for n-pentane/O2/ ison of computed results using this mechanism and the mecha-
argon mixtures [40] and for n-pentane/air mixtures [41], as well nisms from Wang et al. [47] and Curran et al. [17] is shown in
as at high pressures for lean n-pentane/air mixtures in shock tubes the Supplementary material. It is seen that the combined mecha-
[42]. nism achieves better agreement with the experimental data at high
In this study, the chemistry model for n-pentane has been built temperatures than the mechanism of Wang et al. [47], which can
in a way similar to Curran et al. [3]. Instead of deriving the classes be attributed to the differences in the C0–C4 base mechanism
manually, they were adopted from available published mecha- [48]. Moreover, the combined mechanism and the mechanism
nisms [17,47]. First, a well-validated C0–C4 kinetic mechanism from Curran et al. [17] both fail to predict the ignition delay times
was chosen from Narayanaswamy et al. [48], which was developed in the low temperature regime accurately.
based on the detailed mechanism by Blanquart et al. [49]. This The two original mechanisms [17,47] were established using
mechanism is well-characterized and describes the oxidation of the concept of reaction classes and rate rules. Hence, the combined
fuels spanning from C1 to C8 hydrocarbon species. The reactions mechanism applied in this study is based on the same principles.
of various substituted aromatic compounds, and those of n-hep- The mechanism is attached in the Supplementary material. The re-
tane and iso-octane were excluded due to their marginal effect verse reaction rates needed in the mechanism are calculated from
on n-pentane oxidation. thermochemistry and equilibrium constants during the simulation.
The high-temperature kinetic mechanism of n-pentane was Note that the mechanism from Curran et al. [17] provides the
taken from Wang et al. [47]. The kinetic mechanism from Wang high-temperature reactions for n-pentane oxidation as well. How-
et al. [47] shows good model performance at high temperatures ever, Wang et al. [47] studied and updated kinetic parameters for
for different experimental setups. Due to the large model size the high-temperature reactions intensively including the fuel
and number of fuels in the mechanism, the mechanism from Wang decomposition, H-atom abstraction from n-pentane and pentene,
 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415 409

as well as pentyl and pentenyl radical decomposition. Therefore, in prediction. A given rule, which has a large sensitivity, could lead
this study, the well-studied reactions from Wang et al. [47] were to a marginal model uncertainty, if the rate uncertainty is neglected.
incorporated in the mechanism and were exempted from calibra- Here, the uncertainty factors of the reaction rate rules were
tion. This is consistent with our calibration approach described assumed equal to four, as no literature values were found, and
earlier, which only considers the reactions that have not been stud- large uncertainties are expected as discussed in Section 2.1. As
ied in detail. mentioned earlier, the C0–C4 base mechanism and the reactions
with accurate rate constants were not considered in the calibration
3.2. Computational details process. These reaction rates were assumed to be exact and their
uncertainties were thus negligible in this study.
The chemistry model described above was calibrated in a fur- The response surface coefficients were determined by the so-
ther step. The flame speeds of hydrocarbon species depend called sensitivity analysis based (SAB) method [25]. The numerical
strongly on the reactions involving small molecules as found in calculations were performed using the 0-dimensional configura-
numerous reports [3,17]. Since these reactions are well-validated tion of isochoric homogeneous reactor in the FlameMaster code
and included in the C0–C4 base mechanism, we focus on the [52]. An adiabatic premixed combustion is assumed to take place
ignition of n-pentane for the calibration process, for which the fuel and the species and energy equations are solved. Various defini-
oxidation pathways are dominant. An overview of the experimen- tions of the ignition delay time can be selected in FlameMaster
tal measurements of ignition delay times used here are shown in code [52], e.g. the onset of OH⁄ or CH⁄ emission and the time to
Table 3. These experimental data were obtained from measure- reach a given temperature increase. In this study, the ignition delay
ments in rapid compression machines [36,37] and shock tubes time is defined as the time to reach the maximum temperature
[40,41] in a pressure range from 2 bar to 12 bar. An overall uncer- change. The total computation time was roughly two days on an
tainty of 10% in the ignition delay time measurements has been DELL T1600 workstation.
reported by Davidson et al. [40]. However, values for uncertainty
estimates are often lacking in such experimental studies
[36,37,41]. Therefore, the standard deviation of these experimental 4. Results and discussion
measurements was assumed equal to the uncertainty reported in
Ref. [40]. In this section, the calibration of reaction rate rules is first
For the present study, the rate rules that showed the highest applied to n-pentane ignition measured in the shock tube [40] to
optimization potential in the range of experimental conditions demonstrate the methodology. The numerical results obtained
were chosen as the uncertain parameters. The optimization poten- using the calibrated model based on rate rules are compared with
tial of rate rule is introduced here as the model performance from the calibration process, in which the
elementary reactions are directly optimized. Thereafter, a cali-
t ig t ig;0
t
brated model is developed with respect to experimental data span-
Optimization potential ¼ fi  Sensitivity ¼ fi Pn ig;0A A ning from the low to the high temperature regime. The pressure
1 i i;0
n i¼1 Ai;0 dependence of the low temperature chemistry is further optimized
Ai;0 tig  tig;0 for better model agreement with the experimental data.
¼ fi ; ð7Þ
Ai  Ai;0 t ig;0

