Characterization and Modeling of Premixed Turbulent N-Heptane Flames in The Thin Reaction Zone Regime

Characterization and modeling of premixed turbulent
n-heptane flames in the thin reaction zone regime
Thesis by
Bruno Savard
In Partial Fulfillment of the Requirements
for the Degree of
Doctor of Philosophy
California Institute of Technology
Pasadena, California
2015
(Defended May 12, 2015)

ii

c 2015
Bruno Savard
All Rights Reserved

iii
Acknowledgments
I gratefully acknowledge funding from the Air Force Office of Scientific Research (Award FA9550-
12-1-0144) under the supervision of Dr. Chiping Li, and from the Natural Sciences and Engineering
Research Council of Canada (NSERC Postgraduate Scholarship D). This research used resources
of the National Energy Research Scientific Computing Center, which is supported by the Office of
Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231.
I would also like to thank Dr. Andy Aspden, who kindly shared with our group the DNS data
presented in Aspden et al. J. Fluid Mech. 680 (2011) 287-320 (used in Chapers 5 and 6).
Finally, I would particularly like to thank my advisor, Prof. Blanquart, for his continuous
support and his motivation for this research project; the committee members, Prof. McKeon, Prof.
Colonius, and Prof. Shepherd, for their valuable comments and suggestions to improve the quality
of this work; and finally, (Prof.) Yuan Xuan, Brock Bobbitt, Simon Lapointe, and Nicholas Burali,
my (ex-)office mates who have contributed to parts of this work.

iv
Abstract
n-heptane/air premixed turbulent flames in the high-Karlovitz portion of the thin reaction zone
regime are characterized and modeled in this thesis using Direct Numerical Simulations (DNS)
with detailed chemistry. In order to perform these simulations, a time-integration scheme that can
efficiently handle the stiffness of the equations solved is developed first. A first simulation with unity
Lewis number is considered in order to assess the effect of turbulence on the flame in the absence of
differential diffusion. A second simulation with non-unity Lewis numbers is considered to study how
turbulence affects differential diffusion. In the absence of differential diffusion, minimal departure
from the 1D unstretched flame structure (species vs. temperature profiles) is observed. In the
non-unity Lewis number case, the flame structure lies between that of 1D unstretched flames with
“laminar” non-unity Lewis numbers and unity Lewis number. This is attributed to effective Lewis
numbers resulting from intense turbulent mixing and a first model is proposed. The reaction zone is
shown to be thin for both flames, although large chemical source term fluctuations are observed. The
fuel consumption rate is found to be only weakly correlated with stretch, although local extinctions
in the non-unity Lewis number case are well correlated with curvature. All these results explain
the apparent turbulent flame speeds. Other variables that better correlate with this fuel burning
rate are identified through a coordinate transformation. It is shown that the unity Lewis number
turbulent flames can be accurately described by a set of 1D (in progress variable space) flamelet
equations parameterized by the dissipation rate of the progress variable. In the non-unity Lewis
number flames, the flamelet equations suggest a dependence on a second parameter, the diffusion of
the progress variable. A new tabulation approach is proposed for the simulation of such flames with
these dimensionally-reduced manifolds.

v
Contents
Acknowledgments iii
Abstract iv
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Turbulent premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Laminar premixed flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Regimes of turbulent premixed flames . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Direct numerical simulation of high Karlovitz premixed turbulent flames . . . . . . . 9
1.4 Time-integration for stiff chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Flame structure and differential diffusion . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6 Reaction zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.7 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.8 Objectives and outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2 Governing equations and numerical solver 21
2.1 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.1 Fluid mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.2 Chemical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 Numerical algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

vi
2.2.1 Overview of the numerical solver . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.2 Extension to multi-dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3 Application to turbulent reacting flows . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 Semi-implicit time-integration for stiff chemistry 32
3.1 Species vs. temperature equation stiffness . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Preconditioned iterative method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 Semi-implicit preconditioning for stiff chemistry . . . . . . . . . . . . . . . . . . . . . 34
3.3.1 Proposed preconditioner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3.2 Extension to multi-dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4.1 One-dimensional premixed flame . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4.2 Three-dimensional turbulent premixed flame . . . . . . . . . . . . . . . . . . 39
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5.1 Theoretical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5.2 Eigenvalue analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5.3 Convergence of the sub-iterations . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.4 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.5.4.1 Theoretical stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.5.4.2 Numerical stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.5.4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5.5 Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5.5.1 Order of accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5.5.2 Magnitude of errors . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5.6 Mass conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5.7 Computational efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
vii
3.6.1 Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.6.1.1 Premixed flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.6.1.2 Non-premixed flames . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.6.1.3 0D ignition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.6.2 Advantages over other methods . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.6.2.1 Fully-implicit method . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.6.2.2 Operator-splitting methods . . . . . . . . . . . . . . . . . . . . . . . 67
3.6.2.3 Stiffness removal through QSSA . . . . . . . . . . . . . . . . . . . . 69
3.6.3 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4 Direct numerical simulations 76
4.1 Numerical approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.1.1 Flow configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.1.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.1.3 Turbulence forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2 Chemistry and transport models validation . . . . . . . . . . . . . . . . . . . . . . . 82
4.2.1 Reduced mechanism performance . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2.2 Transport model validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3 Grid refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4 Turbulent flame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4.1 Thickened flame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4.2 Reaction zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4.3 Lower-Ka flame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5 Characterization and modeling of the flame structure 91
5.1 Structure of the n-heptane flame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.1.1 Turbulent flame structure in the absence of differential diffusion . . . . . . . 92

viii
5.1.2 Turbulent flame structure with differential diffusion . . . . . . . . . . . . . . 94
5.2 Effective species Lewis numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2.1 Turbulent flame structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2.2 Proposed model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2.2.1 Species transport equation . . . . . . . . . . . . . . . . . . . . . . . 101
5.2.2.2 Temperature equation . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.2.2.3 Effective Lewis numbers . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.2.2.4 Turbulent model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.2.3 Evaluation of the effective Lewis numbers and model validation . . . . . . . . 105
5.2.3.1 Sensitivity analysis to species Lewis numbers . . . . . . . . . . . . . 105
5.2.3.2 Computing the effective Lewis numbers . . . . . . . . . . . . . . . . 106
5.2.4 Model dependence on Reynolds versus Karlovitz number . . . . . . . . . . . . 108
5.2.4.1 RANS-based approach . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2.4.2 Flame thickness-based approach . . . . . . . . . . . . . . . . . . . . 110
5.2.4.3 Time scale-based approach . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.4.4 Integral length scale limit . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.4.5 Comparison with DNS results . . . . . . . . . . . . . . . . . . . . . 113
5.2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.2.5.1 Model’s sensitivity to unburnt pressure, temperature, and equiva-
lence ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2.5.2 Practical use of the model . . . . . . . . . . . . . . . . . . . . . . . 115
5.3 Use of the effective species Lewis number model with the n-heptane flame . . . . . . 116
5.4 Limitations of the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6 Characterization of the reaction zone 119
6.1 Turbulent reaction zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.1.1 Source term fluctuations and differential diffusion effect . . . . . . . . . . . . 120
6.1.2 Fuel consumption at Tpeak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

ix
6.2 Stretching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.2.1 Curvature and strain rate distributions . . . . . . . . . . . . . . . . . . . . . 125
6.2.2 Propagating vs. material surface . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2.3 Fuel consumption rate vs. curvature and strain rate . . . . . . . . . . . . . . 129
6.2.4 Fuel consumption rate limiting species vs. curvature and strain rate . . . . . 131
6.3 Turbulent flame speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.3.1 Computing the turbulent flame speed . . . . . . . . . . . . . . . . . . . . . . 134
6.3.2 Link between differential diffusion effects and turbulent flame speed . . . . . 135
6.4 Summary and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7 Modeling of the reaction zone 140
7.1 Progress variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.2 Coordinate transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.3 Transformed transport equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
7.4 Unity Lewis number limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.4.1 Flamelet equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.4.2 Fuel burning rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.4.2.1 Dependence of the burning rate on c and χ . . . . . . . . . . . . . . 150
7.4.2.2 Flamelet-generated manifold . . . . . . . . . . . . . . . . . . . . . . 153
7.5 Non-unity Lewis numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.5.1 Flamelet equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.5.2 Fuel burning rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.5.2.1 Dependence of the burning rate on c, χ, and ξ . . . . . . . . . . . . 159
7.5.2.2 Flamelet-generated manifold . . . . . . . . . . . . . . . . . . . . . . 162
7.6 Tabulation approach with the reduced-dimension manifold . . . . . . . . . . . . . . . 163
7.6.1 Unity Lewis number limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.6.2 Non-unity Lewis numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.7 Summary and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

x
8 Conclusions 170
8.1 Time-integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
8.2 Direct numerical simulations of high Karlovitz n-heptane/air flames . . . . . . . . . 171
8.3 Flame structure and differential diffusion . . . . . . . . . . . . . . . . . . . . . . . . 171
8.4 Reaction zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.5 Reduced-chemistry models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.6 Limitations and suggestions for future work . . . . . . . . . . . . . . . . . . . . . . . 174
A Impacts of the effective Lewis number model 177
A.1 Impact on SL and lF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
A.2 Impact on the regime diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
A.3 Impact on ST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
A.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
B Correlations 183
C Intermediate step in the transformation of the transport equations 186
C.1 Species equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
C.2 Temperature equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
D Transformed transport equations in the unity Lewis number limit 188
D.1 Species equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
D.2 Temperature equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
E Joint probability density function of ω̇c vs. χ at cpeak 190
E.1 Unity Lewis number limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
E.2 Non-unity Lewis numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

xi
List of Figures
1.1 Energy density comparison of several transportation fuels (indexed to gasoline = 1).
Figure taken from the U.S. Energy Information Administration official website (URL:
http://www.eia.gov/todayinenergy/detail.cfm?id=14451). . . . . . . . . . . . . . . . . 2
1.2 Representation of a n-heptane/air flame (equivalence ratio of 0.9, unburnt temperature
of 298 K, and unburnt pressure of 1 atm) in physical space and in temperature space.
YF is the fuel mass fraction and ω̇F is its chemical consumption rate, while T is the
temperature and ω̇T is its chemical production rate (heat release rate). Each quantity
is normalized by its respective maximum value in the flame. . . . . . . . . . . . . . . . 4
1.3 Normalized fuel species mass fraction vs. temperature profiles with non-unity Lewis
numbers and unity Lewis numbers (for all species), for a n-heptane/air flame (same as
in Fig. 1.2) and a hydrogen/air flame (equivalence ratio of 0.4, unburnt temperature
of 298 K, and unburnt pressure of 1 atm). . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Normalized thermodynamic properties vs. temperature profiles for the n-heptane/air
flame (same as in Fig. 1.2). ρ is the density and ν is the kinematic viscosity. . . . . . 5
1.5 Regime diagram for turbulent premixed combustion as suggested by Peters [142]. The
approximate location in this regime of turbulent premixed flames relevant to industrial
applications has been added. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 DNS of high Karlovitz turbulent premixed flames with finite-rate chemistry. . . . . . . 10
3.1 Schematic of the computational domain and initial condition. . . . . . . . . . . . . . . 39
3.2 Schematic of the three-dimensional turbulent premixed flame configuration. . . . . . . 40

xii
3.3 Contours of temperature on a two-dimensional slice of the three-dimensional turbulent
premixed flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Comparison of the chemical timescale (τ ) of the full chemical Jacobian to the species
lifetime of the preconditioned chemical Jacobian. . . . . . . . . . . . . . . . . . . . . . 45
3.5 Relative magnitude of the elements of each row compared to the element on the re-
spective diagonals. ∆t = 5 × 10−6 s. . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.6 Evolution of the density residual as a function of sub-iterations over two time steps for
the proposed semi-implicit time-integration scheme. The dashed lines correspond to
fitted exponential curves averaged over several time steps. . . . . . . . . . . . . . . . . 47
3.7 Rate of convergence of the sub-iterations: comparison between theoretical (largest spec-
tral radius of A00 ) and numerical (rate of convergence of the largest density residuals)
results for the one-dimensional flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.8 Spectral radius of A00 and A00,exp as a function of temperature in the one-dimensional
premixed flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.9 Scatter plot of the spectral radius of A00 , with ∆t = 5.7 × 10−7 s, as a function of
temperature in the three-dimensional turbulent flame. The one-dimensional profile is
added for comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.10 Temporal accuracy of the method as a function of the time step size for the one-
dimensional, propagating flame. The errors for the various species mass fractions are
evaluated as the absolute difference of their integrated value in temperature space
compared with a reference solution obtained with ∆t = 2 × 10−8 s. . . . . . . . . . . 53
3.11 Impact of time step size and number of sub-iteration on the accuracy of 1D propagating
flames. When not mentioned, four sub-iterations are used (Q = 4). . . . . . . . . . . . 54
3.12 Comparison between the joint probability density function of the C2 H4 mass fractions
vs. temperature obtained with different time step sizes. . . . . . . . . . . . . . . . . . 55

xiii
3.13 Maximum deviation in the domain from the inlet elemental mass fractions vs. sim-
ulation time. The results shown are for the three-dimensional turbulent flame with
∆t = 5.7 × 10−7 s and four sub-iterations. . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.14 Maximum deviation in the domain from the inlet elemental mass fractions vs. number
of sub-iterations, Q. The results shown are for the one-dimensional, propagating flame
using ∆t = 2 × 10−6 s. The theoretical convergence rate is added for comparison. . . 57
3.15 Comparison of the chemical timescale (τ ) of the full chemical Jacobian to the species
lifetime of the preconditioned chemical Jacobian at the peak rate of heat release, con-
sidering various fuels and chemical mechanisms. . . . . . . . . . . . . . . . . . . . . . 60
3.16 Spectral radius of A00 vs. the inverse of the time step size for various cases presented in
Table 3.3. The red cross symbols correspond to the profile for the n-C7 H1 6/air flame
with the stiff reaction (Eq. 3.20) removed from CaltechMech. . . . . . . . . . . . . . . 62
3.17 Profiles of n-C3 H7 mass fraction vs. temperature in the one-dimensional, n-C7 H16
propagating flame (similar to the test case of Section 3.4.1) with CaltechMech. The
solutions from using three different time step sizes, each with 4 sub-iterations, are
compared. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.18 Evolution of the density residual as a function of sub-iterations over two time steps for
the n-C7 H16 flame with CaltechMech using the proposed time-integration scheme with
∆t = 2 × 10−6 s. The convergence rates given by the first (red) and the third (black)
largest eigenvalues are shown for comparison. k references to the sub-iteration index. 64
3.19 Contours of temperature (top) and spectral radius of A00 (bottom) from the two-
dimensional coflow laminar flame, obtained with a time step of 4.0 × 10−6 s. . . . . . 65
3.20 n-C3 H7 mass fraction of from the 1D (freely propagating) flame solution. Solutions
from using Godunov splitting (GS) are compared, for different time step sizes, to the
solution using the proposed preconditioning method. . . . . . . . . . . . . . . . . . . . 68

xiv
3.21 Computational cost of 1D stationary flame simulation for different chemical integration
methods. Cost is calculated as cpu time per point (s/pt), per second of simulation time
(s). All cases use four sub-iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.22 Scatter plots of species mass fractions computed from the algebraic expression assuming
QSS vs. their actual value in the three-dimensional turbulent flame. A straight line
(y = x) is expected for perfect QSS species. . . . . . . . . . . . . . . . . . . . . . . . . 71
3.23 Chemical consumption timescale associated with 1-CH2 , 2-C7 H15 , and n-C3 H7 vs. tem-
perature in the one-dimensional flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.1 Schematic diagram of the flow configuration. . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Time-averaged (over 10 eddy turnover times), planar Favre-averaged TKE vs. x for
the non-unity Lewis number flame. The average flame position is shown by the dashed
line (more details on the flame position are given in Section 6.3.1). . . . . . . . . . . . 82
4.3 Normalized energy and dissipation spectra for the higher- and lower-Ka non-unity
Lewis number flames. The two-dimensional three components spectra taken in a y-z
plane in the unburnt gases (averaged over time). κ = 2kπ/L, for k = 1, 2, ..., Ny /2,
with Ny the number of points in the y- or z-direction, is the wavenumber. . . . . . . . 83
4.4 Laminar flame speed vs. equivalence ratio profiles obtained numerically with the 35-
species mechanism and experimentally [51, 170]. The numerical simulations were per-
formed with full transport (solid line) and constant Lewis numbers (dashed line). . . . 84
4.5 Comparison of species mass fraction (top row) and source term (bottom row) vs.
temperature profiles for n-C7 H16 , CH2 O, and H2 O between the one-dimensional, un-
stretched, flame solutions obtained with full transport (solid lines) and constant Lewis
numbers (dashed lines). The unburnt conditions are the same as in Section 4.1. . . . . 85
4.6 Comparison of the fuel burning rate statistics in the non-unity Lewis number turbulent
flame with mixture-averaged diffusivities and constant Lewis numbers. . . . . . . . . . 86
4.7 Comparison of the fuel burning rate statistics in the non-unity Lewis number turbulent
flame with refined and nominal grids. . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

xv
4.8 Contours of vorticity, n-C7 H16 and CH2 O mass fractions, and temperature through the
flame brush on a two-dimensional horizontal slice. The laminar flame thickness lF is
4.9 Contours of the source terms of n-C7 H16 (top) and H2 O (bottom) on the same two-
dimensional horizontal slice as in Fig. 4.8. The isoterm T = 1500 K (white) is also
shown. The laminar reaction zone thicknesses (full width at half-height) of n-C7 H16
(δC7 H16 ) and H2 O (δH2 O ) are also shown for comparison. . . . . . . . . . . . . . . . . 89
4.10 Contours of vorticity, density, and fuel burning rate through the flame brush on a two-
dimensional horizontal slice of the non-unity Lewis number flame. The T = 1240 K
isocontour is also shown (red). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.11 Two-dimensional slices showing the vorticity and temperature for the non-unity Lewis
number flames. The yellow line indicates the T = 940 K isocontour and the red line
indicates the T = 1540 K isocontour. The vorticity ranges are satureated at [0,8e4]
(s−1 ) and [0,1.6e6] (s−1 ). The temperature ranges are both [298,2200] K. . . . . . . . 90
4.12 Two-dimensional slices showing the fuel consumption rate for the non-unity Lewis
number flames. The fuel consumption rate range is saturated at [0,650] (kg m−3 s−1 ).
The T = 1240 K isocontour is also shown (white). . . . . . . . . . . . . . . . . . . . . 90
5.1 Contours of vorticity, n-C7 H16 and CH2 O mass fractions, and temperature through the
flame brush on a two-dimensional horizontal slice. The laminar flame thickness lF is
5.2 Joint PDF and conditional mean (solid line) of the n-C7 H16 (top left), C2 H4 (top right),
and CO2 (bottom) mass fraction vs. temperature from the unity Lewis number DNS.
The unity Lewis number flamelet solution is also shown (dashed line). . . . . . . . . . 93
5.3 Flamelet solutions of the H2 mass fraction vs. temperature with varying diffusivity
coefficients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
xvi
5.4 Joint PDF and conditional mean (solid line) of the C2 H4 mass fraction vs. temperature
from the non-unity Lewis number DNS. The non-unity and unity Lewis number flamelet
solutions are also shown (dashed line). . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.5 Schematic diagram of the flow configuration used by Aspden et al. [10]. Diagram taken
from Ref. [10]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.6 Contours of hydrogen mass fraction and temperature on a two-dimensional vertical
slice of the DNS case D [10]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.7 Conditional mean hydrogen mass fraction profiles as a function of temperature for the
DNS cases A through D and unstretched laminar flames with full transport and unity
Lewis numbers. DNS from Aspden et al. [10] . . . . . . . . . . . . . . . . . . . . . . . 99
5.8 Hydrogen mass fraction profiles as a function of temperature for flamelets with full
transport and constant Lewis numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.9 Contribution from each term in the species transport equation for H2 as a function
of temperature, obtained for an unstretched laminar flamelet with constant, non-unity
Lewis numbers. Each contribution is normalized by the largest absolute value of all
contributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.10 Contribution from each term in the temperature equation as a function of temperature,
obtained for an unstretched laminar flamelet. Each contribution is normalized by the
largest absolute value of all contributions. . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.11 Comparison of H2 and OH mass fraction profiles between DNS conditional means and
flamelets with modified H2 (left), H (center), and all species but H2 and H (right) Lewis
numbers. DNS from Aspden et al. [10] . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.12 Effective Lewis number for H2 versus turbulent Reynolds number obtained by L2-
norm minimization from the DNS where only H2 (black circles) or all Lewis numbers
(red triangles) are modified in the flamelet simulations (three instantaneous snapshots
separated by several eddy turnover times for each case). . . . . . . . . . . . . . . . . . 109