where fi is the uncertainty factor of rate rule i, tig,0 is the ignition de- 4.1. Mechanism calibration based on rate rules
lay time calculated using the unoptimized mechanism, Ai,0 is the
nominal value of the pre-exponential factor of the rate coefficient To illustrate the calibration of reaction rate rules, the experi-
in the mechanism, and n is the number of the reactions using this mental data from Davidson et al. [40] at high temperatures are
rule. tig indicates the ignition delay time calculated using Ai, which used in a first step. Davidson et al. [40] measured ignition delay
is obtained by multiplying Ai,0 with a factor of two. Alternatively, a times behind reflected shock waves over a range of initial temper-
factor of fi can be employed to numerically obtain the sensitivity, in atures of 1261–1490 K at p = 1.9 atm and / = 1.0 in oxygen/argon
which case the optimization potential is expressed as mixtures. The sensitivity analysis of rate rules at 1360 K is shown
in Fig. 1 in comparison with the analysis of elementary reactions
fi tig  t ig;0 in Fig. 2. The analysis reveals that the most sensitive rate rule is
Optimization potential ¼ : ð8Þ
fi  1 t ig;0 the one specified for the H-atom abstraction from secondary (2°)

For each experimental condition considered in the calibration pro-

cess, a sensitivity analysis of rate rules was performed. A positive Fuel + H → R(2°) + H2
sensitivity characterizes rules which enhance the reactivity of the Fuel decomposition
fuel, while rules with a negative sensitivity inhibit the reactivity. Fuel + H → R(1°) + H2

Here, the optimization potential is obtained by multiplying the sen- Fuel + OH → R(2°) + H2O

sitivity by the uncertainty factor of the reaction rate constant. The R(2°) → Alkene + H

sensitivity, combined with uncertainty factors, indicates the contri- Alkene + O → Alkenyl + OH

Alkene + H → Alkenyl + H2
bution of the uncertainty of a rule to the uncertainty of the model
R(1°) → Alkene + H

Fuel + CH3 → R(2°) + CH4

Table 3 Fuel + OH → R(1°) + H2O

Measurements of ignition delay times. Fuel + O → R(2°) + OH

Alkene decomposition
Pressure / Temperature (K) Number Source
Fuel + O → R(1°) + OH
5 bar 1.0 678–913 14 Minetti et al. [36]
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15
8 bar 1.0 648–875 15 Ribaucour et al. [37]
12 bar 1.0 647–879 12 Ribaucour et al. [37] Sensitivity [-]
1.9 atm 1.0 1261–1490 9 Davidson et al. [40]
10 atm 1.0 1070–1383 9 Marks et al. [41] Fig. 1. Sensitivity analysis of reaction rate rules at p = 1.9 atm, T = 1360 K
and / = 1.0.
410 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415

N-C5H12 → P-C4H9 + CH3 Based on the results from the sensitivity analyses, the estab-
N-C5H12 + H → S2X-C5H11 + H2 lished n-pentane mechanism is subjected to the calibration using
N-C5H12 + H → SX-C5H11 + H2
the measured ignition delays from Davidson et al. [40]. Two
N-C5H12 → N-C3H7 + C2H5
calibrations are performed. An improved model (ModelRR) is
N-C5H12 + H → PX-C5H11 + H2

S2X-C5H11 → C5H10-2 + H
obtained through calibration of rate rules. In comparison, a model
N-C5H12 + OH → S2X-C5H11 + H2O (ModelER) is developed using the standard approach of treating the
C5H10-2 + O → C5H9 + OH elementary reactions as uncertain parameters. Ten rate rules and
N-C5H12 + OH → SX-C5H11 + H2O 13 reactions that show high sensitivity (>1%) in the range of
C5H10-2 + H → C5H9 + H2
experimental conditions were chosen for the corresponding opti-
PX-C5H11 → H + C5H10
mization processes.
N-C5H12 + OH → PX-C5H11 + H2O
N-C5H12 + CH3 → H + C5H10
The results obtained with the optimized mechanisms are shown
N-C5H12 + O → S2X-C5H11 + OH and compared with experiments in Figs. 3 and 4. It can be seen that
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 the model predictions were significantly improved with both mod-
Sensitivity [-] els and the ignition delay times computed using the optimized
mechanisms agree well with the measurements. In order to dem-
Fig. 2. Sensitivity analysis of elementary reactions at p = 1.9 atm, T = 1360 K and / onstrate the efficiency of model calibration, the L2error norms of
= 1.0.
the model predictions are employed
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2
2 1 Xn gr ðxÞ  gobs r
carbon sites of fuel by H radicals. In the n-pentane mechanism, this L ¼ : ð9Þ
rule is employed for the two reactions n r¼1 robs
r