xvii
5.13 Comparison of the effective Lewis number for H2 versus Karlovitz number for three
different models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.14 Conditional mean (solid line) of the n-C7 H16 (top left), C2 H4 (top right), and CO2
(bottom) mass fraction vs. temperature. The non-unity, the unity, and the effective
Lewis number flamelet solutions are also shown (dashed line). . . . . . . . . . . . . . . 117
5.15 Normalized effective diffusivity obtained by solving Eq. 5.18 for the respective fuels
3/2 −1/2
(hydrogen and n-heptane) vs. Karlovitz number (Ka = (u0 /SL ) (l/lF ) to be
consistent with the analysis in Section 5.2). The solid line represents DT /α = aKa Ka
(corresponding to Eq. 5.34) with aKa = 0.05. . . . . . . . . . . . . . . . . . . . . . . . 118
6.1 (a,b) Contours of n-C7 H16 source term normalized by its peak laminar value on a
two-dimensional horizontal slice. Also shown are three temperature isocontours: 600
K (white, left of the reaction zone), temperature of peak source term Tpeak (black),
and 1850 K (white, right of reaction zone). The laminar reaction zone thicknesses of
n-C7 H16 , δC7 H16 is also shown for comparison. (c) Normalized n-C7 H16 source term vs.
distance along the isocontour T = Tpeak . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2 Probability density, at T = Tpeak , of n-C7 H16 source term, ω̇C7 H16 , normalized by its
peak laminar value for both the non-unity and the unity Lewis number cases. . . . . . 121
6.3 (a)Joint PDF and conditional mean (solid line) of n-C7 H16 source term, ω̇C7 H16 , nor-
malized by its peak laminar value vs. temperature from the unity Lewis number simu-
lation. Joint PDF of n-C7 H16 source term, ω̇C7 H16 , normalized, in the direction normal
to the isoterm T = Tpeak , by its value on this isoterm vs. temperature for the unity
Lewis number (b) and the non-unity Lewis number DNS (c). The non-unity and unity
Lewis number flamelet solutions are also shown (dashed line). . . . . . . . . . . . . . . 123
6.4 Contours of ω̇C7 H16 normalized by its laminar value on the isosurface corresponding to
T = Tpeak . (a,b) Le = 1; (c,d) Le 6= 1. (a,c) View from unburnt side; (b,d) view from
burnt side. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

xviii
6.5 Probability density function, on the reaction surface, of κ, normalized by the laminar
reaction zone thickness (top), and at , normalized by the ratio of the laminar reaction
zone thickness to the laminar flame speed (bottom). . . . . . . . . . . . . . . . . . . . 126
6.6 Probability density function, on the isoterm T = Tpeak , of at , normalized by the stan-
dard deviation of the distributions. Distributions from the H2 /air flame of Aspden
et al. [10] with Ka = 28 and Ka = 4200 and from the non-unity Lewis number n-
C7 H16 /air presented in this thesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.7 Joint probability density function, on the isoterm T = Tpeak , of n-C7 H16 source term,
ω̇C7 H16 , normalized by its peak laminar value vs. the normalized tangential strain rate.
(a) Le = 1; (b) Le 6= 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.8 Joint probability density function, on the isoterm T = Tpeak , of n-C7 H16 source term,
ω̇C7 H16 , normalized by its peak laminar value vs. the normalized mean curvature. (a)
Le = 1; (b) Le 6= 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.9 Joint probability density function, on the isoterm T = Tpeak , of n-C7 H16 mass fraction,
YC7 H16 , normalized by its peak laminar value vs. the normalized mean curvature for
the non-unity Lewis number flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.10 Flame drift over time. Recall that the x-axis is positively oriented downstream. A
constant inflow of 1 m/s was imposed at the inlet. The straight black lines correspond
to the turbulent flame speeds obtained from Eq. 6.6. . . . . . . . . . . . . . . . . . . . 135
6.11 Schematic drawing to illustrate Eq. 6.13. SFeff = hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak · SL0 is used. 137
6.12 Flow chart illustrating the mechanism through which turbulence affects the flame speed
for the flames presented in this thesis. SFeff = hω̇F /ω̇F ,lam iTpeak SL0 is used. . . . . . . . 139
7.1 Joint PDF of the progress variable c = YH2 + YH2 O + YCO + YCO2 vs. temperature in
the higher-Ka flame. One-dimensional unstretched flame profiles are shown by dashed
lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
xix
7.2 Joint PDF and conditional mean (solid line) of the fuel burning rate, ω̇F , normalized by
its peak laminar value vs. progress variable c for both the higher-Ka and the lower-Ka
flames. The unity, the non-unity, and the effective Lewis number flamelet solutions are
also shown (dashed line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.3 Mean of the terms in Eq 7.15 (for three different species) conditioned on c for both
unity Lewis number flames, < term|c >. The bars correspond the standard deviation
of these terms conditioned on c,< (term− < term|c >)2 |c >1/2 . The red dashed line
corresponds to the residual of the equation (right hand side of Eq. 7.11 in the unity
Lewis number limit). Each term is normalized by the absolute value of the largest
term (corresponding to the same equation) in a unity Lewis number one-dimensional
unstretched flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.4 Mean of the terms in Eq 7.16 conditioned on c for both unity Lewis number flames,
< term|c >. The bars correspond the standard deviation of these terms conditioned
on c,< (term− < term|c >)2 |c >1/2 . The red dashed line corresponds to the residual
of the equation (right hand side of Eq. 7.12 in the unity Lewis number limit). Each
term is normalized by the absolute value of the largest term (corresponding to the same
equation) in a unity Lewis number one-dimensional unstretched flame. . . . . . . . . . 149
7.5 Joint PDF of the fuel burning rate vs. the dissipation rate of the progress variable,
normalized by their peak laminar values, on the isosurface of c = cpeak in the higher-Ka
unity Lewis number flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.6 Joint PDF of the comparison between the predicted fuel burning rate and the actual
burning rate in the higher-Ka unity Lewis number flame. . . . . . . . . . . . . . . . . 152
7.7 Comparison of the fuel burning rate vs. progress variable profile between the 1D flame
solved in physical space and the flamelet solved in c-space (Eq. 7.15 and 7.16 with
χ (c) = χphys (c)). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

xx
7.8 Comparison between the fuel burning rate predicted by ω̇F,FGM (c, χ) and < ω̇F | (c, χ) >
as a function of the dissipation rate, normalized by their peak laminar values, given
c = cpeak . The joint PDF on the isosurface of c = cpeak in the higher-Ka unity Lewis
number flame is also shown for comparison. . . . . . . . . . . . . . . . . . . . . . . . . 154
7.9 Mean of the terms in Eq 7.22 (for three different species) conditioned on c for both non-
unity Lewis number flames, < term|c >. The bars correspond the standard deviation
of these terms conditioned on c,< (term− < term|c >)2 |c >1/2 . The red dashed line
corresponds to the residual of the equation (right hand side of Eq. 7.11). Each term
is normalized by the absolute value of the largest term (corresponding to the same
equation) in a non-unity Lewis number one-dimensional unstretched flame. . . . . . . 158
7.10 Mean of the terms in Eq 7.24 conditioned on c for both non-unity Lewis number flames,
< term|c >. The bars correspond the standard deviation of these terms conditioned
on c,< (term− < term|c >)2 |c >1/2 . The red dashed line corresponds to the residual
of the equation (right hand side of Eq. 7.12). Each term is normalized by the absolute
value of the largest term (corresponding to the same equation) in a non-unity Lewis
number one-dimensional unstretched flame. . . . . . . . . . . . . . . . . . . . . . . . . 159
normalized by their peak laminar values (at c = clam,peak ), on the isosurface of c = cpeak
in the higher-Ka non-unity Lewis number flame. . . . . . . . . . . . . . . . . . . . . . 160
normalized by their peak laminar values (at c = clam,peak ), on the isosurface of c =
cpeak in the higher-Ka non-unity Lewis number flame, conditional on ξ ∈ Ξk , with
k = low, mid, high. Ξlow = − 43 ξlam,peak , − 23 ξlam,peak , Ξmid = 23 ξlam,peak , 43 ξlam,peak ,

8 10

Ξhigh = 3 ξlam,peak , 3 ξlam,peak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.13 Joint PDF of the comparison between the predicted fuel burning rate and the actual
burning rate in the higher-Ka non-unity Lewis number flame. . . . . . . . . . . . . . . 162

xxi
7.14 Joint PDF of the thermodynamic properties vs. progress variable for the higher-Ka
unity Lewis number flame. The 1D flame profiles are also shown (dashed lines). . . . 165
7.15 Joint PDF of the thermodynamic properties vs. progress variable for the higher-Ka
non-unity Lewis number flame. The unity, non-unity, and effective Lewis number 1D
flame profiles are also shown (dashed lines). . . . . . . . . . . . . . . . . . . . . . . . . 167
7.16 Joint PDF and conditional mean (solid line) of the progress variable diffusivity vs.
progress variable for both non-unity Lewis number flames. The unity, non-unity, and
effective Lewis number 1D flame profiles are also shown (dashed lines). . . . . . . . . 168
A.1 Effective laminar flame speed and flame thickness versus Karlovitz number from flamelet
simulations (solid lines) and analytical expressions (dotted lines). . . . . . . . . . . . . 178
A.2 Regime diagram taking into account the effective Karlovitz number, considering the
Ka-based model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
A.3 Normalized turbulent flame speed profiles for the models given by Eq. A.7-A.10 for
Le =0.3 (solid lines), 1 (symbols), and 3 (dotted lines). . . . . . . . . . . . . . . . . . 182
B.1 Joint probability density function, on the isoterm T = Tpeak , of the normalized strain
rate vs. the normalized mean curvature. (a) Le = 1; (b) Le 6= 1. . . . . . . . . . . . . 184
E.1 Comparison between the fuel burning rate predicted by ω̇c,FGM (c, χ) and < ω̇c | (c, χ) >
as a function of the dissipation rate, normalized by their peak laminar values, given
c = cpeak . The joint PDF on the isosurface of c = cpeak in the turbulent flames is also
shown for comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
E.2 Joint PDF of the progress variable source term vs. the dissipation rate of the progress
variable, normalized by their peak laminar values (at c = clam,peak ), on the iso-
surface of c = cpeak in the higher-Ka non-unity Lewis number flame, conditional

5 3

on ξ ∈ Ξk , with k = low, mid, high. Ξlow = 4 ξlam,c-peak , 2 ξlam,c-peak , Ξmid =
7 9
, Ξhigh = 21 ξlam,c-peak , 34 ξlam,c-peak . . . . . . . . . . . . . . . 192

8 ξlam,c-peak , 8 ξlam,c-peak
xxii
List of Tables
3.1 Largest stable time step size for the proposed semi-implicit scheme and the explicit
time-integration of the chemical source terms for the 1D flame test case. Numerical
and theoretical results (see Section 3.5.4.1) are compared. . . . . . . . . . . . . . . . . 50
3.2 Laminar flame speed obtained from simulations with various time step sizes and number
of sub-iterations. ρmax (A00 ) is the theoretical maximum spectral radius of A00 in the
flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3 Theoretical largest stable time step size for the proposed semi-implicit scheme and
the explicit time-integration of the chemical source terms with various unstretched
one-dimensional premixed flames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4 Theoretical largest stable time step size for the proposed semi-implicit scheme and
the explicit time-integration of the chemical source terms with non-premixed flamelets.
χst is the scalar dissipation rate at stoichiometry, Tf the temperature on the fuel side,
and To the temperature on the oxidizer side. The oxidizer is air and the chemical
mechanism considers 47 species and 290 reactions [17]. . . . . . . . . . . . . . . . . . . 65
3.5 Theoretical largest stable time step size for the proposed semi-implicit scheme and the
explicit time-integration with a 0D isobaric ignition case. T0 is the initial temperature. 65
3.6 Comparison of the ignition delay time tign and the burnt temperature obtained with
FlameMaster, and with the proposed framework using two different time step sizes.
Four sub-iterations are used. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

xxiii
4.1 Parameters of the simulation. ∆x is the grid spacing (uniform), η the Kolmogorov
length scale in the unburnt gas, nF the number of grid points through the laminar
flame thickness, ∆t the time step, φ the equivalence ratio, and Ret the turbulent
Reynolds number in the unburnt gas. . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.2 Constant species Lewis numbers used in the turbulent flame simulations (evaluated in
the burnt gas from a one-dimensional flame solution). . . . . . . . . . . . . . . . . . . 80
5.1 Parameters for the series of turbulent premixed hydrogen flame DNS performed in
Ref. [10]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2 Lewis numbers used in the constant, non-unity Lewis numbers model. These Lewis
numbers are evaluated at 723 K from a full transport flamelet solution. . . . . . . . . 100
5.3 Effective Lewis numbers obtained from the DNS (average from three instantaneous
snapshots separated by several eddy turnover times). . . . . . . . . . . . . . . . . . . . 108
6.1 SL0 /vη ratios (see Ref. [185]) and curvature and strain rate statistics for the flames
presented in this thesis and for two lean H2 /air (φ = 0.4) flames presented in [10]. vη
is evaluated at Tpeak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.2 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between strain rate
and fuel consumption, and between curvature and fuel consumption at T = Tpeak for
both flames presented the unity and the non-unity Lewis number flames. . . . . . . . 130
6.3 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between strain
rate and species mass fractions, and between curvature and species mass fractions at
T = Tpeak for both (higher-Ka) flames presented in this thesis. . . . . . . . . . . . . . 133
6.4 Ratios relevant to the turbulent flame speed (see Eq. 6.13). Confidence intervals cor-
respond to plus or minus one standard deviation. . . . . . . . . . . . . . . . . . . . . . 138
7.1 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between strain rate
and fuel consumption rate, curvature and fuel consumption rate, and dissipation rate
and fuel consumption rate at c = cpeak for the higher-Ka unity Lewis number flame. . 151
xxiv
7.2 Prediction error (Eq. 7.17) for various estimators of the fuel burning rate normalized
by the prediction error for f = 0 in both unity Lewis number flames. . . . . . . . . . . 152
7.3 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between fuel burn-
ing rate and various variables at c = cpeak for the higher-Ka non-unity Lewis number
flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.4 Prediction error (Eq. 7.17) for various estimators of the fuel burning rate normalized
by the prediction error for f = 0 in both non-unity Lewis number flames. . . . . . . . 163
7.5 Prediction error (Eq. 7.17) for various estimators of the progress variable source term
normalized by the prediction error for f = 0 in both unity Lewis number flames. . . . 166
7.6 Prediction error (Eq. 7.17) for various estimators of the progress variable source term
normalized by the prediction error for f = 0 in both non-unity Lewis number flames. . 168
B.1 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between strain rate
and curvature at T = Tpeak for both (higher-Ka) flames presented in this thesis. . . . 185
E.1 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between dissipation
rate and progress variable source term at c = cpeak for the higher-Ka unity Lewis
number flame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
E.2 Pearson’s correlation coefficient, r, and distance correlation, dCorn , between dissipation
rate and progress variable source terms for various intervals of ξ at c = cpeak for the
higher-Ka non-unity Lewis number flame. . . . . . . . . . . . . . . . . . . . . . . . . . 191

1
Chapter 1
Introduction
1.1 Background
While considerable worldwide efforts have been made towards the development of sustainable energy
sources, the U.S. Energy Information Administration projects that in 2040, more than 90 % of the
world’s energy consumption in the transportation sector (and more than 30 % of the world’s total
energy consumption) will still come from petroleum and other liquid fuels1 . Liquid fuels remain
ideal for transportation because of their high energy density, as illustrated in Fig. 1.1. This is a
consequence of the large hydrocarbon species these fuels contain. Unfortunately, the combustion
products of such fuels can include pollutants such as NOx and soot.
In an effort to reduce the emission of these pollutants, most combustion devices recently developed
rely on the combustion of premixed fuel/air mixtures. As opposed to non-premixed flames, in
which the fuel and the oxidizer react at a local stoichiometric equivalence ratio, premixed fuel/air
mixtures can be set to a low equivalence ratio (lean mixture), i.e. the oxidizer is in a higher relative
concentration compared to stoichiometry. With lean mixtures, the burnt temperature is reduced,
which results in a reduction of the NOx concentration, and no soot is produced, as these particles
are only produced under rich conditions (higher relative concentration of fuel).
The success of many of these combustion devices (in jet engines, gasoline internal combustion
(IC) engines, and some homogeneous compression charge ignition (HCCI)-like engines) relies on the
1 From the Annual Energy Outlook 2015 report (can be found at URL:
http://www.eia.gov/forecasts/aeo/index.cfm/).
2
Figure 1.1: Energy density comparison of several transportation fuels (indexed to gasoline =
1). Figure taken from the U.S. Energy Information Administration official website (URL:
http://www.eia.gov/todayinenergy/detail.cfm?id=14451).
presence of highly turbulent premixed flames. Turbulence enhances the flame propagation speed,
which allows for a larger fuel mass flow rate in a jet engine or for larger rotation speed in IC engines.
This results in an increased power output. In military applications, scramjets often use a cavity
flame holder on which a turbulent premixed flame is anchored. Their performance is also affected
by the turbulent flame speed.
The development process of such combustion devices relies on numerical simulation as an increas-
ingly important tool. However, state-of-the-art numerical tools used in the industry rely on models
validated at much lower turbulence intensities and with simple fuels (e.g. methane). The extent to
which these models are adequate for highly turbulent premixed flames with large hydrocarbon fuels
remains an open question.
1.2 Turbulent premixed combustion
Combustion is characterized by the following schematic reaction:
fuel + oxidizer → products + heat (1.1)