The L2error norms of the models are calculated based on logarith-

Class 2 : N—C5 H12 þ H ! SX—C5 H11 þ H2 ; mic ignition delays and are shown in Table 4. In comparison with
Class 2 : N—C5 H12 þ H ! S2X—C5 H11 þ H2 : the unoptimized model, the L2error norms of optimized models
are strongly decreased. ModelER is found to have marginal predic-
For n-pentane, two pentyl isomers can result by H-atom abstraction tion advantages. However, ModelRR shows better agreement at rel-
from secondary carbon sites. Here, SX and S2X refer to the first and atively high temperatures. The slightly disadvantageous L2error
second secondary site respectively. The rate leading to SX–C5H11 is norm of ModelRR is attributed to the large absolute deviations at
twice that of the one leading to S2X–C5H11 because of the higher higher ignition delay times, at around 1250 K, as shown in Fig. 3.
degeneracy of the first secondary site in n-pentane. As shown in The rate parameter uncertainties were improved after the cali-
Fig. 2, the two reactions are the second and third most sensitive bration, which leads to improved prediction uncertainties as
reactions with a sensitivity of 9.0% and 7.1%, respectively. This is shown in Fig. 4. The prediction uncertainties of ModelER are not
consistent with the large sensitivity (14.4%) of the rate rule found shown here, as the differences between ModelER and ModelER are
in Fig. 1. The fuel decomposition class shows the largest negative marginal. The standard deviations of the models are of similar
sensitivity in Fig. 1, while also the two individual decomposition magnitude as the experimental errors. It should be mentioned that
reactions are found to inhibit the fuel reactivity in Fig. 2. This the standard deviation of the predictions shown in Fig. 4 is a lower
shows, as expected, that the sensitivity analyses between rate rules bound of the actual uncertainty because of the exclusion of the cal-
and elementary reactions are consistent. However, the different ibration of the C0  C4 sub-mechanism and well-studied reactions
evaluation for both approaches points to one potential problem, from Ref. [47].
when considering individual reactions as compared to reaction clas- The similar performance accuracy for both optimized models
ses. If there are many reactions of one class, then each one could has demonstrates the possibility to improve model predictions and
a relatively low sensitivity compared to some other reactions, the prediction precisions for a given set of experimental data by cali-
effect of all reactions could still be large, such that this as a reaction brating reaction rate rules. It should be noted that the calibration
class could have the highest sensitivity. of rate rules is incapable of achieving further improvements of pre-
Note that the class of fuel decomposition was not considered in diction precision over the traditional calibration of the elementary
the calibration, although it plays a major role for the ignition delay reactions, except, if the sensitivity analysis finds reaction classes, in
time. This is due to the constraint of the calibration approach
where well-studied reaction rates [47] are not to be changed in
the optimization process. In this study, two reactions are included 1.9 atm, φ = 1.0, 4% O2/Ar
in this class.

1
Class 1 : C4 H9 þ C2 H5 ! N—C5 H12 ;
τig [ms]

Class 1 : C3 H7 þ C2 H5 ! N—C5 H12 :

The rate constants of both reactions [47] are different from
those suggested by Curran et al. [3] and Sarathy et al. [19] and
Davidson
are specified based on similar well-studied reactions for n-propane ModelER
and n-butane, for which very small uncertainty factors of 1.5 and 0.1 ModelRR
1.2, respectively, have been reported [54,55]. Although it could 0.68 0.72 0.76 0.8
results in inevitable uncertainties by taking these thoroughly-stud- -1
1000/T [K ]
ied rates as approximate rate rules for n-pentane, the introduced
uncertainties are expected to be small, as the decomposition Fig. 3. Predictions of ignition delay times of n-pentane–oxygen–argon mixtures.
products from n-pentane are kinetically very similar to these from Symbols denote experimental data [40]. Dashed line shows results of the
unoptimized model. Black line shows the results of the calibrated model ModelER
n-propane and n-butane. In comparison to a factor of four used for and red line shows predictions using calibrated model ModelRR. (For interpretation
the uncertainty of rate rules in this study, these rates are relatively of the references to color in this figure legend, the reader is referred to the web
well-studied and therefore exempted from modification. version of this article.)
 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415 411

1.9 atm, φ = 1.0, 4% O2/Ar 1.9 atm, φ = 1.0, 4% O2/Ar

1 1
τig [ms]

τig [ms]
0.1 Davidson 0.1 Davidson
0.68 0.72 0.76 0.8 0.68 0.72 0.76 0.8
-1 -1
1000/T [K ] 1000/T [K ]

Fig. 4. Predictions of ignition delay times of n-pentane–oxygen–argon mixtures. Symbols denote experimental data [40]. Solid lines show the predictions using the models
and dashed lines present the standard deviation of the predictions. Top: unoptimized model. Bottem: optimized model ModelRR.

Table 4 of initial temperature of the gas mixture in Fig. 5. It is clearly seen

L2error norm of models. that the model fails to reproduce the experimental data. The sim-
Model L2 error norm ulated ignition delay times in the high temperature regime show
Unoptimized model 13.9732
better agreement with the experiments in comparison with the
ModelRR 2.7719 results in the low temperature regime. The observed differences
ModelER 2.4884 indicate the necessity to investigate or calibrate the rate constants
of reactions for low temperature combustion.
It should be mentioned that the deviations between experimen-
which each elementary reaction has a low sensitivity but the effect tal measurements from different experimental configurations
of all reactions cannot be neglected. In this sense, the number of could be attributed to the differences in the individual setups. Fur-
rules included in the calibration increases and model predictions thermore, accurate estimates of experimental uncertainties are
can be further improved. However,this was not the case in the important here, because they allow for a more rigorous compari-
present example. son of different experimental observations. However, here is as-
The present approach leads to a kinetically reasonable and sumed that the experimental setups are perfect and the obtained
appropriate model optimization. The number of the uncertain measurements are reliable within the assumed experimental
parameters is reduced, because kinetically similar reactions are uncertainty.
categorized into a class of reactions with an assigned rate rule. The computed ignition delay times using the calibrated model
Therefore, the computational cost is minimized by following such ModelHL are presented in Fig. 5. In the optimization, 27 rules were
a categorization. This provides the possibility to calibrate also the calibrated and detailed information about this can be found in the
oxidation at low temperatures, where many reactions are of partic- Supplementary material. As expected, the optimization process
ular importance. improves the computed ignition delays at the considered experi-
mental conditions significantly. In particular, improved agreement
between the simulations and the measurements at the low and
4.2. Model calibration including low temperature oxidations intermediate temperatures can be observed. For gas mixtures with
high initial temperature, ModelHLachieves very good agreement.
In this section, the calibration of the n-pentane model is de- Interestingly, the differences between the computed ignition
scribed that was carried out using all of the experiments listed in delays at different pressures in the low temperature range
Table 2, such that the model is applicable also to the oxidation at vanishes with decreasing temperature as shown in Fig. 5. Similar
low temperatures. The rate rules for the low temperature classes model performance in the low temperature range can be observed
are now included in the optimization. The ignition delay times in the literature [18,3,56] for different hydrocarbon fuels. However,
simulated using the unoptimized model are shown as a function the measurements from the rapid compression machines show a

100
1
τig [ms]

τig [ms]

10
0.1
Minetti 5 bar
Ribaucour 8 bar Davidson 1.9 atm
Ribaucour 12 bar Marks 10 atm
1 1.1 1.2 1.3 1.4 1.5 1.6 0.64 0.74 0.84 0.94
-1 -1
1000/T [K ] 1000/T [K ]

Fig. 5. Predictions of ignition delay times of n-pentane using different models. Solid lines show the predictions using the calibrated model ModelHL and dashed lines present
the results of the unoptimized model. Symbols denote experiments: Stoichiometric n-pentane/air mixtures by Minetti et al. [36], Ribaucour et al. [37] and Marks et al. [41];
oxygen/argon mixture at / = 1.0 in 4% by Davidson et al. [40].
412 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415