3
In most practical applications of combustion, the oxidizer is the oxygen contained in air. A flame
can be seen as a region (in a flow field) of self-sustained combustion. Premixed flames are flames
in which the reactants are mixed prior to reacting. They are characterized by a flame speed (speed
at which the flame front moves with respect to the unburnt gas) and a flame thickness. Besides the
nature of the fuel and the oxidizer, premixed flames are characterized by the following conditions
in the unburnt side: the temperature, the pressure, and the equivalence ratio. A unity equivalence
ratio corresponds to a stoichiometric mixture. If it is larger than unity, the mixture is rich, i.e. the
oxidizer is the limiting reactant. Inversely, if the equivalence ratio is less than unity, the mixture is
lean, and the fuel is the limiting reactant. A brief description of turbulent premixed combustion is
provided below. Accent is put on concepts relevant to the present thesis.
1.2.1 Laminar premixed flames
For laminar flames, the flame speed is denoted SL and the flame thickness can be defined as
Tb − Tu
lF = , (1.2)
|∇T |max
with Tb the temperature in the burnt gas, Tu the temperature in the unburnt gas, and |∇T |max the
maximum temperature gradient in the flame. Figure 1.2 is representative of a large hydrocarbon
one-dimensional laminar premixed flame. Representation in both physical space and temperature
space is provided. The flame corresponds to a n-heptane/air flame with equivalence ratio of 0.9,
unburnt temperature of 298 K, and unburnt pressure of 1 atm. YF is the fuel mass fraction and
ω̇F is its chemical consumption rate, while T is the temperature and ω̇T is its chemical production
rate (heat release rate). Three regions can be identified: the preheat zone, where no reaction occurs
(only diffusion and convection), the reaction zone, where the fuel is burnt and most of the heat is
released, and the oxidation layer, where slow oxidation reactions take place to form the products
(conversion of CO to CO2 ).
Representing the flame structure in temperature space is particularly useful to highlight the
4
1
1
Normalized magnitude
0.8
Normalized magnitude 0.8
YF
0.6 -ωF 0.6
T
ωT
0.4 0.4
YF
0.2 0.2 -ωF
ωT
0 0
-1 -0.5 0 0.5 1 1.5 0 0.2 0.4 0.6 0.8 1
x/lF (T-Tu)/(Tb-Tu)
(a) physical space (b) temperature space
Figure 1.2: Representation of a n-heptane/air flame (equivalence ratio of 0.9, unburnt temperature
of 298 K, and unburnt pressure of 1 atm) in physical space and in temperature space. YF is the fuel
mass fraction and ω̇F is its chemical consumption rate, while T is the temperature and ω̇T is its
chemical production rate (heat release rate). Each quantity is normalized by its respective maximum
value in the flame.
1 1
LeF=1 LeF=1
0.8 LeF>1 0.8 LeF<1
0.6 0.6
YF/YF,u
YF/YF,u
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(T-Tu)/(Tb-Tu) (T-Tu)/(Tb-Tu)
(a) n-heptane/air (b) hydrogen/air
Figure 1.3: Normalized fuel species mass fraction vs. temperature profiles with non-unity Lewis
numbers and unity Lewis numbers (for all species), for a n-heptane/air flame (same as in Fig. 1.2)
and a hydrogen/air flame (equivalence ratio of 0.4, unburnt temperature of 298 K, and unburnt
pressure of 1 atm).
effects of differential diffusion. Differential diffusion is characterized by non-unity Lewis numbers
Lei (one Lewis number for each species i), which are the ratios of the thermal diffusivity α to the
species mass diffusivities Di . Figure 1.3 shows typical fuel mass fraction vs temperature profiles for
5
fuels with different Lewis numbers. Differential diffusion is an important aspect of premixed flames.
First, it has a considerable impact on the laminar flame speed. For the two flames considered (Fig.
1.4), SL changes from 29 cm/s to 36 cm/s for n-heptane and from 45 cm/s to 22 cm/s for hydrogen,
respectively between unity and non-unity Lewis numbers. Second, lean premixed flames with less
than unity fuel Lewis numbers are thermodiffusively unstable, whereas lean premixed flames with
greater than unity fuel Lewis numbers are thermodiffusively stable (for rich flames, the Lewis number
of the oxidizer controls the stability). Thermodiffusively unstable laminar flames do not maintain a
planar shape: cells and cusps are formed. This strongly affects the propagation speed and regions
of local extinctions can be created. More details can be found in Ref. [113, 133, 114, 4, 53]. Finally,
the response of flames to stretch (strain rate or curvature) is largely affected by differential diffusion
(more details in Section 1.6).
Another aspect of premixed flames which will be particularly relevant to turbulence, is the evolu-
tion of the thermodynamic properties through the flame. Since the temperature and the composition
of the species in the gas evolve through the flame, the thermodynamic properties are not constant,
as illustrated in Fig 1.4.
1
ρ
α
Normalized magnitude
0.8 ν
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
(T-Tu)/(Tb-Tu)
Figure 1.4: Normalized thermodynamic properties vs. temperature profiles for the n-heptane/air
flame (same as in Fig. 1.2). ρ is the density and ν is the kinematic viscosity.
Note that for the purpose of this thesis, a “flamelet” will be defined as a one-dimensional un-
stretched flame.
6
1.2.2 Turbulence
As mentioned in Section 1.1, flames in practical combustion devices are always subject to a turbulent
flow, which allows for an increase in the turbulent flame speed and consequently, a larger energy
conversion rate, or a smaller engine. For the purpose of this thesis, only a brief description of
turbulent quantities is provided. More details can be found in Ref. [178, 152].
Turbulence can be seen as a random collection of coherent structures, often called eddies. The
largest turbulent length scale is called the integral length scale l. It is associated with the rms
velocity fluctuations u0 and the eddy turnover time
l
τ= . (1.3)
u0
The Turbulent Kinetic Energy (TKE) is defined as
3 02
k= u . (1.4)
2
The smallest scale, the Kolmogorov length scale η, is the scale at which the kinetic energy is converted
into heat, through viscous dissipation. The length scale η is related to the dissipation rate and the
kinematic viscosity ν as
1/4
ν3

η= . (1.5)

This dissipation rate is related to the largest scales by
3
u0
= . (1.6)
l
The velocity associated with the Kolmogorov length scale is
1/4
vη = (ν) (1.7)
7
and the associated time scale is

ν 1/2
τη = . (1.8)

The ratio between the largest and the smallest scales is related to the turbulent Reynolds number
ReT as
l 3/4
= ReT , (1.9)
η
with
u0 l
ReT = . (1.10)
ν
1.2.3 Regimes of turbulent premixed flames
Turbulent premixed flames are typically characterized by the extent to which turbulence is “ex-
pected” to penetrate/disrupt the flame. The Karlovitz number
τF
Ka = , (1.11)
τη
which is the ratio of the flame time scale τF = lF /SL to the Kolmogorov time scale, provides
such information [142]. The larger the Karlovitz number, the more turbulence is expected to pene-
trate/disrupt the flame. In order to push the characterization further, several regime diagrams have
been proposed in the literature [20, 137, 1, 150]. Amongst them, Peters’ regime diagram [142] is
one of the most widely used (Fig. 1.5). A brief description similar to the one found in Ref. [142] is
provided in this section.
For scaling purposes, it is assumed for the rest of this section only that the viscosity and the
thermal diffusivity are equal, i.e. ν = α (unity Schmidt number), and the laminar flame thickness
is equal to the ratio of the thermal diffusivity and the laminar flame speed, i.e. lF = α/SL . With
these assumptions, the turbulent Reynolds number can be written as
u0 l
ReT = , (1.12)
SL lF
8
and the Karlovitz number simplifies to

lF2
Ka = . (1.13)
η2
An additional Karlovitz number that is argued to be more appropriate for the reaction zone is
defined as
2
δ2

δ
Kaδ = = Ka, (1.14)
η2 lF
with δ the laminar reaction zone thickness. Using Eq. 1.9, the ratio u0 /SL can be related to the
ratio l/lF through the turbulent Reynolds number or the Karlovitz number
−1
u0

l
= ReT (1.15)
SL lF
1/3
−2/3 l
= Ka .
lF
Using Eq. 1.15, boundaries delimiting different regimes can be drawn in the regime diagram
(Fig. 1.5). First, setting ReT = 1 separates laminar flames from turbulent flames. Second, the
limit of Ka = 1, known as the Klimov-William criterion, delimitates the thin reaction zone regime
from the corrugated flamelet regime. In the thin reaction zone regime, turbulent eddies are smaller
than the flame thickness and are expected to thicken the preheat zone through turbulent mixing.
In the corrugated flamelet regime, eddies are too large to penetrate the flame, but are sufficiently
intense to significantly corrugate the flame front. For fluctuating velocities smaller than the laminar
flame speed, u0 /SL < 1, the flame is said to be wrinkled (as opposed to be corrugated), hence the
wrinkled flamelet regime. Finally, the limit of Kaδ = 1 separates the broken/distributed reaction
zone regime from the thin reaction zone regime. For Kaδ > 1, the smallest eddies are sufficiently
small to penetrate the reaction zone and are expected to thicken it (distributed reaction zone) or
disrupt it (broken reaction zone). δ is often assumed to be ten times smaller than lF (for scaling
purposes) and hence the Kaδ = 1 line corresponds to Ka = 100.
The wrinkled and the corrugated flamelet regimes have been widely studied (theoretically, ex-
perimentally, and numerically) [113, 114, 133, 4, 53]. The thin reaction zone has been explored in
9
3
10
broken reaction
zones
2
10
industrial
applications
u’/sL
101 thin reaction zones
corrugated
0 flamelets
10
laminar
flames wrinkled flamelets
10-1
10-1 100 101 102
l/lF
Figure 1.5: Regime diagram for turbulent premixed combustion as suggested by Peters [142]. The
approximate location in this regime of turbulent premixed flames relevant to industrial applications
has been added.
several studies, but most of them focused on the lower-Ka part of the regime (due to the experimen-
tal challenges and the computational cost at higher Ka) [38, 78, 160, 176, 70, 11]. The higher-Ka
part of the reaction zone regime and the transition to the broken/distributed reaction zone regime
have been investigated in a only a few studies [10, 9].
Most practical applications fall in the thin reaction zone regime or at the transition between the
thin and broken/distributed reaction zone regimes, as shown in Fig. 1.5. Some military applications
fall even further up in the broken/distributed reaction zone regime [153].
It is important to note that the regime diagram is based on scaling arguments. For instance, it
does not account for any potential effect of differential diffusion, or the nature of the fuel. In addition,
it does not account for the changes in thermodynamic properties through the flame (Fig 1.4).
1.3 Direct numerical simulation of high Karlovitz premixed
turbulent flames
As industrial applications of turbulent premixed (and partially premixed) flames fall in the thin/broken
reaction zone regimes, understanding how a flame behaves in these regimes is critical [148]. Experi-
10
103
Aspden et al., JFM 2011, PCI 2011, PCI2015
Sankaran et al., PCI 2007
broken reaction Wang et al., Energ. Fuel 2012
zones Hawkes et al., CNF 2012
102 ,
CH4 C3H8
present work
H2 n-C7H16
H2 thin reaction zones
u’/sL
101
H2 H2
CH4 corrugated
flamelets
100
laminar
10-1
10-1 100 101 102
l/lF
Figure 1.6: DNS of high Karlovitz turbulent premixed flames with finite-rate chemistry.
ments are difficult to conduct at high Karlovitz numbers and a limited number of them are available
in the literature [38, 78]. Equivalently, due to their expensive computational costs (more details in
the next section), very few Direct Numerical Simulations (DNS) of turbulent premixed flames in the
broken reaction zone or the thin reaction zone regimes have been performed [10, 9, 160, 176, 70, 11].
Figure 1.6 presents all of these simulations (to the best of the authors’ knowledge) in the context
of the regime diagram (as proposed by Peters [142]). Note that only simulations performed with
detailed finite-rate chemistry are presented (an additional study in the context of astrophysics can
be found in Ref. [8]).
The only fuels that have been considered in previous simulations are hydrogen [10, 176], methane [9,
160], and propane [9]. While these fuels are used in several ground-based applications [41], most of
the fuels used for transportation contain larger hydrocarbons [44, 117]. Since their chemical path-
ways are far more complex, and a wide range of stable species are present through the flame front,
it remains unclear how turbulence influences their chemistry at high Karlovitz number.
Moreover, heavy hydrocarbons have large Lewis numbers (e.g. LeC12 H26 ≈ 3.5). It has also been
observed that for sufficiently high Karlovitz number differential diffusion effects were negligible for
both H2 (LeH2 ≈ 0.3) and C3 H8 (LeC3 H8 ≈ 2) [9]. However, this similar behavior between smaller
and larger than unity Lewis number fuels cannot be generalized to intermediate Karlovitz numbers
11
as pertinent to the transition between the thin/broken reaction zone regimes. Furthermore, while
the series of lean H2 /air premixed flames performed by Aspden et al. [10] have provided information
on how turbulence affects differential diffusion over a wide range of Karlovitz numbers, there is no
such information available for heavy hydrocarbon fuels.
To tackle these questions, a series of DNS of a premixed n-C7 H16 turbulent flames in the thin
reaction zone regime and close to the transition between the thin/broken reaction zones regimes are
targetted in this thesis. n-C7 H16 is chosen for the fuel because it is used in surrogates for gasoline [46,
117]. In addition, it corresponds to a step towards the simulation of n-decane or n-dodecane flames
(larger chemical mechanism), which species are often used in surrogates for kerosene [129, 47].
Two sets of simulations at different Karlovitz numbers are targetted. For each of these sets
(higher-Ka and lower-Ka), a first simulation with unity Lewis number is considered in order to assess
the effect of turbulence on the flame in the absence of differential diffusion. A second simulation
with non-unity Lewis numbers (amongst which LeC7 H16 = 2.8) is considered to study how turbulence
affects differential diffusion.
More specifically, the following two aspects need to be characterized with these target flames:
1. the effect of turbulence on the flame structure and the impact of turbulent mixing on differential
diffusion (more details in Section 1.5),
2. the effect of turbulence on the reaction zone with and without differential diffusion (more
details in Section 1.6).
Modeling these effects for larger scale simulations corresponds to the subsequent challenge (more
Before the characterization and the modeling can be achieved, the DNS need to be performed
within reasonable computational time. As detailed in the next section, the time-integration of
chemically reacting flows is particularly expensive because of the stiffness of the equations solved.
A wide range of methods have been developed over the past few decades to overcome this issue.
However, most of these methods are not optimal for the simulation of highly turbulent flames. A
12
detailed discussion follows.
1.4 Time-integration for stiff chemistry
Simulations of reacting flow systems using detailed finite-rate chemistry are extremely challeng-
ing [164]. The expensive nature of the chemical source terms integration comes from four main
challenges: 1) their high non-linearity in the Arrhenius form of the chemical reaction rate constants
(i.e. high computational cost for each function evaluation) [16], 2) the typically large number of
species involved, 3) the strong coupling between chemistry and transport processes (convection and
diffusion) [128], and 4) their very large magnitude (or equivalently small timescales) [112, 85]. As
a result of all these challenges, detailed chemical mechanisms including a large number of species
(above 50) and reactions (above 200) have been included in the numerical simulations of reacting
flows only for relatively simple geometries (e.g. homogeneous or stratified reactors and statistically
one-dimensional flames) [186, 135, 187, 15, 182, 9]. The number of species (and number of reac-
tions) included in the numerical simulations of two-dimensional and three-dimensional turbulent
flames has been relatively limited [17, 105, 104, 37, 9, 10]. Most of these simulations have been
focused on investigating the combustion of relatively simple fuels (e.g. hydrogen, methane, and
ethylene). Beside the work presented in this thesis, only one study has considered a turbulent flame
with large hydrocarbon fuels (propane) [9] due to the large inherent simulation cost. In order to
perform numerical simulations with detailed chemical kinetics, robust, accurate, and efficient nu-
merical algorithms are needed for solving the coupled, highly non-linear, multi-dimensional, Partial
Differential Equations (PDE) governing the unsteady evolution of these complex reacting systems.
There exist various methods to improve the efficiency of chemical source term integration in reacting
flow problems. These are reviewed in the following paragraphs and are organized according to the
challenges mentioned above.
First, the computational cost associated with the evaluation of exponential functions in the
chemical source terms (challenge #1 mentioned above) can be reduced by using single precision
calculations, or by tabulating the exponential functions [124]. However, the associated efficiency gain
13
is not significant and the loss of accuracy might be problematic given the large range of timescales
in a reacting flow simulation. Alternatively, the non-linear chemical source terms may be expanded
using some type of low-order expansion [49, 156, 130]. However, it has been shown that these
methods may be subject to severe time step size restrictions and stability issues when applied to
combustion simulations [49], since they do not fully account for the non-linearity of the system [156].
Second, regardless of the chosen time-integration scheme, the cost of the chemical models could
be alleviated by reducing the total number of species (challenge #2 mentioned above). This can be
accomplished using Quasi-Steady-State (QSS) assumptions and Partial-Equilibrium (PE) approxi-
mations [107, 110], or more advanced methods such as Directed Relation Graph (DRG) [106] and
DRG with Error Propagation (DRGEP) [134] before being applied to the simulation [110, 139]. Even
with these techniques, it has been pointed out in previous work that the size of reduced mechanisms
for practical hydrocarbon fuel surrogates is still too large to be used directly in Direct Numeri-
cal Simulations (DNS) with finite-rate chemistry [66]. Alternative chemistry reduction techniques
based on separation of chemical timescales could be applied. Such techniques include the Com-
putational Singular Perturbation (CSP) method [102] and the Intrinsic Low Dimensional Manifold
(ILDM) [112] method. However, these methods require significant computational efforts to conduct
chemical Jacobian decomposition and mode separation [66], which makes them not suited for the
DNS of multi-dimensional reacting flows.
Third, to avoid the cost associated with the coupled reactive-transport system (challenge #3
mentioned above), most numerical frameworks rely on some variant of splitting techniques (e.g.
Godunov [12] or Strang [166] splitting) followed by the chemical source term integration in a zero-
dimensional setting. These techniques have been widely applied in the numerical simulations of
turbulent reacting flows, for instance in the code developed at Lawrence Berkeley National Labo-
ratory [52, 3]. The lagging errors introduced by the operator splitting treatment are unimportant
for steady-state configurations, but may be substantial in some circumstances, for instance in the
proximity of unsteady premixed flame fronts [128, 95], and become more severe when running with
larger time step sizes. The effects of these errors have been the subject of many previous stud-
14
ies [95, 190, 189, 93, 90], and have been found to be case-dependent. Note that, with the application
of these techniques, the resulting Ordinary Differential Equations (ODE) associated with the chem-
istry (instead of the coupled PDEs) remain stiff. To alleviate the high computational overhead
associated with the integration of these stiff ODEs, methods relying on implicit numerical schemes
based on Backward-Differentiation Formulas (BDF) have been developed [159, 158] and implemented
in packages such as VODE [25] and DASSL [143, 23]. These packages integrate stiff chemical kinetics
using BDFs with a modified iterative Newton procedure [143, 23, 155, 161], and have been widely
adopted in numerical simulations of chemically reacting flows [128, 95, 190]. Despite the signifi-
cant computational efficiency gain brought by the stiff chemistry integration techniques discussed
above, it is important to recall that these techniques are designed for time-dependent ODE systems
(and not PDE), which arise from the application of time-splitting techniques to separate the reactive
(chemical kinetic) part of the PDE system (species transport equations) from the convective-diffusive
part [6].
Forth, time-integration techniques designed for the coupled PDEs governing the unsteady evolu-
tion of complex reacting systems are discussed in the following. First and foremost, species chemical
source terms can be integrated explicitly. For instance, an explicit time-integration scheme, along
with QSS assumptions, is used in S3D [72], a massively parallel DNS solver developed at Sandia
National Laboratories for the simulations of compressible, turbulent reacting flows. This code has
been applied to the simulation of turbulent flames with relatively simple fuels, for instance hydro-
gen [121, 103], methane [160], and ethylene [188]. Heavier fuels have been considered in ignition
simulations of HCCI-like systems (e.g. n-heptane [186], iso-octane [187], and ethanol [15]). The ap-
plication of explicit time-integration methods are commonly limited by prohibitively small time step
sizes to resolve the smallest chemical timescales present in the system (challenge #4) [66, 25]. This
is the reason why the S3D code relies on “stiffness removal” techniques such as QSSA [107, 160, 110].
On the other hand, implicit time-integration methods generally yield better stability characteristics
and allow for larger integration time step sizes than explicit methods [26]. Unfortunately, fully-
implicit methods are generally prohibitively expensive [66], especially for unsteady problems [39],
15
due to the large computational overhead for chemical Jacobian inversion within each time step.
This makes fully-implicit time-integration methods prohibitive for simulations of reacting flows with
large hydrocarbon fuels (e.g. n-heptane, kerosene, and diesel), where up to hundreds of species
and reactions are typically required [175, 109, 19]. Further, Krylov-based iterative methods have
been proposed [16, 116, 169] to reduce the computational burden associated with the construction,
storage, and inversion of large, often non-sparse, Jacobian matrices [6, 54]. Alternatively, chemi-
cal Jacobian diagonal-preconditioning has also been proposed for the time-integration of the PDE
system [85, 131, 57].
While these simple diagonal preconditioners have been used for the simulation of steady-state
chemically reacting flows [85, 131, 57], they were argued to be inappropriate for time-accurate
simulations of unsteady flows [85]. That is why large efforts have been put in the development of
iterative preconditioning methods for solving the ODEs describing 0D chemical systems. Examples
of efficient methods can be found in Ref. [135, 116]. However, these methods are typically tailored for
very large chemical mechanisms (thousands of species) and are considerably more computationally
expensive than explicit time-integration for mechanisms of small to medium sizes (tens to hundreds of
species, with hundreds of reactions). In addition, these preconditioning methods rely on a decoupling
of the chemistry and transport. In other words, the inversion of the sparse-chemical Jacobian
is spatially local and not global. Consequently, computationally less expensive preconditioning
iterative methods applied to the PDE system are desirable. Development of such a method would
allow the target flames (Section 1.3) to be simulated with reasonable computational resources.
1.5 Flame structure and differential diffusion
For non-premixed turbulent flames, the flame structure is often assumed to be similar to that of a
unity Lewis number flame, even when the species have Lewis numbers far from unity [45, 144, 149].
The underlying reason is that at high turbulence levels, diffusion of species and temperature is
dominated by turbulent mixing, resulting in an effective unity Lewis number. This gives good
results at sufficiently high turbulence, as experimentally observed [62, 14]. To the best of the authors’
16
knowledge, there is no premixed flame experiment equivalent to those conducted in Ref. [62, 14] in
the literature.
Aspden et al. [10] recently performed a series of DNS of lean (φ = 0.4) premixed hydrogen flames
(LeH2 ≈ 0.3) at Karlovitz numbers ranging from 30 to 4200. Their results clearly show that the
turbulent flame structure varies significantly between the lowest and the largest Ka flames. In agree-
ment with the results for non-premixed flames, their largest Ka flame has a structure comparable to
that of a methane flame, i.e. the flame behaved as an effective unity Lewis number flame. Similarly,
they found that their high Karlovitz propane flame (LeC3 H8 ≈ 2)(distributed reaction zone regime)
also had a unity-Lewis-number-like structure [9]. These results suggest that a transition for these
“effective” Lewis numbers between their laminar values and unity may be identified. Peters first
suggested this concept of effective Lewis numbers [142]. However, no quantitative model has been
suggested yet.
In addition, there is no such information in the literature on the flame structure of larger hydro-
carbon turbulent premixed flames (LeC7 H16 = 2.8). Based on the results of Aspden et. al. [10, 9],
it is likely that differential diffusion will not be fully suppressed by turbulence in the target flames
considered in this thesis. This would be easily verified by comparing the unity with the non-unity
Lewis number target flames. Assessing if differential diffusion effects are still present in such flames
has important implications for modeling, as will be discussed in Section 1.7.
1.6 Reaction zone
The reaction zone of turbulent flames previously simulated has been shown to be affected by turbu-
lence. In particular, fluctuations in the fuel consumption rate have been identified [73, 35, 36, 71].
Understanding the physics behind these source term fluctuations is of high interest as they directly
affect the turbulent flame speed, which is one of the most important (and difficult to predict) global
characteristic of a premixed turbulent flame.
Such source term fluctuations in turbulent flames have commonly been attributed to stretching
effects, i.e. effects of curvature and strain rate [71, 22]. The idea comes from the well known
17
equation [133, 114]
SL = SL0 − L SL0 κ + at ,