pressure dependence [57,37]. As shown in Fig. 5, the experimental The rate constants at other pressures within the specified pres-
measurements at 8 bar and 12 bar have discrepancies of about sure range can then be determined through logarithmic
5–10 ms for the temperature range from 600 K to 700 K, where interpolation
the differences between simulation results are negligible. The
two measurements reported by Minetti et al. [36] at around ln p  ln pi
ln k ¼ ln ki þ ðln kiþ1  ln ki Þ : ð11Þ
676 K indicate a strong increase of ignition delay times in the ln piþ1  ln pi
low temperature regime. This is not unexpected in rapid compres-
sion machines. As the induction period increases at lower temper- In this study, the discrete rate coefficients are specified for the pres-
ature, the effect of heat loss in rapid compression machines could sures of interest, which are 5, 8, and 12 bar. An example is given in
grow and lead to longer ignition delays [58,59]. Table 6.
As discussed in Refs. [3,17,19], the low temperature chemistry The three sets of rate parameters are calibrated independently
depends on several combination and decomposition steps, which within the uncertainty limit for the experimental pressures of
are pressure-dependent. For instance, the dissociation of perox- interest. In a consistent calibration process, the rate rules in the
yalkylhydroperoxide and ketohydroperoxide yields a significant ef- PLOG function for a certain pressure is only considered at the
fect on the ignition of the fuels. In this sense, the model predictions experimental condition with the same pressure, while the rules
can be further improved if the option of treating pressure-depen- for non-pressure-dependent reactions are considered over the
dent rate coefficients in the calibration process is possible. This is entire pressure range of interest. For example, the rules specified
demonstrated in the next section. for class 28 at 5 bar are calibrated using the corresponding
measurements from Minetti et al. [36] and rules at 8 bar using
the measurements from Ribaucour et al. [37] at 8 bar. In contrast,
4.3. Pressure-dependent low temperature chemistry
both measurements are considered for the calibration of class 15
for RO2 isomerization.
As mentioned earlier, the pressure dependence of the low
The model is then calibrated for all measurements using pres-
temperature pathways plays an important role for the accurate
sure-dependent rate rules. The calibrated rate rule for class 28 is
computation of ignition delay times. Concerning the results from
shown as an example in Table 6. Different pre-exponential factors
the sensitivity analysis, the reaction steps with large sensitivities
are obtained after the calibration process for discrete pressures.
(listed in Table 5) are further studied incorporating pressure-
As shown in Fig. 6, the computed ignition delay times using the
dependent rate constants. Class 20 describes the recombination
calibrated model (ModelPLOG) appear satisfactory for the experi-
of RO2 radicals, which is competing with the isomerization of
mental conditions. For high initial temperatures, excellent agree-
RO2 to QOOH radicals and thus slows down the reactivity of n-pen-
ment between ModelPLOG and the experiments is obtained. The
tane in the low temperature regime. The same inhibiting effect can
integration of the pressure dependence into the low temperature
be observed from the cyclization of hydroperoxyalkylradical QOOH
chemistry has no effect on high temperature oxidation. As
to form a cyclic ether and an OH radical with a ring size of the cyc-
expected, the computed ignition delay times in the intermediate
lic ether equal to 4. This reaction class is only chain propagation
and low temperature regime are improved, especially at low
and therefore attenuates reaction progress. Classes 27 and 28 are
temperatures. The prediction uncertainties of the unoptimized
two important steps to release OH radicals leading to the low tem-
and calibrated model are presented in Figs. 7 and 8. It can immedi-
perature chain branching process. The rate rules of the four chosen
ately be seen that the uncertainties of ModelPLOG are efficiently
steps and the number of the elementary reactions using these rules
minimized through the calibration and agree very well with the
are provided in Table 5.
experimental observations.
Due to the lack of accurate measurements and calculations, only
As for the calibration described in Section 4.2, 27 rate rules were
one set of reaction parameters is used for the reactions derived for
considered in this optimization process, as no additional experi-
these four classes. Even though especially class 28 is generally con-
mental conditions were included. Detailed information about the
sidered the rate-determining step for ignition at very low temper-
optimized rate parameters and their uncertainties is shown in
atures, limited information for all rates shown in Table 5 can be
Fig. 9. It is clearly seen that the rate parameters for the isomeriza-
found in the recent literature and effect of isomerization ring size
tion of alkyl peroxy radicals (class 15) and peroxy-alkyl-hydroper-
and broken C–H bond are typically not quantified. However, due
oxide radicals (class 27), typical for low temperature chain
to their importance, these rate parameters are frequently modified
branching pathway, are increased to accelerate the ignition pro-
to obtain good agreement with experiments [19].
cess. While the unoptimized model shows a large deviation from
Here, the dependence of a rate coefficient on pressure is pre-
the experiments in the intermediate temperature range, it can pre-
sented based on direct interpolation of rate coefficients specified
dict ignition delay times at very low temperatures with moderate
at individual pressures, which is the so-called PLOG function devel-
accuracy. Therefore, the modification of rates for classes 15 and 27
oped by Miller and Lutz [60]. For the pressure pi, the pre-exponen-
could lead to the underprediction of ignition delays at low temper-
tial factor Ai, the temperature exponent ni and the activation
atures. In order to compensate this degradation of the model
energy Ei are specified.
  performance, rate coefficients used for the ketohydroperoxide
Ei decomposition (class 28) were decreased to retard the fuel ignition
ki ðT; pi Þ ¼ Ai T ni exp  i ¼ 1; 2; 3 . . . ð10Þ
RT

Table 6
Calibrated rate rule for the pentylhydroperoxide decomposition, class 28.
Table 5 p (bar) A (1/s) n (–) E (kJ/mol)
Pressure-dependent reaction classes considered.
C28: Ketohydroperoxide ? Products + OH
No. of class Reaction class Number of reactions 5.000 3.750E+15 0.000 175.728
8.000 7.593E+15 0.000 175.728
20 RO2 + RO2 ? RO + RO + O2 6
12.00 1.318E+16 0.000 175.728
23 QOOH ? Cyclic ether + OH (RS 4) 3
27 O2QOOH ? Ketohydroperoxide + OH 3 Originally assigned rate rule
28 Ketohydroperoxide ? products + OH 3 1.500E+16 0.000 175.728
 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415 413

100
1

τig [ms]

τig [ms]
10
0.1
Minetti 5 bar
Ribaucour 8 bar Davidson 1.9 atm
Ribaucour 12 bar Marks 10 atm
1 1.1 1.2 1.3 1.4 1.5 1.6 0.64 0.74 0.84 0.94
-1 -1
1000/T [K ] 1000/T [K ]

Fig. 6. Predictions of ignition delay times of n-pentane using different models. Solid lines show the predictions using the calibrated model ModelPLOGand dashed lines present
the results of the unoptimized model. Symbols denote experiments: Stoichiometric n-pentane/air mixtures by Minetti et al. [36], Ribaucour et al. [37] and Marks et al. [41];
oxygen/argon mixture at / = 1.0 in 4% by Davidson et al. [40].

100 100
τig [ms]

τig [ms]

10 10

Minetti 5 bar
Ribaucour 12 bar Ribaucour 8 bar
1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6

1000/T [K-1] 1000/T [K-1]

1
τig [ms]

0.1

Davidson 1.9 atm

Marks 10 atm
0.64 0.74 0.84 0.94

1000/T [K-1]

Fig. 7. Prediction uncertainty of ignition delay times of n-pentane using the unoptimized model. The shaded bands show the prediction uncertainties. Symbols denote
experiments: Stoichiometric n-pentane/air mixtures by Minetti et al. [36], Ribaucour et al. [37] and Marks et al. [41]; oxygen/argon mixture at / = 1.0 in 4% by Davidson et al.
[40].

100
1
τig [ms]

τig [ms]

10
0.1
Minetti 5 bar
Ribaucour 8 bar Davidson 1.9 atm
Ribaucour 12 bar Marks 10 atm
1 1.1 1.2 1.3 1.4 1.5 1.6 0.64 0.74 0.84 0.94

1000/T [K-1] 1000/T [K-1]

Fig. 8. Prediction uncertainty of ignition delay times of n-pentane using the calibrated model ModelPLOG. The shaded bands show the prediction uncertainties. Symbols
denote experiments: Stoichiometric n-pentane/air mixtures by Minetti et al. [36], Ribaucour et al. [37] and Marks et al. [41]; oxygen/argon mixture at / = 1.0 in 4% by
Davidson et al. [40].
414 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415

Fig. 9. Calibrated rate rules in ModelPLOG. x⁄ and a⁄ denote the calibrated normalized rate parameters and their uncertainties respectively.

at very low temperatures, since these reactions are only sensitive measurement of Ribacour et al. [37] (at 650 K and 8 bar) appear
in the low temperature regime and their influence attenuates with as outliers in comparison with the calibrated model. If the heat loss
the increase in the initial temperature. effect at low temperatures is reproduced during the simulation,
As mentioned by Curran et al. [17], the pre-exponential factors better agreement is expected.
used for class 15 of the heptyl peroxy radical isomerization are
approximately an order magnitude lower than those recom- 5. Conclusions
mended in the literature [61–63]. For iso-octane, these factors
were further reduced by another factor of 0.3. In the present study, In this work, a stochastic calibration methodology is success-
the factors specified for n-heptane by Curran et al. [17] were em- fully applied to calibrate the rate rules used for kinetic model
ployed in the unoptimized mechanism for n-pentane and in the development, while also estimating uncertainties in the rules and
calibration, these were increased by a factor of 4, which is closer models. The calibration applied only to rate rules provides equally
to the recommended rates in the literature [61–63]. good performance in comparison with the optimization of the sin-
As demonstrated, the PLOG functions allow a practical option to gle elementary reactions, yet it leads to a chemically consistent
integrate the pressure dependence for low temperature combus- mechanism. Further, the methodology reduces the number of
tion. In addition, this procedure provides computational advanta- parameters that need to be considered and therefore enables also
ges. The computational effort of the calibration process depends optimization of cases, where a large number of reactions appears
strongly on the generation of the response surface. For each exper- as important, as is the case for low temperature auto-ignition of
imental condition, the response surface is developed with respect larger aliphatic fuels. The calibration approach is first applied to
to the sensitive reactions at the condition. The number of the high temperature auto-ignition and then extended to the complex
sensitive reactions is therefore quite important for the size of the low temperature chemistry. Further refinement through the incor-
response surface and thus for the calibration effort. The implemen- poration of pressure dependent rules provides a significant
tation of PLOG functions increases the calibration efficiency by improvement of the prediction capability at low temperatures.
using accurate rates for discrete pressures. Nevertheless, it does The demonstrated calibration methodology enables a simpler
not lead to additional computation cost, since the number of reac- and more consistent calibration of rate parameters that are diffi-
tions considered for each condition remains unchanged. cult to be obtained from computational chemistry or experimental
Note that the optimized model underpredicts the ignition delay measurements, or which have simply not been studied with these
times at very low temperatures of 675 K at 5 bar and 650 K at 8 bar, approaches. This method is not only useful for providing an opti-
as shown in Fig. 6, while an excellent agreement between the sim- mized model estimation, but also allows for the possibility to
ulated results with calibrated model and the experiments is investigate which crucial rules should be studied in greater depth
achieved at higher temperatures. As discussed in Section 4.2, the in the future.
ignition delays at very low temperatures could be affected by the
heat loss effect in rapid compression machines and thus the mea- Acknowledgments
sured ignition delay times are potentially larger than their actual
values. This seems to be confirmed in the optimization procedure, This work was performed as part of the Cluster of Excellence
as two measurements of Minetti et al. [36] (at 675 K) and one ‘‘Tailor-Made Fuels from Biomass’’, which is funded by the
 L. Cai, H. Pitsch / Combustion and Flame 161 (2014) 405–415 415