(1.16)
where L is the Markstein length, κ is the mean curvature, at is the tangential strain rate, SL0 is
the unstretched laminar flame speed, and SL is the stretched laminar flame speed. The Markstein
length is known to be dependent on the unburnt mixture conditions, and in particular on the fuel
Lewis number [114]. In 1992, Haworth and Poinsot [73] investigated the effect of the fuel Lewis
number on the correlation between local consumption speed and tangential strain rate/curvature
in two-dimensional flames at the edge of the corrugated flamelets regime, using one-step chemistry.
They found that, for non-unity Lewis number flames, the local consumption speed is more correlated
with curvature than strain rate, while the opposite was found for the unity Lewis number flame.
Extending these results to high Karlovitz number turbulent flames, and, more specifically, to
flames at the edge of the broken reaction zones regime, is not straightforward. First, as turbulence
increases, turbulent eddies eventually penetrate the preheat zone [142, 151] and it was observed
that differential diffusion effects are suppressed [10, 142, 22] (as discussed for the flame structure in
Section 1.5). In addition, it is also important to note that Eq. 1.16 is only valid for small stretch [114],
i.e. L (κ + at /SL ) 1, which may very well not be the case at high Karlovitz numbers.
At high Karlovitz numbers, similar correlations have been investigated, but only in the mean
sense. Recently, using the leanest (φ = 0.31) series of hydrogen/air (LeH2 ≈ 0.3) flames presented in
Ref. [10], Amato et al. [5] showed that, in terms of averaged quantities, a correlation between local
consumption speed and curvature is maintained up to very high Karlovitz numbers (up to 3400).
In a series of two-dimensional DNS of lean (φ = 0.7) methane/air (LeCH4 ≈ 1) turbulent premixed
flames, Hawkes and Chen [71] showed that (in the mean sense) a correlation between local flame
consumption speed and tangential strain rate could be found up to high Karlovitz numbers (Ka ≈
300). Unfortunately, there is no such information available in the literature for fuels with larger
than unity Lewis numbers at high Karlovitz number.
Furthermore, source term fluctuations leading to local extinctions have not been previously stud-
ied using DNS data. The only local extinctions identified are for lean hydrogen/air flames which are
18
thermodiffusively unstable [113]. Under laminar conditions, cells (characterized by positive curva-
ture) of intensified burning rate are formed, while being surrounded by regions of local extinctions
(characterized by high negative curvature) [53]. These extinctions tend to disappear with increasing
Karlovitz number and are not caused by turbulence [10, 11].
Characterization of the reaction zone of the target flames and comparison with the results (ob-
tained for other flames) discussed above will provide valuable information. In particular, it could be
determined if different modeling approaches are needed or if the existing ones are sufficient (mored
1.7 Modeling
Direct numerical simulations of turbulent premixed flames or even non-reacting turbulent flows are
limited to small turbulent Reynolds numbers only. This is due to the inherent cost of resolving a
wide range of scales [152]. A popular alternative approach is the use of Large Eddy Simulations
(LES), which consider only the scales larger then a specified cut-off. The smaller scales are assumed
to be universal. More details on the extent of the validity of this assumption and how to properly
identify this cut-off can be found in Ref. [152]. The governing equations are therefore filtered and
closure for the non-resolved scales needs to be provided. These are called Sub-Grid-Scale (SGS)
models. SGS models are required for both non-reacting and reacting flows.
A particular challenge for the development of SGS models in the context of turbulent flames is
the number of closures required (N species) and the complexity of the terms that require closure.
For instance, each of the species source term is, in general, a function of all the other species
mass fractions. In practice, LES of turbulent premixed flames is done with reduced chemistry
models [165, 63, 64, 98].
Chemistry reduction can be done by the use of reduced chemical mechanisms, bringing the
number of transported species down to a tractable number [29], or by tabulation, reducing the
number of transported scalar variables to one or two [171, 65]. Variants of the first method are
described in Section 1.4 (challenge # 2). However, these methods assume a separation between
19
the chemical and the diffusive or convective time scales. Particularly in high Karlovitz flames, not
all chemical timescales are smaller than the flow timescales. For this reason, the detailed chemical
mechanisms can only be reduced down to relatively large mechanisms with these techniques.
In order to further reduce the cost associated with the chemistry, Gicquel et al. [65] proposed
a Flame Prolongation of ILDM (FPI), i.e. the chemistry is represented by the ILDM at high
temperatures (where the chemical timescales are small) and by flamelets at lower temperatures. van
Oijen et al. [171] have then proposed to represent the chemistry by a Flamelet-Generated Manifold
(FGM). This type of manifold “is not only based on chemical assumptions, but also takes the most
important transport processes into account” [171]. With the FGM approach, only a few controlling
variables (typically one or two) are transported and closure is provided through the FGM (table
with as many entries as the number of controlling variables). The FGM is generated a priori by
computing one or a series of one-dimensional flame solutions, depending on the number of controlling
variables.
FGM methods are computationally very attractive for obvious reasons. However, the validity
of tabulated chemistry lies on an important hypothesis: the turbulent flame structure and the
chemical source terms are assumed to be similar to that of laminar flamelets. Single flamelets [63]
and stretched (strained or curved) flamelets [98] have been considered to generated the FGM for
LES simulations of premixed flames. These were shown to provide good accuracy at low Karlovitz
numbers, namely in the wrinkled/corrugated flamet regime, and at the lower edge of the thin reaction
zone regime. Based on the discussion provided in the previous two sections, it is not clear if the
FGM approach should be valid at Karlovitz numbers corresponding to the flames considered in this
thesis. This highlights the importance of characterizing the flame structure and the reaction zone
(especially the relation with strain rate and curvature) of these flames.
Another challenge for tabulated chemistry is the choice of Lewis numbers. As discussed in
Section 1.5, for non-premixed turbulent flames, tabulation is usually done with unity Lewis number
flamelets, even when the species have Lewis numbers far from unity [45, 144, 149]. However, for
premixed turbulent flames, chemistry tabulation is generally done by computing flamelets with full
20
transport [98] or constant “laminar” Lewis numbers [171]. There is clearly an inconsistency between
what is done for premixed and non-premixed chemistry tabulation, hence the need for an adequate
effective Lewis number model. While this may have limited impact on the simulation of methane/air
flames, it is crucial for large hydrocarbon flames which exhibit important differential diffusion effects.
1.8 Objectives and outline
In view of the above discussion, the objectives of the current thesis are as follows:
1. to propose a time-integration method specifically-designed for the coupled, highly non-linear,
PDEs governing the evolution of unsteady reacting flows such as highly turbulent flames;
2. to perform a series of high Karlovitz n-heptane/air premixed flames in the upper portion of
the thin reaction zone regime;
3. to determine if the flame structure of these turbulent premixed flames can be mapped onto that
of flamelets and to provide a model for the effective Lewis numbers that should be considered
in these flamelets;
4. to assess the effects of turbulence and differential diffusion on the reaction zone characteristics,
the source term fluctuations, and the turbulent flame speed;
5. to provide a modeling approach for the source term fluctuations and to propose a tabulated
chemistry method for large-hydrocarbon flames in the thin reaction zone regime.
The governing equations considered and the numerical solver used are presented in Chapter 2.
The new time-integration method is described and validated in Chapter 3. In Chapter 4, the config-
uration and the parameters of the turbulent flame simulations are presented, along with qualitative
results. The flame structure is characterized and modeled in Chapter 5. The reaction zone is charac-
terized in Chapter 6 and a modeling approach is discussed in Chapter 7. Conclusions and suggestions
for future work are presented in Chapter 8.

21
Chapter 2
Governing equations and numerical

solver
In this chapter, the governing equations considered are presented first. The numerical solver, with the
exception of the modifications made on the time-integration scheme (decribed in the next chapter),
is then detailed.
2.1 Governing equations
The equations governing the unsteady evolution of the chemically reacting flows considered are
described in the following.
2.1.1 Fluid mechanics
The reacting mixture is assumed to contain a total number of N species and their chemistry is
assumed to be given by a chemical kinetics mechanism involving K reactions, with forward and
backward reactions counted separately. The chemically reacting flows of interest in the current
study are of relatively low Mach number (Ma ), typically below 0.3 [55, 162]. Under this condi-
tion, the acoustic waves can be ignored and the pressure field can be decomposed into a spatially-
invariant, but (potentially) time-dependent component, P0 (t), and a fluctuating hydrodynamic pres-
22
sure, p (x, t) [55, 162, 180, 142], with
p (x, t)
= O Ma2 .

(2.1)
P0 (t)
Since the focus of this thesis is on turbulence-chemistry interaction, Soret and Dufour effects,
body forces, and radiative heat transfer are ignored [128, 95, 162, 56]. While non-negligible Soret
and Dufour effects have been observed for specific hydrogen/air flames [59], limited effects have been
observed in laminar stretched n-heptane/air flames [181]. In particular, these effects where found to
be of opposite signs in lean vs. rich mixtures, and are expected to be smaller with near stoichiometric
mixtures. The assumption that Soret/Dufour effects can be negleceted for the flames considered in
this thesis will be tested in Section 4.2.2. In addition, the species molecular diffusion is assumed to
be described by the Fickian law [128, 95, 162, 180, 142]. Under these assumptions, the evolution
of the system is governed by the following conservation equations of mass, momentum, energy, and
species density [182, 180, 142]:
∂ρ
+ ∇ · (ρu) = 0 (2.2)
∂t
∂ρu
+ ∇ · (ρu ⊗ u) = −∇p + ∇ · τ (2.3)
∂t

∂ρT X α
cp + ∇ · (ρuT ) = ∇ · (ρcp α∇T ) + cp,i ρ ∇Yi + Yi Vc,i · ∇T + ω̇T (2.4)
∂t i
Lei

∂ρYi α
+ ∇ · (ρuYi ) = ∇ · ρ ∇Yi + ∇ · (ρYi Vc,i ) + ω̇i . (2.5)
∂t Lei
In the above equations, ρ is the density, u is the velocity vector, T denotes the temperature of the
mixture, and Yi is the mass fraction of species i. In the momentum equation (Eq. 2.3), τ is the
deviatoric stress tensor, defined as
2
τ = µ ∇u + (∇u)T − µ(∇ · u)I,

(2.6)
3
where I is the identity matrix and µ is the fluid viscosity. In the energy conservation equation
23
(Eq. 2.4), ω̇T includes heat source terms due to chemical reactions, α is the thermal diffusivity, and
cp is the specific heat at constant pressure of the mixture, given by
N
X
cp = Yi cp,i , (2.7)
i=1
where cp,i is the specific heat at constant pressure of species i. In the species conservation equations
(Eq. 2.5), ω̇i is the chemical source term of species i, and Lei is the Lewis number of species i,
defined as
α
Lei = , (2.8)
Di
with Di the mass diffusivity for species i. The correction velocity Vc,i in Eq. 2.5 accounts for
gradients in the mixture molecular weight as well as ensures zero net diffusion flux. It has the
following expression [180, 142]:
   
N N
α ∇W X ∇Y j ∇W X Yj 
Vc,i = − α −α  , (2.9)
Lei W j=1
Lej W j=1
Lej
where  −1
N
X Yj 
W = (2.10)
j=1
W j
is the local mean molecular weight of the mixture, and Wj is the molecular weight of species j.
The above set of equations is complemented by the equation of thermodynamic state
P0 W
ρ= , (2.11)
RT
b
where P0 is the thermodynamic pressure (see Eq. 2.1) and R

b is the universal gas constant.
24
2.1.2 Chemical model
The overall rate of change of species i, ω̇i , in Eq. 2.5 can be split into a production term, ω̇i+ , and a
consumption term, ω̇i− , as
ω̇i = ω̇i+ − ω̇i− . (2.12)
It is important to note that both the production term ω̇i+ and the consumption term ω̇i− are positive.
The production rate of species i, ω̇i+ , is given by the sum of the contributions from all elementary
chemical reactions leading to the formation of this species:
K
" N ν #
X Y ρYs js
ω̇i+ = Wi kj , (2.13)
j=1 s=1
Ws
νji >0
where νjs is the stoichiometric coefficient of species s in reaction j. In the above expression, the rate
constant of reaction j, kj , is given by the Arrhenius form, kj (T ) = Aj T bj exp−Ta,j /T , where Ta,j is
the activation temperature of this reaction. Similarly, the consumption rate of species i, ω̇i− , is given
by the sum of the contributions from all elementary chemical reactions leading to the destruction of
this species:
K
" N ν #
X Y ρYs js
ω̇i− = Wi kj . (2.14)
j=1 s=1
Ws
νji <0
The local heat release rate is given by
N
X
ω̇T = − hi ω̇i , (2.15)
j=1
where
Z T
hi = h0i + cp,i dT (2.16)
T0
is the specific enthalpy of species i, and h0i denotes its value under standard and reference conditions.
25
2.2 Numerical algorithm
In this section, a description of the flow solver and the numerical algorithms used is provided.
2.2.1 Overview of the numerical solver
The simulations in this work are performed using the structured, multi-physics and multi-scale finite-
difference code NGA [55]. The NGA code allows for accurate, robust, and flexible simulations of both
laminar and turbulent reactive flows in complex geometries and has been applied in a wide range
of test problems, including laminar and turbulent flows [182, 183, 184] and constant and variable
density flows [55, 31, 173], as well as Large-Eddy Simulations (LES) [183, 122] and Direct Numerical
Simulations (DNS) [17, 173, 32]. This numerical solver has been shown to conserve discretely mass,
momentum, and kinetic energy, with arbitrarily high order spatial discretization [55].
NGA uses both spatially and temporally staggered variables [55]. All scalar quantities
(ρ, p, T, Yi , Di , α, cp , cp,i , µ) are stored at the volume centers, and the velocity components are stored
at their respective volume faces. The convective term in the species transport equations is discretized
using the bounded quadratic upwind biased interpolative convective scheme (BQUICK) [74], and
the diffusive term is discretized using a second-order centered scheme. The variables are advanced
in time using the second-order semi-implicit Crank-Nicolson scheme of Pierce and Moin [145].
An iterative procedure is applied to fully cover the non-linearities in the Navier-Stokes equations.
This iterative procedure has been found to be of critical importance for stability and accuracy
considerations [55, 162, 145]. The numerical algorithmic sequence for one time step is described
below, where a uniform time step ∆t is employed. The density, pressure, and scalar fields are
advanced from time level tn+1/2 to tn+3/2 , and the velocity fields are advanced from time level tn to
tn+1 . A total number of Q sub-iterations is assumed.
0. Upon convergence of the previous time step, the density, ρn+1/2 , pressure, pn+1/2 , velocity
fields, un , and scalar fields, Yn+1/2 , are stored, where Y represents the vector of species mass
fractions (Y1 , ..., YN ). The solutions for pressure, species mass fractions, and momentum (from
26
the previous time step) are used as initial best guesses for the forthcoming iterative procedure:
n+3/2 n+3/2 n+1 n

p0 = pn+1/2 , Y0 = Yn+1/2 , and (ρu)0 = (ρu) , (2.17)
where the subscript indicates the index of the sub-iteration. The Adams-Bashforth prediction
is used for the initial density evaluation:
n+3/2
ρ0 = 2ρn+1/2 − ρn−1/2 . (2.18)
This ensures that the continuity equation is discretely satisfied at the beginning of the iterative
n+3/2
procedure. The vector of chemical source terms is denoted by Ω = (ω̇1 , ..., ω̇N ), and Ω0 is
evaluated using the thermochemical quantities obtained at the conclusion of the previous time
step (explicit prediction).
For the sub-iteration k = 1, . . . , Q
1. The scalar fields are advanced in time using the semi-implicit Crank-Nicolson method [55, 145]
for the convective and diffusive terms, and explicit integration for the chemical source terms:
n+3/2
Yn+1/2 + Yk
Y∗k = , (2.19)
2
n+3/2 n+3/2
ρn+1/2 Yn+1/2 + ∆t Cn+1 + Dn+1 · Y∗k + Ω∗k