Excellence Initiative of the German federal state governments to [29] D. Sheen, H. Wang, Combust. Flame 158 (2011) 645–656.
[30] Y.M. Marzouk, H.N. Najm, J. Comput. Phys. 228 (6) (2009) 1862–1902.
promote science and research at German universities, and as part
[31] H.N. Najm, B.J. Debusschere, Y.M. Marzouk, S. Widmer, O.P. Le Matre, Int. J.
of the Collaborative Research Center (SFB) 1029 which is funded Numer. Methods Fluids 80 (2009) 789–814.
by the German Research Foundation (DFG). [32] S.G. Davis, A.V. Joshi, H. Wang, F. Egolfopoulos, Proc. Combust. Inst. 30 (2005)
1283–1292.
[33] T. Russi, A. Packard, R. Feeley, M. Frenklach, J. Phys. Chem. A 112 (2008) 2579–
Appendix A. Supplementary material 2588.
[34] S.M. Sarathy, S. Vranckx, K. Yasunaga, M. Mehl, P. Owald, W.K. Metcalfe, C.K.
Westbrook, W.J. Pitz, K. Kohse-Hinghaus, R.X. Fernandes, H.J. Curran, Combust.
Supplementary data associated with this article can be found, in Flame 159 (2012) 2028–2055.
the online version, at http://dx.doi.org/10.1016/j.combustflame. [35] K. Yasunaga, J.M. Simmie, H.J. Curran, T. Koike, O. Takahashi, Y. Kuraguchi, Y.
2013.08.024. Hidaka, Combust. Flame 158 (2011) 1032–1036.
[36] R. Minetti, A. Roubaud, E. Therssen, M. Ribaucour, L.R. Sochet, Combust. Flame
118 (1999) 213–220.
References [37] M. Ribaucour, R. Minetti, L.R. Sochet, H.J. Curran, W.J. Pitz, C.K. Westbrook,
Proc. Combust. Inst. 28 (2000) 1671–1678.
[1] Z. Hong, D.F. Davidson, R.K. Hanson, Combust. Flame 158 (4) (2011) 633–644. [38] J.F. Griffiths, P.A. Halford-Maw, D.J. Rose, Combust. Flame 95 (1993) 291–306.
[2] M.P. Burke, M. Chaos, Y. Ju, F.L. Dryer, S.J. Klippenstein, Int. J. Chem. Kinet. 44 [39] J.F. Griffiths, P.A. Halford-Maw, C. Mohamed, Combust. Flame 111 (1997) 327–
(7) (2012) 444–474. 337.
[3] H. Curran, P. Gaffuri, W. Pitz, C. Westbrook, Combust. Flame 114 (1998) 149– [40] D.F. Davidson, S.C. Ranganath, K.Y. Lam, M. Liaw, Z. Hong, R.K. Hanson, J.
177. Propul. Power 26 (2010) 280–287.
[4] B. Eiteneer, M. Frenklach, Int. J. Chem. Kinet. 35 (2003) 391–414. [41] B. Marks, O. Mathieu, R. Archuleta, E.L. Petersen, G. Bourque, Spring Technical
[5] Y. Hidaka, Y. Henmi, T. Ohonishi, T. Okuno, Combust. Flame 130 (2002) 62–82. Meeting of the Central States Section of the Combustion Institute, 2012.
[6] A. Laskin, H. Wang, C. Law, Int. J. Chem. Kinet. 32 (2000) 589–614. [42] V.P. Zhukov, V.A. Sechenov, A. Yu Starikovskii, Combust. Flame 140 (2005)
[7] D.L. Baulch, C.J. Cobos, R.A. Cox, C. Esser, P. Frank, T. Just, J.A. Kerr, M.J. Pilling, J. 196–203.
Troe, R.W. Walker, J. Warnatz, J. Phys. Chem. Ref. Data 21 (1992) 411–734. [43] D. Healy, D.M. Kalitan, C.J. Aul, E.L. Petersen, G. Bourque, H.J. Curran, Energy
[8] D.L. Baulch, C.T. Bowman, C.J. Cobos, R.A. Cox, T. Just, J.A. Kerr, M.J. Pilling, D. Fuels 24 (2010) 1521–1528.
Stocker, J. Troe, W. Tsang, R.W. Walker, J. Warnatz, J. Phys. Chem. Ref. Data 34 [44] F. Buda, R. Bounaceur, V. Warth, P.A. Glaude, R. Fournet, F. Battin-Leclerc,
(2005) 757–1397. Combust. Flame 14 (2005) 170–186.
[9] C. Zhou, J.M. Simmie, H.J. Curran, Combust. Flame 158 (2011) 726–731. [45] E. Ranzi, A. Frassoldati, R. Grana, A. Cuoci, T. Faravelli, A.P. Kelley, C.K. Law,
[10] C. Zhou, J.M. Simmie, H.J. Curran, Phys. Chem. Chem. Phys. 12 (2010) 7221– Prog. Energy Combust. Sci. 38 (2012) 468–501.
7233. [46] F. Battin-Leclerc, Prog. Energy Combust. Sci. 34 (2008) 440–498.