ρk Yk+1 = k k
n+1
∆t ∂C ∂D n+3/2 n+3/2

+ + · Yk+1 − Yk . (2.20)
2 ∂Y ∂Y k
To simplify the discrete notations for spatial differential operators, the operators corresponding
to the convective and diffusive terms in the scalar equations (Eq. 2.5) are written as C and
∂C ∂D
D, respectively. ∂Y and ∂Y are the Jacobian matrices corresponding to the convective and
∂C
diffusive terms, respectively. C and ∂Y are functions of the density and the velocity, while D
∂D
and ∂Y are functions of the density and the kinematic viscosity. They are consistently updated
27
at each sub-iteration. Depending on the order of discretization, these operators are generally
banded diagonal matrices (e.g. tridiagonal for 2nd order discretization and pentadiagonal
for 3rd order discretization). It is important to note that the semi-implicit Crank-Nicolson
method proposed by Pierce and Moin [145] is not applied to the time-integration of the species
chemical source terms, Ω∗k . As mentioned in the introduction, this is due to the extremely high
∂Ω n+1

computational cost associated with the calculation of the chemical Jacobian matrix, ∂Y k ,
and the even more expensive inversion of this matrix.
The temperature equation is advanced in time in a similar fashion:
n+3/2
T n+1/2 + Tk
Tk∗ = , (2.21)
2
n+3/2 n+3/2
= ρn+1/2 T n+1/2 + ∆t Ckn+1 + Dkn+1 · Tk∗ + Ω∗T,k + ST,k
∗

ρk Tk+1
n+1
∆t ∂C ∂D n+3/2 n+3/2

+ + · Tk+1 − Tk , (2.22)
2 ∂T ∂T k
P cp,i
∗ ρα α
with ST,k being the discretized form of cp ∇cp ·∇T + i cp ρ Lei ∇Yi + Yi Vc,i ·∇T evaluated
at half time step. Note that the temperature equation is treated the exact same way as the
∗
species transport equations, with the exception of the additional source term ST,k .
2. The density field is predicted from thermodynamics using
N n+3/2
−1
P Yi,k+1
P0 Wi
n+3/2 i=1
ρk+1 = . (2.23)
b n+3/2
RTk+1
It is important to note that this density evaluation does not ensure conservation of the species
densities, ρYi , since no density rescaling such as the one proposed by Shunn et al. [162] is used.
However, upon convergence of the sub-iterations, this formulation is equivalent to the density
treatment proposed by Shunn et al.
3. The momentum equation is advanced in time using a similar semi-implicit Crank-Nicolson

28
method as for the scalar fields
un + un+1
u∗k = k
, (2.24)
2
n+3/2
ρn+1/2 + ρk+1 ρn−1/2 + ρn+1/2 n
b n+1
u k+1 = u
2 2
h i
n+1/2 n+1/2 n+3/2
+ ∆t Cu,k + Du,k · u∗k + ∇pk
n+1/2
∆t ∂Cu ∂Du
+ + · u b n+1
k+1 − un+1
k , (2.25)
2 ∂u ∂u k
where Cu and Du are discrete operators associated with the convective and the viscous terms,
respectively. u
b is the predicted velocity field used to compute the fluctuating hydrodynamic
pressure (Step 4).
4. A Poisson equation is then solved for the fluctuating hydrodynamic pressure:
n+3/2 n+3/2
" ! #
n+3/2 1 ρn+1/2 + ρk+1 ρk+1 − ρn+1/2
∇ 2
δpk+1 = ∇· b n+1
u k+1 + . (2.26)
∆t 2 ∆t
The Poisson equation is solved using the high-fidelity HYPRE package [55, 61]. The predicted
velocity field is then updated through a projection step:
2∆t
n+3/2

n+3/2 n+3/2 n+3/2
un+1 b n+1
k+1 = u k+1 − n+3/2
∇δpk and pk+1 = pk + δpk+1 . (2.27)
ρn+1/2 + ρk+1
5. Upon convergence of the sub-iterations, the new solutions are updated:
n+3/2 n+3/2 n+3/2 n+3/2

ρn+3/2 = ρQ , pn+3/2 = pQ , un+1 = un+1
Q , and Y = YQ . (2.28)
It is important to note that the above formulation becomes equivalent to the fully-implicit Crank-
Nicolson time-integration scheme upon convergence of the sub-iterations [145].

29
2.2.2 Extension to multi-dimensions
For clarity, the extension to multi-dimensions is presented here only for the species equations
(Eq. 2.20). An identical procedure is applied to the temperature (Eq. 2.22) and the momentum
equations (Eq. 2.25).
For simpler implementation, the set of equations (Eq. 2.20) is solved in practice in its residual
form:
" n+1 #
n+3/2 ∆t ∂C ∂D n+3/2 n+3/2

ρk I − + · Yk+1 − Yk
2 ∂Y ∂Y k
n+3/2 n+3/2
n+1/2 n+1/2
Cn+1 + Dn+1 · Y∗k + Ω∗k .

= ρ Y − ρk Yk + ∆t k k (2.29)
The above equation is equivalent to
n+3/2 n+3/2
Yk+1 = Yk − ∆tJ−1 · Θk , (2.30)
where the matrix J is defined as
n+1
n+3/2 ∆t ∂C ∂D
J = ρk I− + , (2.31)
2 ∂Y ∂Y k
and the vector
n+3/2 n+3/2
ρk Yk − ρn+1/2 Yn+1/2 n+1
− Ck + Dn+1 · Y∗k + Ω∗k

Θk = k (2.32)
∆t
is the error (residual) made on the species transport equation at the previous sub-iteration. When
the sub-iterations are fully-converged, the residual, Θk , is zero.
In multi-dimensional numerical simulations, inverting the system of discrete equations presented
can be of high computational cost, despite its sparse nature. The method of Approximate Factor-
ization (AF) is therefore used to convert the single, multi-dimensional problem into multiple, one-
∂C ∂D
dimensional problems that can be solved efficiently [55, 145]. The transport operator F = ∂Y + ∂Y
30
can be split exactly into directional transport operators Fx , Fy , and Fz , leading to
n+3/2 ∆t n+1
J = ρk I− F
2 k
n+3/2 ∆t n+1 ∆t n+1 ∆t n+1
= ρk I− F − F − F (2.33)
2 x,k 2 y,k 2 z,k
in a general three-dimensional orthogonal coordinate system. Accordingly, the following factorization
is used:
n+3/2 ∆t n+1 ∆t n+1 ∆t n+1

ρk I− F − F − F
2 x,k 2 y,k 2 z,k
n+3/2
= ρk I
!
∆t 1
· I− · Fn+1
x,k
2 ρn+3/2
k
!
∆t 1
· I− · Fn+1
y,k
2 ρn+3/2
k
!
∆t 1
· Fn+1 + O ∆t2 .

· I− z,k (2.34)
2 ρn+3/2
k
The AF does not degrade the temporal accuracy of the time-integration scheme, since it introduces
a second-order error in time, the same order as the one introduced by the temporal discretization
(Eqs. 2.19 and 2.20).
2.3 Application to turbulent reacting flows
As previously mentionned, NGA is tailored for the simulation of low Mach number variable density
turbulent flows. In particular, it is optimized for massively parallel computations of such flows.
These characteristics are necessary for the simulation of turbulent reacting flows such as turbulent
flames. However, the solver as presented in this section integrates the scalar source terms in an
explicit fashion. As mentionned in the introduction, the scalar transport equations in reacting flows
are stiff and an explicit time-integration would make the simulation of such flows impracticable.
Therefore, the time-integration scheme has to be adapted to treat the stiff scalar source terms in
31
some type of implicit fashion. This is the focus of the next chapter.
32
Chapter 3
Semi-implicit time-integration for

stiff chemistry1
A semi-implicit preconditioned iterative method is proposed for the time-integration of the stiff
chemistry in simulations of unsteady reacting flows, such as turbulent flames, using detailed chemical
kinetic mechanisms. Emphasis is placed on the simultaneous treatment of convection, diffusion, and
chemistry, without using operator splitting techniques. As such, a diagonal-preconditioned iterative
(to account for the non-linearity of the system) method for the efficient integration of stiff chemistry
in the numerical simulation of unsteady chemically reacting flows is proposed in a multi-dimensional
setting.
The implementation of the method is presented first, followed by a description of the numerical
test cases. Then, a detailed theoretical analysis of the stability and the accuracy of the scheme
is provided, along with a numerical validation using the test cases. Finally, the extension of the
method to other chemically reacting flows, its advantages over other methods, and its limitations
are discussed.
1 The work presented in this chapter is published in B. Savard, Y. Xuan, B. Bobbitt, and G. Blanquart, J. Comput.
Phys. (2015). Yuan Xuan has contributed to the formulation of the scheme in the form of a preconditioner. He has also
acted as a second advisor during the publication process. Brock Bobbitt has performed most of the one-dimensional
flame simulations with NGA found in this chapter (along with the corresponding figures). The author of this thesis
has contributed to the rest of the work presented in this chapter, i.e. the formulation of the approximate factorization
with the diagonal preconditioner, the comparison of the method to the diagonal preconditioners previously found
in the literature, the 3D turbulent flame, the theoretical analysis, the comparison with the numerical results, the
extension of the results to other canonical flows, and the discussion.
33
3.1 Species vs. temperature equation stiffness
As discussed in the introduction (Section 1.4), in chemically reacting flows, the very large mag-
nitude (or equivalently small timescales) of the source terms are responsible for the stiffness of
the scalar transport equations [112, 85]. This stiffness is generally believed to be due to the stiff
source terms in the species transport equations, but not due to the temperature transport equa-
tion [66, 112, 101]. An estimate for the temperature time scale in a n-C7 H16 /air premixed flame
(test case presented in Section 3.4.1) gives τT ∼ (Tb − Tu ) / (ω̇T / (ρcp ))max ∼ 10−4 s, where Tb and
Tu are the burnt and the unburnt temperatures, respectively. This time scale is approximately
an order of magnitude larger than the time step corresponding to a unity convective CFL in such
a laminar flame. Therefore, the temperature equation (Eq. 2.4) is advanced in time in the exact
same fashion as described in the previous chapter (Eqs. 2.21 and 2.22) without any further implicit
treatment. Cases for which the temperature time scale may not be considered large are discussed in
Section 3.6.3.
Consistently, improvement of the numerical procedure detailed in the previous chapter is based
on modifying the time-marching step for the species mass fraction fields only (step 1 in the procedure
described in Section 2.2.1). All other intermediate steps are left unchanged.
3.2 Preconditioned iterative method
The time-marching for species transport equations described in Eq. 2.30 resembles the standard
preconditioned Richardson-type iterative method [154], where the matrix J acts as a preconditioner.
More precisely, the choice of the preconditioner, J, can be arbitrary and does not modify the discrete
form of the equations to solve (i.e. Θk = 0). It only changes the convergence characteristics of the
iterative method. For instance, setting

n+3/2
J = ρk I (3.1)
is equivalent to the fully-explicit integration of the convective, diffusive, and chemical source terms
34
in the species transport equations, while setting
n+1
n+3/2 ∆t ∂C ∂D ∂Ω
J = ρk I− + + (3.2)
2 ∂Y ∂Y ∂Y k
corresponds to the full-implicit integration of the convective, diffusive, and chemical source terms in
the species transport equations.
Clearly, there is a trade-off in the choice of the preconditioner. Since it is applied at each step
of the iterative method, it is preferable to have a preconditioning matrix, J, with low computing
and inversion cost. The cheapest preconditioner would therefore be the one described by Eq. 3.1
(fully-explicit integration), which leads to poor convergence performance requiring extremely small
time step sizes. On the other hand, the optimal preconditioner would be the one leading to the
fully-implicit integration of the various terms (Eq. 3.2). Unfortunately, since the chemical source
terms of most species are generally dependent on a large number of other species, the chemical
∂Ω
Jacobian matrix, ∂Y , is usually not sparse [135]. Therefore, its construction and inversion may
become prohibitively expensive especially when a large number of species is considered.
Note that preconditing methods have been used in the literature for other types of problems.
An exemple is the inversion of the nonlinear systems resulting from higher-order implicit Runge-
Kutta steps, in which the stiffness is not specifically due to the chemistry. A wide range of iterative
preconditiong methods have been proposed for such systems, including diagonal preconditioners [2,
83, 89, 167].
3.3 Semi-implicit preconditioning for stiff chemistry
3.3.1 Proposed preconditioner
In an attempt to achieve better convergence characteristics while keeping the form of the precon-
ditioner as simple as possible, a preconditioning method for stiff chemistry, which lies between the
fully-explicit and fully-implicit extremes, is proposed.

35
The proposed preconditioner is written as
n+1
n+3/2 ∆t ∂C ∂D
J = ρk I− + −Λ , (3.3)
2 ∂Y ∂Y k
where Λ is a diagonal matrix defined as
ω̇i−
Λi,i = . (3.4)
Yi
The matrix Λ may be regarded as a very good approximation of the negative of the diagonal of the
chemical Jacobian:
∂ ω̇i ∂ ω̇i+ ∂ ω̇i− ω̇ −
= − ≈0− i (3.5)
∂Yi ∂Yi ∂Yi Yi
as the production rate of species i (Eq. 2.13) is not a function of the species mass fraction and its
consumption rate (Eq. 2.14) is often linear in Yi . The ith element of Λ represents an approximation of
the inverse of the timescale corresponding to the chemical consumption of species i (approximation
of the inverse of the consumption characteristic times as used in Ref. [85, 66]). The proposed
preconditioner aims to suppress the small timescales due to the fast consumption of the different
species in the system with stiff chemistry.
As the matrix Λ is diagonal, the preconditioner corresponds to a species-wise relaxation, similar
to the Jacobi preconditioner. The proposed method accounts for the non-linearities by coupling the
transport equations through the iterative procedure (Eq. 2.32).
The proposed method is inspired by the work of Eberhardt and Imlay [57], who first introduced
a diagonal preconditioning in a point-implicit algorithm for the simulation of steady-state reacting
flows. The diagonal elements were found by computing the L2 -norm of the corresponding row of the
chemical Jacobian. However, this was found to lead to an improper approximation of the chemical
time scales and resulted in a lack of elemental conservation in time-marching algorithms [85, 28, 27].
In an effort to improve the accuracy of the method, Ju [85] replaced the jth element of the diagonal
preconditioner by the maximum between the inverse of the consumption characteristic time of species
36
j and the inverse of the production time of elementary reaction in which species j is the product.
It was argued that this type of preconditioning is suited for the simulation of steady flows, but
should fail to provide time-accurate solutions of unsteady flows [85]. An implicit correction with the
diagonal approximation of the Jacobian would introduce errors that would accumulate over time.
In the present algorithm, the diagonal preconditioning is applied within an iterative procedure
for each time step. This iterative method allows further reduction of the residuals and this is the
reason why the method is suitable for the simulation of unsteady reacting flows.
3.3.2 Extension to multi-dimensions
The proposed preconditioning is compatible with the method of Approximate Factorization described
in Section 2.2.2. Equation 2.33 becomes
n+3/2 ∆t n+1
J = ρk I− (F − Λ)k
2
n+3/2 ∆t n+1 ∆t n+1 ∆t n+1 ∆t n+1
= ρk I− F − F − F + Λ . (3.6)
2 x,k 2 y,k 2 z,k 2 k
Accordingly, the following factorization is proposed:
n+3/2 ∆t n+1 ∆t n+1 ∆t n+1 ∆t n+1

ρk I− F − F − F + Λ
2 x,k 2 y,k 2 z,k 2 k

n+3/2 ∆t n+1
= ρk I+ Λ
2 k
−1 !
∆t n+3/2 ∆t n+1 n+1
· I− ρk I+ Λ · Fx,k
2 2 k
−1 !
∆t n+3/2 ∆t n+1
· I− ρk I+ Λ · Fn+1
y,k
2 2 k
−1 !
∆t n+3/2 ∆t n+1 n+1
· Fz,k + O ∆t2 .

· I− ρk I+ Λk (3.7)
2 2
Similar to Section 2.2.2, this proposed factorization does not degrade the temporal accuracy of the
preconditioned time-integration scheme, since it introduces a second-order error in time, the same
order as the one introduced by the temporal discretization (Eqs. 2.19 and 2.20).
37
n+3/2 ∆t n+1
In the above factorization, the inversion of ρk I+ 2 Λk is computationally trivial since it
is a diagonal matrix. As such, three simpler one-dimensional inversion problems in the x, y, and z
directions can be solved sequentially using tridiagonal (2nd order spatial discretization schemes) or
pentadiagonal (3rd order spatial discretization schemes) matrix inversion algorithms analytically, in
a serial or parallel fashion. This is very important, as it keeps the overall cost of any sub-iteration
linear with the number of grid points and linear with the number of species.
3.3.3 Summary
A semi-implicit preconditioning strategy is proposed, in combination with an iterative method, for
the time-integration of the stiff chemistry. The proposed method takes advantage of the iterative
structure of the NGA code and, more precisely, its semi-implicit formulation of the transport terms.
In addition, the method is compatible with the approximate factorization (see Section 2.2.2), which
is necessary to maintain the performance of the code. Note that this preconditioning method could
be applied as well within any other iterative algorithm.
The proposed semi-implicit preconditioning method is based on an approximation of the diagonal
of the chemical Jacobian. The hypothesis, which will be tested in Section 3.5, is that the smallest
chemical timescales are well approximated by this diagonal preconditioner, allowing the use of larger
integration time step sizes. Another assumption that will be tested in Section 3.5 is that a good
convergence rate of the sub-iterations can be obtained with sufficiently large time step sizes. This
would allow the total number of operations per time step to be very similar to that for the explicit
time-integration of the chemistry. As mentioned in the introduction, other diagonal preconditioners
were found to be associated with a lack of robustness and elemental conservation in non-iterative
time-marching algorithms [28, 27] or were argued to lack time-accuracy [85]. However, the proposed
method is iterative and, upon convergence of the sub-iterations, the fully-implicit formulation is
recovered. Therefore, any issues would be alleviated by the convergence of the sub-iterations.
38
3.4 Test cases
The performance of the proposed iterative method will be tested in Section 3.5 on two flow con-
figurations: a one-dimensional unstretched premixed flame and a three-dimensional high Karlovitz
turbulent premixed flame.
3.4.1 One-dimensional premixed flame
The one-dimensional laminar unstretched premixed flame is selected as the first test case since it is
the most canonical configuration (that includes convection, diffusion, and chemistry) and it is well
suited for the quantitative evaluation of the stability and accuracy of a numerical scheme [128, 66, 95].
The condition of the present test case corresponds to a n-heptane/air flame with an equivalence ratio
of 0.9 and an unburnt temperature and pressure of 298 K and 1 atm, respectively. N -heptane is
used in this study as a representative of heavy hydrocarbons of relevance to transportation fuels.
A reduced finite-rate chemistry model is used in the present work. The mechanism developed in
Ref. [17] was reduced from 47 species and 290 reactions to 35 species and 217 reactions in an
effort to alleviate the computational cost. Since the gas mixture is slightly lean, species that are
only produced under rich conditions (and their associated reactions), namely C5 H5 , C5 H6 , and all
aromatic species (benzene, naphthalene...) were removed from the mechanism. As the focus of the
present chapter is placed on the time-integration, the Lewis numbers of all the species are set to unity
(i.e. no differential diffusion) for the present simulation. Under these conditions, the laminar flame
speed is SL = 29 cm/s, and the flame thickness is lF = 0.43 mm, with lF = (Tb − Tu )/(∂T /∂x)max .
It is expected that using different fuels, mixture compositions, and unburnt conditions will lead to
qualitatively similar results. This is discussed in more detail in Section 3.6.
The computational domain is initialized with a fully-converged solution of a stationary flame.
The flame front is initially located near the center of the domain (x0 = 5.5 mm) to reduce effects
of the boundaries. The left boundary is set to be a wall, and the right boundary is an open flow.
Once the simulation is started, the flame front moves towards the unburnt gases (left of the domain)
at a speed which corresponds to the laminar flame speed. The length of the 1D computational
39
2000
1500 SL
T [K]
1000
500
0
0 2 4 6 8 10 12
x [mm]
Figure 3.1: Schematic of the computational domain and initial condition.
domain is approximately 30 times the laminar flame thickness. The domain is discretized with a
uniform grid cell spacing (∆x = 18 µm) except behind the initial flame location (x > 5.5 mm) where
it is stretched with a factor of 1.1 (ratio of the grid cell size to its neighbor). A schematic of the
configuration is shown in Fig. 3.1.
While arbitrarily high order (for the convective and viscous terms) is available in the NGA code,
the simulations throughout this thesis rely on second-order spatial discretization of the viscous
and convective terms of the Navier-Stokes equations. Grid convergence tests were performed and
revealed a 2nd order accuracy in space (not shown). This spatial order of accuracy was found to be
independent of the proposed time-integration scheme. These tests also determined that 24 points
across the flame front (lF ) is sufficient to achieve satisfactory grid independence. In the following,
this grid resolution is used for all numerical tests.
3.4.2 Three-dimensional turbulent premixed flame
The second configuration chosen corresponds to the unity Lewis number flame presented in details
in Chapter 4 (other simulations with differential diffusion will be presented/analyzed in Chapters 4
to 7). A brief description is provided in this section. Figure 4.1 presents a schematic of the flow
configuration. The left and right ends of the domain correspond to an inflow of unburnt gases and
an outflow of burnt gases, respectively. The position of the flame is statistically steady as an inflow
velocity equal to the turbulent flame speed is used. Turbulent forcing is employed to avoid a fast
decay of turbulence due to viscous dissipation [31]. This forcing is not used at the inlet and the outlet
to avoid negative velocities. Forcing starts at 0.5L after the inlet and is switched off at a distance of
40
forcing forcing
begins flame ends
inflow 0 outflow
0 0.5L 8L 11L
Figure 3.2: Schematic of the three-dimensional turbulent premixed flame configuration.
3L from the end of the domain, with L being the domain width and height (Fig. 4.1). This distance
is found sufficient for the turbulence to decay without forcing such that no negative axial velocities
are found at the outlet. Note that the decay of turbulence is due to viscous dissipation.
The chemical mechanism and the flame unburnt conditions are the same as for the one-dimensional
premixed flame detailed in Section 3.4.1. The Karlovitz number is 280 and the turbulent Reynolds
number in the unburnt gases is 190. Other important parameters of the simulation are listed in
Table 5.1. Note that prior to this simulation, the flow field is established without any flame; homo-
geneous isotropic turbulence is obtained in the forced region. Then, the turbulent premixed flame
is simulated with finite-rate chemistry for several eddy turnover times. The simulation is performed
in parallel using 1920 processors on the cluster Hopper at the National Energy Research Scientific
Computing Center (NERSC). Fig. 3.3 provides visual information about the flame simulated.
The simulation is performed with a time step size of ∆t = 5.7 × 10−7 s, which corresponds
to a convective CFL condition of 0.8. With the proposed semi-implicit scheme, the stiffness of the
chemical model was found not to impact the stability of the turbulent reacting flow simulation. More
details on the stability of the numerical framework and the choice of time step size are provided in
the following section.
3.5 Results
First, a theoretical analysis on the convergence of the sub-iterations for the species transport equa-
tions is presented in this section. Second, this analysis is further discussed in light of the eigenvalue
content of the proposed precondition matrix. Third, the convergence of the sub-iterations is evalu-
41
Figure 3.3: Contours of temperature on a two-dimensional slice of the three-dimensional turbulent

premixed flame.
ated numerically and compared to the theoretical analysis. Fourth, the performance of the proposed
method in terms of stability is presented both theoretically and numerically, using the test cases
previously introduced. Fifth, since stability does not imply accuracy, the numerical accuracy of
the proposed method is presented and its dependence on the time step size and the number of
sub-iterations is discussed. Sixth, the performance of the proposed method in terms of elemental
conservation is presented. In particular, the effects of the iterative procedure (and the number of
sub-iterations used) are assessed. Finally, the computational efficiency of the method is discussed.
3.5.1 Theoretical analysis
To simplify the analysis, it is assumed transport is integrated explicitly (i.e. not modified by the
sub-iterations). For the sub-iteration k + 1 within a single iteration, Eq. 2.30 and 2.32 take the form
Jk · (Yk+1 − Yk ) = ρ0 Y0 − ρk Yk + ∆t (C + D) · Y0 + ∆tΩ∗k , (3.8)
where
∆t
Jk = ρ k I + Λk . (3.9)
2
The superscripts (n and n + 1) have been dropped for clarity. The subscript 0 corresponds to
42
the final solution of the previous iteration. The terms in the equation can be reorganized as

∆t 1
I+ Λ · (Yk+1 − Yk ) = − Yk (3.10)
2 ρ k
1
+ [ρ0 Y0 + ∆t (C + D) · Y0 ]
ρk
∆t ∗
+ Ω .
ρk k

1 1 ∗ 1
Expanding ρΛ , ρΩ , and ρk around Y0 gives
k k

∆t1
I+ Λ · (Yk+1 − Yk ) = −Yk (3.11)
2 ρ 0

1 ∂ 1
+ + · (Yk − Y0 ) [ρ0 Y0 + ∆t (C + D) · Y0 + ∆tΩ0 ]
ρ0 ∂Y ρ 0

∆t 1 ∂Ω
· (Yk − Y0 ) + O |Yk − Y0 |2 .

+
2 ρ ∂Y 0
Subtracting Eq. 3.11 evaluated at two consecutive sub-iterations, and neglecting the second-order
terms yields

∆t 1
I+ Λ · (Yk+1 − 2Yk + Yk−1 ) = − (Yk − Yk−1 ) (3.12)
2 ρ 0

∂ 1
+ (ρY + ∆t (C + D) · Y + ∆tΩ) ⊗ · (Yk − Yk−1 )
∂Y ρ 0

∆t 1 ∂Ω
+ · (Yk − Yk−1 ) .
2 ρ ∂Y 0
A simpler expression reads
Yk+1 − Yk = A0 · (Yk − Yk−1 ) , (3.13)
with
−1
∆t 1 ∆t 1 ∂Ω
A0 = I − I + Λ I− − T0 , (3.14)
2 ρ 0 2 ρ ∂Y 0
43
where

∂ 1
T0 = (ρY + ∆t (C + D) · Y + ∆tΩ) ⊗ . (3.15)
∂Y ρ 0
The convergence of the sub-iterations is assured as long as the absolute values of all eigenvalues
of A0 are less than unity, i.e. the spectral radius of A0 is less than unity. The opposite implies a
divergence (in the linear sense) of the sub-iterations which is likely to be associated with an unstable
simulation. Without surprise, in the limit of ∆t → 0, the spectral radius of A0 goes to zero and the
convergence of the sub-iterations is ensured. For practical time step sizes, large eigenvalues of A0

can be due to the large magnitude of the chemical Jacobian ρ1 ∂Y ∂Ω
or the matrix T0 . However,
0
it can be easily shown that the only eigenvalue of T0 is a ratio of densities (density of the explicit
prediction vs. initial density) which is of order one and is negligible compared to the large eigenvalues
of the chemical Jacobian (more details in Section 3.5.2). Consequently, the matrix T0 is neglected
in the present theoretical analysis. This simplification will be further justified by the numerical tests
in Section 3.5.4.2. Therefore, the rest of the theoretical analysis will consider the following matrix:
−1
∆t 1 ∆t 1 ∂Ω
A00 =I− I+ Λ I− , (3.16)
2 ρ 0 2 ρ ∂Y 0
which is only a function of the full chemical Jacobian, and the approximate (diagonal) Jacobian.
The convergence of the sub-iterations is then assured as long as the spectral radius of A00 is less
than unity.
Alternatively, one can consider the residuals in the relative species mass fractions instead of the
residuals in their absolute values. These relative residuals are evaluated as follows:

00
Yrel rel rel rel
k+1 − Y k = A0 · Y k − Y k−1 , (3.17)
where Yrel = G−1 Y, with G = diag (Y0,1 , . . . , Y0,N ). Y0,i is the mass fraction of species i obtained
at the end of the previous iteration. The matrix A000 reads

44
−1
∆t 1 ∆t −1 1 ∂Ω
A000 = G−1 A00 G = I − I + Λ I− G G . (3.18)
2 ρ 0 2 ρ ∂Y 0
Note that both the absolute and the relative value system of equations are analytically equivalent.
In particular, the eigenvalues of A000 are identical to those of A00 .
3.5.2 Eigenvalue analysis
It is clear from Eq. 3.16 that if the approximation of the diagonal of the chemical Jacobian (Eq. 3.4),
∂Ω
further referred to as the precondition matrix, were to be equal to the full chemical Jacobian, ∂Y ,
the scheme would be, in the linear sense, unconditionally stable (recall that T0 = 0 is assumed).
In all other cases, the first pertinent analysis to justify the choice of the preconditioner (Eq. 3.3)
is to evaluate the eigenvalue content of the precondition matrix and compare it to that of the full
chemical Jacobian. To ensure a good convergence rate of the sub-iterations, the eigenvalue content
of the precondition matrix should be close to that of the full Jacobian. This assumption is tested in
the following.
∂Ω
More specifically, the eigenvalues of the chemical Jacobian ∂Y correspond to the inverse of the
chemical timescales (τ ) present in the system. These timescales are associated with the rate of
change of the species (or a combination of species) mass fractions in the system in the absence of
transport. The idea of using a preconditioner is to allow the use of a time step larger than the
smallest of these timescales. Consequently, given a time step ∆t, it is important that the chemical
timescales smaller than ∆t be well represented by the precondition matrix, such that the source
terms associated with these timescales are properly integrated over ∆t. Note that the eigenvalues of
the proposed diagonal precondition matrix correspond to an approximation of the reciprocal species
lifetimes.
The species lifetimes obtained for the precondition matrix (Eq. 3.4) and the chemical timescales
obtained with the full Jacobian are compared in Fig. 3.4. These are evaluated with the mixture
composition and the temperature (T = 1615 K) corresponding to the peak rate of heat release in a
one-dimensional flame similar to that presented in Section 3.4.1. For this a priori analysis, the one-
45
dimensional flame is computed with FlameMaster [146]. Note that, a posteriori, virtually identical
results were obtained when the NGA solution was considered (Section 3.4.1). This is not surprising,
as the precondition matrix and the Jacobian only depend on the local mixture composition and
temperature, which should not be dependent on the solver used. It is interesting to associate species
to each of the chemical timescales shown in Fig. 3.4. However, it is important to note that not every
chemical timescale corresponds to a species lifetime. The timescales are ordered from the smallest
to the largest. They are then compared, entry by entry, between the preconditioned and the full
Jacobian.
10−4
10−5
τ from diagonal Jacobian (s)
n-C3H7
10−6
1-CH2
10−7
10−8
2-C7H15
10−9
1-C5H11
10−10
10−10 10−9 10−8 10−7 10−6 10−5 10−4
τ from full Jacobian (s)
Figure 3.4: Comparison of the chemical timescale (τ ) of the full chemical Jacobian to the species
lifetime of the preconditioned chemical Jacobian.
It is well known that because of elemental conservation, nk of the eigenvalues of the full Jacobian
are zero, with nk the number of elements in the chemical system [112]. Only one of them is zero for
the precondition matrix, i.e. the one associated with N2 , as its source term is identically zero. This
is to be expected from the definition of the precondition matrix. Therefore, only the 31 smallest
timescales are shown in Fig. 3.4.
Although Fig. 3.4 does not provide direct information on the stability limit, it can be observed
that all the timescales smaller than about 10−5 s are well approximated by the precondition matrix.
In other words, the diagonal matrix represents accurately the smallest chemical timescales in the
46
system. As mentioned previously, the proposed method is similar to the Jacobi method, which is
guaranteed to converge in diagonal-dominant problems (but it can also converge in other cases). It

is therefore interesting to assess if the matrix I − ∆t
2
1 ∂Ω
ρ ∂Y (Eq. 3.16) is diagonal dominant for
0
−5
∆t < 10 s. Figure 3.5(a) shows that the matrix is clearly not diagonal dominant for ∆t = 5 × 10−6

−1 1 ∂Ω
s. However, the corresponding normalized matrix introduced in Eq. 3.18, I − ∆t 2 G ρ ∂Y G is
0
close to be diagonal dominant for the same ∆t, as shown in Fig. 3.5(b), i.e. all the terms of each
row are smaller in magnitude than the corresponding term on the diagonal. Since both of these
matrices have the same eigenvalue content and the same terms on the diagonal, it becomes clear
why the precondition matrix approximates adequately the smaller timescales of the full Jacobian.
35 1000 35 1
30 30
100
25 25
0.2
20 10 20
j
15 1 15
10 10 0.04
0.1
5 5
0 0.01 0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
i i

∆t 1 ∂Ω ∆t −1 1 ∂Ω
(a) I − 2 ρ ∂Y
(b) I − 2
G ρ ∂Y
G
0 0
Figure 3.5: Relative magnitude of the elements of each row compared to the element on the respective
diagonals. ∆t = 5 × 10−6 s.
The results from Fig. 3.5(b) suggest that the Jacobi method may be successfully applied to such
a matrix (it is guaranteed to converge if the matrix is diagonal dominant). This will be verified in
the following sections as the stability and the convergence rate of the method is analyzed considering
the spectral radius of A00 (or, equivalently, A000 ).
3.5.3 Convergence of the sub-iterations
The one-dimensional flame test case is used to evaluate numerically the convergence of the sub-
iterations. The maximum density residual over the whole domain is considered as its convergence is
47
controlled by the convergence of all the chemical species. Figure 3.6 displays the residual through
two complete time steps for four different time step sizes. As the time step size decreases, the
residuals decrease faster. The rate of convergence of the sub-iterations is observed to follow an
exponential relationship, i.e. Resk ∼ rk with r the convergence rate. The numerical convergence
rate r is computed by fitting an exponential curve to the density residuals. Since density is an
analytical function of the species mass fractions, its convergence rate should tend towards that of
the slowest converging species mass fraction. In other words, this convergence rate should be close
to the spectral radius of A00 (Eq. 3.16).
0 0
10 10
∆t = 5.5 · 10 s
−6 ∆t = 2.0 · 10−6 s
∆t = 5.0 · 10−6 s ∆t = 2.0 · 10−7 s
Residual [kg/m3 ]
Residual [kg/m3 ]
−2 −5
10 Simulation unstable 10
−4
10 −10
10
−6
10 −15
10
1 2 1 2
Time Step Time Step
(a) At the stability limit (b) Far from the stability limit
Figure 3.6: Evolution of the density residual as a function of sub-iterations over two time steps for the
proposed semi-implicit time-integration scheme. The dashed lines correspond to fitted exponential
curves averaged over several time steps.
The numerical and theoretical (spectral radius) convergence rates are compared in Fig. 3.7. The
numerical convergence rates are in relatively good agreement with the theoretical values. This further
justifies neglecting the variable density matrix, T0 , in the present theoretical analysis (Section 3.5.1).
Using the spectral radius (i.e. the largest eigenvalue) as a measure of the convergence rate is a worst
case scenario, as the projection of the density residuals on the associated eigenvector might be
identically zero (to machine precision). This might explain partially better numerical convergence
rates than theoretically predicted.