[11] S. Di Tommaso, P. Rotureau, O. Crescenzi, C. Adamo, Phys. Chem. Chem. Phys. [47] H. Wang, E. Dames, B. Sirjean, D.A. Sheen, R. Tangko, A. Violi, J.Y.W. Lai, F.N.
13 (2011) 14636–14645. Egolfopoulos, D.F. Davidson, R.K. Hanson, C.T. Bowman, C.K. Law, W. Tsang,
[12] Z. Zhao, M. Chaos, A. Kazakov, F.L. Dryer, Int. J. Chem. Kinet. 40 (1) (2008) 1–18. N.P. Cernansky, D.L. Miller, R.P. Lindstedt, JetSurF version 2.0 (2010). <http://
[13] A.V. Joshi, H. Wang, Int. J. Chem. Kinet. 38 (2006) 57–73. melchior.usc.edu/JetSurF/JetSurF2.0>.
[14] V. Vasudevan, D.F. Davidson, R.K. Hanson, J. Phys. Chem. A 109 (2005) 3352– [48] K. Narayanaswamy, G. Blanquart, H. Pitsch, Combust. Flame 157 (2010) 1879–
3359. 1898.
[15] S.S. Vasu, Z. Hong, D.F. Davidson, R.K. Hanson, D.M. Golden, J. Phys. Chem. A [49] G. Blanquart, P. Pepiot-Desjardins, H. Pitsch, Combust. Flame 156 (2009) 588–
114 (2010) 11529–11537. 607.
[16] R. Sivaramakrishnan, J.V. Michael, Combust. Flame 156 (2009) 1126–1134. [50] P. Pepiot-Desjardins, H. Pitsch, Combust. Flame 154 (2008) 67–81.
[17] H. Curran, P. Gaffuri, W. Pitz, C. Westbrook, Combust. Flame 129 (2002) 253– [51] P. Pepiot-Desjardins, PhD. Thesis, Department of Mechanical Engineering,
280. Stanford University, 2008.
[18] C.K. Westbrook, W.J. Pitz, O. Herbinet, H.J. Curran, Emma J. Silke, Combust. [52] H. Pitsch, FlameMaster: A C++ Computer Program for 0D Combustion and 1D
Flame 156 (2009) 181–199. Laminar Flame Calculations.
[19] S.M. Sarathy, C.K. Westbrook, M. Mehl, W.J. Pitz, C. Togbe, P. Dagaut, H. Wang, [53] M. Frenklach, H. Wang, C.-L. Yu, M. Goldenberg, C.T. Bowman, R.K. Hanson, D.F.
M.A. Oehlschlaeger, U. Niemann, K. Seshadri, P.S. Veloo, C. Ji, F.N. Egolfopoulos, Davidson, E.J. Chang, G.P. Smith, D.M. Golden, W.C. Gardiner, V. Lissianski,
T. Lu, Combust. Flame 158 (2011) 2338–2357. 1996. <http://www.me.berkeley.edu/gri mech/version12>.
[20] P. Saxena, F.A. Williams, Proc. Combust. Inst. 31 (2007) 1149–1156. [54] W. Tsang, Combust. Flame 78 (1989) 71–86.
[21] D.A. Sheen, H. Wang, Combust. Flame 158 (2011) 2358–2374. [55] W. Tsang, R.F. Hampson, J. Phys. Chem. Ref. Data 15 (1986) 1087.
[22] M. Frenklach, Combust. Flame 58 (1984) 69–72. [56] N. Peters, G. Paczko, R. Seiser, K. Seshadri, Combust. Flame 128 (2002) 38–59.
[23] M. Frenklach, A. Packard, P. Seiler, R. Feeley, Int. J. Chem. Kinet. 36 (2004) 57– [57] U. Pfahl, K. Fieweger, G. Adomeit, in: Twenty-Sixth-Symposium (International)
66. on Combustion, The Combustion Institute, Pittsburgh, 1998, pp. 362–369.
[24] K. Braman, O. Todd, V. Raman, AIAA (2011). [58] G. Mittal, C.J. Sung, Combust. Flame 150 (2007) 355–368.
[25] S.G. Davis, A.B. Mhadeshwar, D.G. Vlachos, H. Wang, Int. J. Chem. Kinet. 36 [59] M. Ribaucour, R. Minetti, E.M. Sazhina, S.S. Sazhin, Eur. Combust. Meet. (2007).
(2004) 94–106. [60] J.A. Miller, A.E. Lutz, 2003.
[26] M.T. Reagan, H.N. Najm, R.G. Ghanem, O.M. Knio, Combust. Flame 132 (3) [61] R.R. Baldwin, M.W.M. Hisham, R.W. Walker, J. Chem. Soc. Faraday Trans. I
(2003) 545–555. (1982) 1615–1627.
[27] M.T. Reagan, H.N. Najm, B.J. Debusschere, O.P. Le Maitre, O.M. Kino, R.G. [62] K.J. Hughes, P.D. Lightfoot, M.J. Pilling, Chem. Phys. Lett. 191 (1992) 581–586.
Ghanem, Combust. Theory Model 8 (2004) 607–632. [63] S.H. Robertson, P.W. Seakins, M.J. Pilling, in: M.J. Pilling (Ed.), Chemical
[28] M.T. Reagan, H.N. Najm, P.P. Pbay, O.M. Knio, R.G. Ghanem, Int. J. Chem. Kinet. Kinetics, 1997, pp. 125–234.
37 (6) (2005) 368–382.