48
1
theoretical
0.8 numerical
Convergence rate
0.6
0.4
0.2
0
105 106 107
1/∆t
Figure 3.7: Rate of convergence of the sub-iterations: comparison between theoretical (largest spec-
tral radius of A00 ) and numerical (rate of convergence of the largest density residuals) results for the
one-dimensional flame.
3.5.4 Stability
The stability of the preconditioned iterative method relies on the stability of the sub-iterations.
By stability, it is meant that the solution remains bounded in time. Note that stability does not
require the sub-iterations to be fully converged, nor does it imply accuracy. Due to the high non-
linearity and the complexity of the governing equations, the stability condition for the proposed
preconditioned iterative method and its relation to sub-iteration convergence rate are investigated
theoretically (simplified system) and numerically in the present section.
3.5.4.1 Theoretical stability
Assuming the residuals of the species mass fractions at each sub-iteration remain sufficiently small
for the linear analysis presented in Section 3.5.1 to be valid, then the sub-iterations do not diverge if
the spectral radius of A00 (Eq. 3.16), is less than 1. In other words, if this spectral radius is less than
1, stability is ensured, independently of the number of sub-iterations used. However, the accuracy
of the solution will be affected by the number of sub-iterations (discussed in Section 3.5.5).
The spectral radius of A00 is plotted as a function of temperature in Fig. 3.8(a) for different values
49

1
of ∆t. The matrices ρΛ and ρ1 ∂Y
∂Ω
are evaluated from the one-dimensional flame solution,
0 0
computed with FlameMaster [146]. Note that the maximum of each curve corresponds to a point in
Fig. 3.7. The spectral radius is shown to be a strong function of temperature. It is not surprising that
the chemical system is more sensitive to perturbations at higher temperatures (where the Arrhenius
rate constants are larger). It is important to remember that the stability limit in a reacting flow
also depends on the nature of the transport terms, which have been ignored for the present analysis,
as well as the coupling between the scalar transport and the Navier-Stokes equations. In particular,
the most unstable location may not systematically occur at the highest temperature in the domain.
For comparison, the matrix corresponding to A00 using an explicit time-integration of the chemical
source term is introduced:

∆t 1 ∂Ω
A00,exp = . (3.19)
2 ρ ∂Y 0
The spectral radius of A00,exp is plotted against temperature in Fig. 3.8(b) for different values of
∆t. The (theoretical) stability limit for the explicit scheme is ∆t = 5.6 × 10−11 s, whereas it is
∆t = 6.1 × 10−6 s for the semi-implicit scheme. Under the present conditions, it can be clearly
observed that the proposed method has the potential to increase the stability limit by several orders
of magnitude.
1.6 1.6
∆t = 10−8 s
Spectral radius of A00,exp
1.4 1.4
Spectral radius of A00
10−7 s
1.2 10−6 s
1.2
1 −5 1
10 s
∆t = 10−11 s
0.8 0.8 10−10 s
0.6 0.6 10−9 s
0.4 0.4 10−8 s
0.2 0.2
0 0
500 1000 1500 2000 500 1000 1500 2000
Temperature (K) Temperature (K)
(a) Proposed semi-implicit (b) Explicit
Figure 3.8: Spectral radius of A00 and A00,exp as a function of temperature in the one-dimensional
premixed flame.
50
3.5.4.2 Numerical stability
The previous section provided a stability criterion through theoretical analysis of a system of reduced
complexity. In this section, the stability of the complete scheme is analyzed numerically. Note that,
while the theoretical analysis was done assuming explicit transport, the test cases are performed with
semi-implicit transport (as described in Section 2.2), which is used in turbulent flame simulations.
It is important to demonstrate the performance of the proposed preconditioning method within the
algorithmic setting used in practical simulations. For the present tests, the same stability limits
were found for the transport terms treated explicitly and implicitly. This is not surprising because
the convective CFL number is less than unity for all test cases. It also confirms that the transport
terms may be neglected in the previous theoretical analysis (Section 3.5.1).
The 1D test case is considered first. For all time step sizes tested, it was found that converging
(as opposed to converged) sub-iterations implied a stable simulation. In other words, unless the
sub-iterations diverge, the simulation remains stable. As shown in Fig. 3.6(a), the largest time
step size that can be used for the simulation to be stable is found to be ∆t = 5 × 10−6 (with
∆t = 5.5 × 10−6 leading to unstable results). This value is very close to the theoretical stability
limit of ∆t = 6.1 × 10−6 s (Section 3.5.4.1). Note that the largest numerically stable time step size
using an explicit time-integration of the chemical source term is ∆t = 2 × 10−10 s, also close to the
theoretical stability limit (∆t = 5.6 × 10−11 s). These stability limits are summarized in Table 3.1.
Numerical Theoretical
Proposed scheme 5 × 10−6 s 6.1 × 10−6 s
Explicit scheme 2 × 10−10 s 5.6 × 10−11 s
Table 3.1: Largest stable time step size for the proposed semi-implicit scheme and the explicit time-
integration of the chemical source terms for the 1D flame test case. Numerical and theoretical results
(see Section 3.5.4.1) are compared.
These results suggest that the stability limit can be well approximated by the theory, i.e. the
maximum ∆t such that the spectral radius of A00 (Eq. 3.16) is less than unity. As such, this should
also correspond to the numerical stability limit for the 3D turbulent premixed flame. However,
∆t = 5 × 10−6 s corresponds to a convective CFL number of approximately 7, which is too large for
51
the overall spatio-temporal scheme to be stable. In other words, for the turbulent case, the largest
stable time step size is constrained not by the time-integration of the chemical source terms but by
the convective CFL condition. This allowed the simulation to be performed with a convective CFL
of 0.8 (see Section 3.4.2), which corresponds to ∆t = 5.7 × 10−7 s.
To understand why such a large time step size could be used for the turbulent flame simulation,
the spectral radius of A00 (Eq. 3.15) is plotted against temperature throughout the whole domain
of the 3D simulation in Fig. 3.9. For reference, the one-dimensional equivalent is added on top of
the scatter plot. This plot suggests that the stability of the chemical system alone (Eq. 3.8) is only
slightly altered by turbulence.
0.6
3D DNS
0.5 1D flame
0.4
0.3
0.2
0.1
0
500 1000 1500 2000
Temperature (K)
Figure 3.9: Scatter plot of the spectral radius of A00 , with ∆t = 5.7 × 10−7 s, as a function of
temperature in the three-dimensional turbulent flame. The one-dimensional profile is added for
comparison.
3.5.4.3 Summary
It was shown in the present section that the stability limit of the proposed method is well approx-
imated by the largest ∆t such that the spectral radius of A00 is less than 1. This stability limit is
independent of the number of sub-iterations used. On the other hand, the number of sub-iterations
does affect accuracy, and it will be discussed in the next section.
An important result is the fact that the stability of the three-dimensional turbulent flame was
52
constrained by the convective CFL limit, rather the time-integration of the chemical source terms.
This means that the proposed method has a great potential, in terms of stability, for such flow
simulations. Theoretical estimates of the stability limits for other types of flows, fuels, conditions,
and chemical mechanisms will be discussed in Section 3.6.1.
3.5.5 Accuracy
In the previous subsection, we established the stability limit(s) of the proposed scheme. We now
investigate the accuracy for a given stable simulation.
3.5.5.1 Order of accuracy
The order of accuracy (i.e. the power-dependence of the error as the time step size is reduced) of
the proposed method is determined for the 1D (freely propagating) flame test case. In the absence
of chemistry, the algorithm used in NGA can be formally shown to be second-order accurate if two
or more sub-iterations are used [145]. In practice, four sub-iterations are typically used (to improve
stability and to achieve adequate accuracy of the fractional-step) [55, 31, 173]. For the present test
case, four sub-iterations are also used to evaluate the order of accuracy. The impact of the number
of sub-iterations on the absolute magnitude of these errors is discussed in the next sub-section.
Simulations with different time steps are performed and results are presented in Fig. 3.10. The
errors for various quantities are evaluated as the absolute difference of their integrated value in
temperature space compared with a reference solution obtained with ∆t = 2 × 10−8 s. The species
n-C7 H16 , OH, CO, and H2 O are chosen as representatives of reactants, radicals, intermediates, and
products. All of these quantities are found to demonstrate second-order accuracy in time, as shown
in Fig. 3.10. It is interesting to note that the expected order of accuracy is already recovered with
only four sub-iterations.

53
0
10
Normalized Error
−2
10
x−2
−4
10
H2 O
−6
10 OH
CO
density
−8 n−C7 H16
10 5 6 7
10 10 10
1/∆t
Figure 3.10: Temporal accuracy of the method as a function of the time step size for the one-
dimensional, propagating flame. The errors for the various species mass fractions are evaluated as
the absolute difference of their integrated value in temperature space compared with a reference
solution obtained with ∆t = 2 × 10−8 s.
3.5.5.2 Magnitude of errors
It is important to distinguish order of accuracy (as the time step size goes to zero) and absolute
magnitude of errors. In order to illustrate the quality of the solution using various time step sizes,
Table 3.2 compares the laminar flame speeds, Fig. 3.11(a) presents the temperature profiles in
physical space (shifted to coincide at T = 400 K), and Fig. 3.11(b) shows the intermediate species n-
C3 H7 mass fraction vs. temperature profiles. The first two quantities are chosen, as they correspond
to the two most important quantities associated with a laminar flame. The n-C3 H7 mass fraction
vs. temperature profile is chosen, as it is the quantity the most sensitive to the time step size (small
reciprocal lifetime where its mass fraction is non-zero). It is clear that, up to a time step size of
∆t = 2 × 10−6 s, errors are negligible even with only four sub-iterations (0.3% error in laminar flame
speed and virtually no difference in the temperature and species profiles). At the stability limit
(∆t = 5 × 10−6 s), the solution is deteriorated when only four sub-iterations are used. However,
using a large number of sub-iterations, the solution reaches similar level of accuracy.
These results demonstrate that, for time step sizes smaller than ∆t = 2 × 10−6 s, sufficiently
accurate solutions for the 1D flame are obtained with as little as four sub-iterations. For time step
sizes between ∆t = 2 × 10−6 s and the stability limit, more sub-iterations are needed to reach
sufficient accuracy. This is a direct consequence of the decreasing convergence rates with increasing
54
∆t (s) Q ρmax (A00 ) SL (cm/s)

2 × 10−7 4 0.42 28.65
2 × 10−6 4 0.65 28.57
5 × 10−6 4 0.92 26.93
5 × 10−6 200 0.92 28.64
Table 3.2: Laminar flame speed obtained from simulations with various time step sizes and number
of sub-iterations. ρmax (A00 ) is the theoretical maximum spectral radius of A00 in the flame.
−6
x 10
20
2000
15
1500
Yn−C3 H7
10
T [K]
1000
5
∆t = 2.0 · 10−7 s ∆t = 2.0 · 10−7 s
500 ∆t = 2.0 · 10−6 s ∆t = 2.0 · 10−6 s
0
∆t = 5.0 · 10−6 s ∆t = 5.0 · 10−6 s
∆t = 5.0 · 10−6 s; Q = 200 ∆t = 5.0 · 10−6 s; Q = 200
0 −5
−5 0 5 10 15 500 1000 1500 2000
x [m] x 10
−4 T [K]
(a) Temperature vs. distance (b) n-C3 H7 mass fraction vs. temperature
Figure 3.11: Impact of time step size and number of sub-iteration on the accuracy of 1D propagating
flames. When not mentioned, four sub-iterations are used (Q = 4).
time step sizes, as presented in Fig. 3.7. There should exist a pair of time step size/number of
sub-iterations such that performance is optimized for a given level of accuracy. However, since the
method is meant to be used with turbulent flames, and not 1D flames, such optimization would be
of limited interest since the largest time step size allowed by the convective CFL condition in the
turbulent flame is smaller than ∆t = 2 × 10−6 s. Because the spectral radius in the turbulent flame
is similar to that of the 1D flame (see Fig. 3.9), four sub-iterations should be sufficient to obtain
accurate solutions. This is further tested below.
As mentioned in the previous section, the turbulent premixed flame simulation was performed
with a convective CFL number of 0.8 (∆t = 5.7 × 10−7 s). In order to evaluate the quality of the
solution, a comparative simulation was performed with a smaller time step size, ∆t = 8 × 10−8 s
(four sub-iterations are used for both simulations). Note that this comparative simulation was also
run until statistically steady state was reached. While the time step size varies by a factor of 7
55
between the two simulations, virtually no differences could be identified between the simulations.
Figure 3.12 shows a representative comparison of joint probability density functions of the species
mass fraction vs. temperature. This type of joint probability density functions is used to evaluate
the impact of turbulence on the chemistry (Chapter 5) and it is important to make sure that any
deviation away from a laminar flame is not due to numerical artifacts. The results obtained using
both time step sizes are virtually identical, which is consistent with the observations made with the
one-dimensional flame test case.
×10−3 ×10−3
8 0.3 8 0.3
< Y |T > < Y |T >
6 6
0.2 0.2
YC 2 H 4
YC2 H4
4 4
0.1 0.1
2 2
0 0 0 0
500 1000 1500 2000 500 1000 1500 2000
T (K) T (K)
(a) ∆t = 5.7 × 10−7 s (CFL= 0.8) (b) ∆t = 8 × 10−8 s
Figure 3.12: Comparison between the joint probability density function of the C2 H4 mass fractions
vs. temperature obtained with different time step sizes.
3.5.6 Mass conservation
As it can be noted from the presentation of the method in Section 2.2, the sum of the species mass
fraction is not implicitly recovered. In other words, with the proposed diagonal preconditioner, the
sum of mass fraction is not guaranteed to remain equal to unity. While element conservation is
not ensured with the proposed scheme using only one sub-iteration, in practice, with the number
of sub-iterations used, elemental mass fractions were found to be adequately conserved. This can
be observed in Fig. 3.13 for the 3D turbulent premixed flame, at a CFL of 0.8 with only four sub-
iterations being used. Under the assumption of unity Lewis number transport, the elemental mass
56
fractions (YC , YH , YO ) should remain perfectly at their inlet values. Any deviations are evidence of
mass conservation errors. The maximum deviations in the domain are only about 1% of the inlet
values. These maximum deviations are found to occur in the oxidation layer of the turbulent flame.
10−1
YH
YC
Maximum relative error YO
10−2
10−3
0 1 2 3 4 5 6
Simulation time (ms)
Figure 3.13: Maximum deviation in the domain from the inlet elemental mass fractions vs. simulation
time. The results shown are for the three-dimensional turbulent flame with ∆t = 5.7 × 10−7 s and
four sub-iterations.
Using more sub-iterations leads to better elemental conservation, as shown in Fig. 3.14 for the 1D
flame test case with ∆t = 2 × 10−6 s. In particular, the rate of convergence of these elemental mass
fraction residuals follows the theoretical spectral radius of 0.65. Note that each of these simulations
were performed with a different number of sub-iterations Q until a constant propagation speed (flame
speed) and a constant flame structure were reached.
3.5.7 Computational efficiency
As presented in Section 2.2, the cost per sub-iteration with the proposed semi-implicit scheme is
virtually identical to that using an explicit time-integration of the chemical source terms. In addition,
with the proposed scheme, a number of four sub-iterations was found to be sufficient in practice to
achieve adequate accuracy and elemental conservation. As previously mentioned, this number of
sub-iterations is typically used for the simulation of non-reacting flows [55, 31, 173] and reacting
57
−2
10
YH
YC
Maximum relative error

YO
−4
10 0.65Q
−6
10
−8
10
0 10 20 30
Q
Figure 3.14: Maximum deviation in the domain from the inlet elemental mass fractions vs. number
of sub-iterations, Q. The results shown are for the one-dimensional, propagating flame using ∆t =
2 × 10−6 s. The theoretical convergence rate is added for comparison.
flows with explicit time-integration. Therefore, the cost per iteration with the proposed scheme is
(in practice) similar to that using an explicit time-integration of the chemical source terms.
Since the proposed scheme does not alter the cost per iteration compared to an explicit time-
integration of the chemical source terms, the increase in efficiency (speed-up) is equal to the increase
in largest stable time step size. For unsteady flames, the optimal time step size corresponds to
the convective CFL limit (the theoretical limit is unity, but 0.8 is the target within our numerical
framework). While the time step size in the 1D flame simulations was limited by the chemistry,
the convective CFL number was 0.2 (∆t = 2 × 10−6 s). This means that for turbulent flames with
umax /SL,b larger than 4, with umax the maximum velocity in the domain and SL,b the laminar
flame speed in the burnt gas (umax /SL,b = 1 for a 1D flame), the time step size will be limited
by the convective CFL limit. This was the case with the 3D turbulent flame test case in which
umax /SL,b ≈ 14. Consequently, the proposed time-integration method is optimally efficient for such
turbulent flames, in the context of the numerical framework of NGA.

58
3.6 Discussion
The proposed preconditioner was shown to exhibit very good performance for the test cases analyzed,
and in particular it allowed the use of a time step size limited by the convective CFL (3D turbulent
flame). In this section, in light of the results presented previously, the theoretical analysis of the
stability limit using the proposed scheme is extended to various flames, fuels, unburnt conditions,
and kinetic mechanisms. An ignition case is also considered. Additional validation of the numerical
stability is provided for a few selected cases. A quantification of the accuracy of the solutions
obtained with the proposed method is also presented for these cases. Then, the advantages of the
proposed preconditioner over alternative methods are highlighted. Subsequently, the limitations of
the present method are discussed. The objective of this section is to help one decide if the proposed
method is well suited for a specific unsteady reacting flow simulation.
3.6.1 Extension
The theoretical analysis presented in Section 3.5.1 is general and does not depend on the fuel, the
chemical mechanism, the flow configuration, or the unburnt conditions. In particular, the eigenvalue
analysis (Section 3.5.2) and the theoretical stability conditions (Section 3.5.4.1) can be applied to any
reacting flow. These analyses, which consider a dependence on the local mixture composition and the
temperature only, were further validated in a one-dimensional flame configuration (Section 3.5.4.2).
The results were also argued to be independent of the transport terms. Finally, it was shown that,
even in a highly turbulent three-dimensional flame, the departure from a one-dimensional flame
solution was not sufficient to significantly influence the stability (both theoretical and numerical).
In summary, the theoretical analysis (without transport) is sufficient to determine the stability and
the convergence rate of any laminar or turbulent reacting flow.
As such, the theoretical results are extended in this section by considering a wide range of one-
dimensional flame and zero-dimensional ignition solutions (each computed with FlameMaster [146]).
First, for premixed flames, the effects of the fuel, the unburnt conditions, and the chemical mech-
anism on the theoretical stability limits are investigated. Then, the analysis is performed for non-
59
premixed flamelets with different scalar dissipation rates and a 2D coflow diffusion flame. Finally, a
homogeneous ignition case at constant pressure is considered.
3.6.1.1 Premixed flames
The theoretical performance of the proposed semi-implicit method is tested on a series of one-
dimensional unstretched premixed flames.
First, the unburnt conditions are kept fixed and various fuels are considered. Four additional fuels
are tested: H2 combined with the chemical model presented in Ref. [77] (9 species 52 reactions),
CH4 with GRI-MECH 3.0 [163] (36 species 422 reactions), i-C8 H18 with CaltechMech v2.1 [18]
(171 species 1835 reactions), and n-C12 H26 , also with CaltechMech. The analysis performed in
Section 3.5.2 is repeated with these fuels. The species lifetimes at the location of the peak heat
release obtained from the semi-implicit precondition matrix (Eq. 3.4) are compared to the chemical
timescales obtained from the full chemical Jacobian in Fig. 3.15. Same as observed in Section 3.5.2,
for all fuels (and mechanisms) tested, the diagonal Jacobian (precondition matrix) approximates
very well almost all the timescales smaller than 10−5 s. Similarly, as presented in Table 3.3, the
stability limits using the proposed scheme are very close to the one found in Section 3.5.4.1.
Second, the same n-C7 H16 premixed flame with the unburnt conditions presented in Section 3.4.1
is computed using CaltechMech. Again, the stability limit (Table 3.3) is only marginally affected by
the chemical mechanism.
Third, a series of unburnt conditions are used for the n-C7 H16 premixed flame (using the mech-
anism introduced in Section 3.4.1). As the unburnt conditions encountered in practical combustion
devices correspond typically to higher temperature and pressure, both these quantities are increased
in this series of tests. Then, the equivalence ratio, which can significantly vary in a combustion
device, is modified in the test cases to cover a wide range centered around stoichiometry. Table 3.3
presents the theoretical stability limit for all these cases. Again, the largest (theoretically) stable
time step size varies only slightly throughout all cases.
An important conclusion can be drawn from the results shown in Table 3.3: the stability limit,
60
10−3
H2 [77]

CH4 [163]
10−5
n-C7 H16
i-C8 H18 [18]
−7
10
n-C12 H26 [18]
10−9
10−11
10−13
10−13 10−11 10−9 10−7 10−5 10−3
τ from full Jacobian (s)
Figure 3.15: Comparison of the chemical timescale (τ ) of the full chemical Jacobian to the species
lifetime of the preconditioned chemical Jacobian at the peak rate of heat release, considering various
fuels and chemical mechanisms.
using the proposed iterative semi-implicit preconditioning method, is only marginally sensitive to
the fuel, the unburnt condition, and the mechanism used. As mentioned in Section 3.5.7, the target
time step size is the convective CFL limit. An effective CFL number is therefore computed for
each of these 1D premixed flame (see Table 3.3), assuming 24 grid points per flame thickness (see
Section 3.4.1) are necessary for accurate simulation of a 1D laminar flame (umax = SL,b ). The
effective CFL number would be larger for turbulent flames and would increase with the turbulent
intensity. From these results, it is obvious that even for moderately turbulent flames, the time step
size would be restricted by the convective CFL, rather than the chemistry.
These results relate to the stability limit only, and do not suggest anything about accuracy. As
shown in Section 3.5.5, the accuracy is a function of the spectral radius, which is a function of
the time step size, and the number of sub-iterations used. Depending on the flow configuration
considered (turbulent, laminar, strained,...), four sub-iterations may be sufficient for the solution
to be accurate, even at the stability limit. However, the opposite is also possible. Under such
circumstance, either the time step size has to be decreased or the number of sub-iterations has to be
increased. This choice depends on the spectral radius vs. time step size profile (previously shown in
Fig. 3.7).
Figure 3.16 presents such profiles for flames corresponding to each of the chemical mechanisms
61
Fuel φ P0 Tu ∆tmax ∆tmax Equivalent

(atm) (K) (s) (s) convective
explicit semi-implicit CFL
H2 [77] 0.9 1 298 5.2 × 10−8 1.6 × 10−6 1.2
CH4 [163] 0.9 1 298 2.7 × 10−9 5.0 × 10−6 0.47
n-C7 H16 0.9 1 298 5.6 × 10−11 6.1 × 10−6 0.76
i-C8 H18 [18] 0.9 1 298 2.3 × 10−14 4.7 × 10−6 0.45
n-C12 H26 [18] 0.9 1 298 2.2× 10−14 4.6 × 10−6 0.55
n-C7 H16 [18] 0.9 1 298 4.5 × 10−14 4.6 × 10−6 0.58
n-C7 H16 0.9 1 400 4.3 × 10−11 5.2 × 10−6 0.89
n-C7 H16 0.9 1 600 2.7 × 10−11 3.8 × 10−6 1.1
n-C7 H16 0.9 2 298 5.2 × 10−11 6.2 × 10−6 1.3
n-C7 H16 0.9 10 298 5.0 × 10−11 6.5 × 10−6 3.1
n-C7 H16 0.7 1 298 2.1 × 10−10 9.5 × 10−6 0.58
n-C7 H16 1.1 1 298 3.9 × 10−11 4.3 × 10−6 0.62
n-C7 H16 1.3 1 298 7.4 × 10−11 4.3 × 10−6 0.43
Table 3.3: Theoretical largest stable time step size for the proposed semi-implicit scheme and the
explicit time-integration of the chemical source terms with various unstretched one-dimensional
premixed flames.
used in this section (and presented in Table 3.3). The first three mechanisms, although used for
different flame conditions, exhibit very similar profiles. This means that, with these mechanisms,
a moderate decrease in time step size from the stability limit translates in an appreciable decrease
in spectral radius. For the 1D test case analyzed in Section 3.4.1, a time step size three times
smaller than the theoretical stability limit was shown to provide sufficient accuracy with only four
sub-iterations.
Interestingly, the spectral radius profile exhibits a plateau just below the stability limit over a
wide range of time step size with CaltechMech. This means that, in order to obtain a minimal level
of convergence of the sub-iterations, either a very large number of sub-iterations or a very small time
step size would be needed. However, this is not what is observed numerically, as shown in Fig. 3.17,
since accurate solutions are obtained with time step sizes as large as half the numerical stability
limit, still with only four sub-iterations.
This result is better understood by considering the density residuals vs. sub-iterations for the
n-C7 H16 flame with CaltechMech, with ∆t = 2 × 10−6 s, presented in Fig. 3.18 (similar to Fig. 3.6).
Their convergence rate is found to be much closer to the third largest eigenvalue of A00 (0.689),
as opposed to its largest one (i.e. spectral radius), corresponding to 0.997. In other words, the
62
2.5
n-C7 H16
H2 [77]
2
CH4 [163]

n-C7 H16 [18]
1.5
n-C7 H16 [18]-modified
0.5
0
104 105 106 107 108 109 1010 1011 1012
1/∆t
Figure 3.16: Spectral radius of A00 vs. the inverse of the time step size for various cases presented in
Table 3.3. The red cross symbols correspond to the profile for the n-C7 H1 6/air flame with the stiff
reaction (Eq. 3.20) removed from CaltechMech.
projection of the species mass fractions residuals on the eigenvectors associated with the two largest
eigenvalues is negligible. This is consistent with the fact that these two largest eigenvalues are only
due to the presence in the chemical mechanism of the following fast reversible reaction:
A1 C2 H2 −C8 H7 A1 C2 H?3 −C8 H7 , (3.20)
which involves species that have negligible mass fractions in the flame simulated (these species are
soot precursors and should not be present in the lean flames considered in this chapter). The high
pressure limit rate constant (the only one available in the literature) was prescribed for this reaction.
Such rate constant is obviously too large, especially for the present atmospheric flames. The spectral
radius vs. time step size profile obtained when this reaction is removed from the chemical mechanism
is shown in Fig 3.16 (red cross symbols). The profile is virtually identical to the one obtained with
the 35-species mechanism introduced in Section 3.4.1. This means that the proposed method is also
efficient with a mechanism as large as CaltechMech, which is far larger than any other mechanism
used for the simulation of three-dimensional turbulent premixed flames [9, 10, 160, 176, 70, 53, 11].
63
−6
x 10
15
∆t = 2 · 10−9 s
∆t = 2 · 10−6 s
∆t = 3.9 · 10−6 s
10
Yn−C3 H7
5
−5
500 1000 1500 2000
T [K]
Figure 3.17: Profiles of n-C3 H7 mass fraction vs. temperature in the one-dimensional, n-C7 H16
propagating flame (similar to the test case of Section 3.4.1) with CaltechMech. The solutions from
using three different time step sizes, each with 4 sub-iterations, are compared.
3.6.1.2 Non-premixed flames
The theoretical stability limit using the proposed scheme is now evaluated for a series of (unity Lewis
number) non-premixed flamelets [136]. These one-dimensional flamelets correspond to solutions
close to the axis of symmetry of counter-flow diffusion flames and to local solutions close to the
stoichiometric isosurface of mixture fraction of turbulent flames.
First, two n-C7 H16 /air flamelets are considered: one with a small scalar dissipation rate (typically
found in laminar co-flow diffusion flames and turbulent diffusion flames, at moderate Reynolds
number) and one with a large scalar dissipation rate, corresponding to half the dissipation rate
leading to extinction. The results are shown in Table 3.4. Once again, the stability limits using
the proposed scheme are very similar to the values found for the series of premixed flames (previous
section). Additionally, the dissipation rate does not seem to have a strong effect on the stability of
the scheme.
Second, two C2 H4 /air flamelets are considered: again, one with a small scalar dissipation rate,
and one with a large dissipation rate. These two flamelets are used to estimate theoretically the
stability limit of a 2D-coflow diffusion flame. The 2D flame corresponds to an International Sooting
Flame Workshop target flame (more details in Ref. [88, 87]). The ethylene fuel (17.6% by mass)
64
10−4
10−6
Residual (kg/m3 )
10−8
0.689k
10−10
10−12
10−14
0.997k
10−16
1 2
Time Step
Figure 3.18: Evolution of the density residual as a function of sub-iterations over two time steps for
the n-C7 H16 flame with CaltechMech using the proposed time-integration scheme with ∆t = 2×10−6
s. The convergence rates given by the first (red) and the third (black) largest eigenvalues are shown
for comparison. k references to the sub-iteration index.
is diluted with nitrogen (82.4% by mass) (in both the flamelets and the 2D flame). The steady-
state solution for the temperature field is shown in Fig. 3.19. The two dissipation rates considered in
Table 3.4 correspond to the maximum and minimum values found in the 2D simulation. As expected,
the more restrictive time step size is encountered at the largest dissipation rate. In contrast, the
2D numerical simulation was found to be stable up to a time step size of ∆t = 4.0 × 10−6 s, which
is larger than the theoretical prediction of ∆t = 2.7 × 10−6 s. With this “practical” time step, the
maximum spectral radius of A00 found in the 2D domain (see Fig. 3.19) is about 0.97. This difference
can be partially explained by the fact that, in the region of largest dissipation rates (at the burner
exit), the flame is extinguished and does not compare well with a flamelet.
3.6.1.3 0D ignition
Although the proposed time-integration scheme was developed primarily for the simulation of multi-
dimensional turbulent flames, it could potentially be applied to the simulation of flows with ignition
events. In order to partially assess the potential of the method for such flows, a canonical 0D,
constant pressure ignition case is considered. The initial conditions as well as the theoretical stability
65
Fuel χst P0 Tf To ∆tmax ∆tmax

(1/s) (atm) (K) (K) (s) (s)
explicit semi-implicit
n-C7 H16 [17] 1 1 400 800 1.4 × 10−11 3.7 × 10−6
n-C7 H16 [17] 320 1 400 800 8.3 × 10−11 2.9 × 10−6
C2 H4 [17] 0.025 4 298 298 4.4 × 10−11 4.9 × 10−6
C2 H4 [17] 138 4 298 298 1.4 × 10−10 2.7 × 10−6
explicit time-integration of the chemical source terms with non-premixed flamelets. χst is the scalar
dissipation rate at stoichiometry, Tf the temperature on the fuel side, and To the temperature on
the oxidizer side. The oxidizer is air and the chemical mechanism considers 47 species and 290
reactions [17].
Figure 3.19: Contours of temperature (top) and spectral radius of A00 (bottom) from the two-
dimensional coflow laminar flame, obtained with a time step of 4.0 × 10−6 s.
limits (computed from a FlameMaster solution) are listed in Table 3.5. These conditions are meant
to be representative of ignition events in HCCI-like engines [186, 86]. All simulations are performed
with the CaltechMech mechanism.
Fuel φ P0 T0 ∆tmax ∆tmax

(atm) (K) (s) (s)
explicit semi-implicit
n-C7 H16 [18] 0.7 30 850 1.1× 10−12 3.7 × 10−7
explicit time-integration with a 0D isobaric ignition case. T0 is the initial temperature.
The same numerical simulation is performed with the present semi-implicit scheme in NGA to
evaluate the practical stability limit. The stability limit identified numerically is ∆t = 5.2 × 10−7
s, which is close to the theoretical limit. Unfortunately, at such large time step size, using four
sub-iterations, the solution is deteriorated: the ignition delay time is over-predicted and the burnt
66
temperature is under-predicted. This is also observed at the theoretical stability limit, as shown in
Table 3.6. Interestingly, with only four sub-iterations and using a time step size of ∆t = 2 × 10−7 s,
the solution is very close to the FlameMaster prediction, as seen in Table 3.6. While only a limited
analysis, the present results show the applicability of the proposed semi-implicit scheme for ignition
events.
Framework ∆t (s) tign (ms) Tb (K)

FlameMaster 1.560 2129
NGA 3.7× 10−7 1.703 2002
NGA 2.0× 10−7 1.566 2126
Table 3.6: Comparison of the ignition delay time tign and the burnt temperature obtained with
FlameMaster, and with the proposed framework using two different time step sizes. Four sub-
iterations are used.
3.6.2 Advantages over other methods
The performance of the proposed preconditioner is compared to that of the fully-implicit precondi-
tioner, operator-splitting methods, and stiffness removal through QSSA in the following.
3.6.2.1 Fully-implicit method
As mentioned in the introduction, the use of fully-implicit time-integration of the chemical source
terms is known to be prohibitively expensive for the simulation of turbulent reacting flows [66],
∂Ω
as the inversion of the full chemical Jacobian, ∂Y , at every point of the domain and at every
time step becomes very expensive when more than 10-20 species are considered. Another problem
that would arise using a fully-implicit preconditioner with the numerical framework presented in
Section 2.2 is that the extension to multi-dimensions using the approximate factorization introduced
in Section 2.2.2 could no longer be applied, as the fully-implicit chemical Jacobian is not a diagonal
matrix. Such a factorization is necessary for the efficiency of the overall procedure.
The major advantage of fully-implicit time-integration of the chemical source terms over the
proposed semi-implicit scheme is the use of a time step size not restricted by the chemistry. For
steady-state problems, this might be justified/useful. However, for turbulent reacting flows (such
67
as the one presented in Section 3.4.2), the characteristic hydrodynamic timescales of the turbulent
flow (relative to the turbulence and the chemistry) need to be resolved. These are often sufficiently
small that the cost increase for the fully-implicit method is not justified anymore.
3.6.2.2 Operator-splitting methods
The preconditioned iterative method integrates simultaneously the chemical, diffusive, and convec-
tive terms at the same effective time level. This guarantees that the numerical scheme used is free of
lagging errors. These errors are of particular importance in unsteady reacting flows, where chemistry,
diffusion, and convection are closely coupled, especially close to the thin flame fronts [128].
Using operator-split formulations, the chemical source terms are decoupled from the diffusive and
the convective terms in order to be integrated using stiff ODE solvers. Therefore, the application of
these methods for the simulation of reacting flow problems leads typically to integration accuracy
degradation [49, 67, 191]. This is demonstrated in Fig. 3.20 for the mass fraction of n-C3 H7 . Using
Godunov splitting, large numerical errors due to operator-splitting are observed with ∆t > 5 × 10−7
s (the ODE solver used is DVODE [25] with 10−8 and 10−20 for the relative and absolute tolerances,
respectively [116]). At this point, the numerical time step size surpasses the diffusion timescales.
In contrast, the proposed preconditioned iterative method does not suffer from these errors, since
the convection, diffusion, and chemistry are all integrated simultaneously. Note that, while Strang
splitting is known to perform generally better than Godunov splitting, its extension to a low Mach
number code based on spatial and temporal staggering is not trivial. This is the reason why Godunov
splitting, easily implementable on a staggered grid, was used for comparison.
Figure. 3.21 presents the computational cost per grid point per simulation time using 1) the
proposed preconditioned iterative method, 2) Godunov splitting, and 3) explicit time-integration of
the chemical source terms. For all cases, four sub-iterations are used. Without surprise, the cost
using small time step sizes (∆t < 5 × 10−8 s) is similar for all methods. For Godunov splitting, it
is computationally as cheap as the explicit method when the chemical source terms are not stiff,
which is the case at small time step sizes. However, using large time step sizes, the chemical source
68
−6
x 10
20 SI 2e-6 s
GS 2e-6 s
GS 5e-7 s
15 GS 1e-7 s
Yn−C3 H7
10
−5
500 1000 1500 2000
T [K]
Figure 3.20: n-C3 H7 mass fraction of from the 1D (freely propagating) flame solution. Solutions
from using Godunov splitting (GS) are compared, for different time step sizes, to the solution using
the proposed preconditioning method.
terms become stiff and the cost associated with solving the stiff ODEs increases. At large time step
sizes, the cost associated with Godunov splitting increases up to twice larger than that associated
with the proposed method. On the other hand, for the proposed preconditioned iterative method,
as mentioned earlier, the total number of operations per time step is virtually the same as that
associated with an explicit time-integration of the chemical source terms. In particular, the cost
per iteration does not vary with time step size. As such, the computational cost per grid point per
simulation time is proportional to the inverse of the time step size. In summary, for the present 1D
flame test case, with a time step of ∆t = 2 × 10−6 s, the computational cost associated with the
proposed method is smaller than that associated with Godunov splitting, while being free of lagging
errors (see Fig. 3.20).
In the simulation of turbulent flames, the chemical source terms are zero almost everywhere
(unburnt/burnt regions in a premixed flame; fuel/oxidizer streams in a non-premixed flame) except
at the flame front. This means that, if the domain is partitioned in the direction perpendicular
to the flame, the cost of a single time step, using an operator-splitting method, will vary between
the different partitions. Unfortunately, a partition cannot advance faster in time than the others.
Therefore, the computational time is dictated by the slowest partition. For the three-dimensional
simulation test case, this made the simulation impracticable using Godunov splitting. This could
69
7
10
Preconditioned iterative method
6 Godunov Splitting
10 Explicit
Cost [pt−1 ]
5
10
Stability limit
4
10
3
10
2
10 5 6 7 8 9 10
10 10 10 10 10 10
1/∆t
Figure 3.21: Computational cost of 1D stationary flame simulation for different chemical integration
methods. Cost is calculated as cpu time per point (s/pt), per second of simulation time (s). All
cases use four sub-iterations.
be partially alleviated by considering load balancing [75] at the cost of making the code more
complicated.
3.6.2.3 Stiffness removal through QSSA
As mentioned in the introduction, a way to remove the stiffness of the species transport equations
(and reduce the number of transport equations) is to put the species with small chemical timescales
in Quasi Steady State (QSS) [107]. It can be assumed that these species concentrations are entirely
controlled by non-QSS species. As such, an algebraic equation for their concentrations is obtained
and these QSS species are not transported in the simulation. This algebraic equation is used to
replace these QSS species in the reactions in which they are involved. More details can be found in
Ref. [140].
Application of this method is particularly interesting for compressible codes, for which the sta-
bility limit is controlled by either chemistry or acoustics [72]. The acoustics timescale is smaller
than the convective timescale (subsonic flows) and may be relatively close to the smallest chemical
timescales. As seen in Fig. 3.4, it is very likely that only a few species (and their associated reac-
tions) are responsible for the small chemical timescales. After removal of these species (and their
associated reactions), using QSSA, the stability of the solver would be limited by the acoustics only.
70
However, putting the species with the smallest associated timescales in quasi-steady state may
not always be justified. With the quasi steady state assumption, algebraic expressions can be
found for these species. In Fig. 3.22, these expressions are compared to their true values in the 3D
turbulent premixed flame (see Section 3.4.2). Several species typically placed in QSS in previous
studies [160, 108] are considered. It is obvious that the QSSA is valid for 1-CH2 , but not for n-C3 H7
nor 2-C7 H15 in the present turbulent premixed flame. This can be explained by the fact that the
timescales corresponding to these species, although very small at high temperature, are very large
at low temperatures, as can be seen in Fig. 3.23. A way to counter this behavior that has been
used in the literature [105] is to preheat the unburnt mixture to make the flame more “robust”, i.e.
to make sure that the species responsible for the stiffness of the system can be put in quasi-steady
state. However, this obviously modifies the nature of the flame simulated. Note that a recently
developed dynamic stiffness removal relies on local, rather than global, QSSA [110]. However, to
the best of the authors’ knowledge, the method has only been applied to the simulation of ignition
problems [186, 110].
3.6.3 Limitations
A first limitation of the method is that it is only efficient for the simulation of unsteady reacting
flows, in which the interplay between the flow field and the chemistry has to be captured through
adequate temporal resolution (small time step size). While one might decide to use the present
method to reach the solution of a steady-state problem, as shown in Section 3.6.1.2 for the 2D-
coflow diffusion flame, the associated cost would be large. In such a case, the use of a large time step
size is desirable to reach the time-independent solution. For all the examples provided, the largest
stable time step size is of the order of 1 × 10−6 s, which makes the method inefficient to reach a
steady-state flow solution.
Second, the method behaves poorly when a fast reversible reaction is present in the chemical
mechanism, since this leads to a spectral radius of A00 close to unity over a wide range of time step
sizes (as discussed in Section 3.6.1.1). Although such fast reaction was found to be unphysical in
71
×10−7 ×10−5
20 10
8
15
YC7 H15 - QSSA

YCH2 - QSSA 6
10
4
5
2
0 ×10−7 0 ×10−5
0 5 10 15 20 0 2 4 6 8 10
YCH2 YC7 H15
(a) 1-CH2 (b) 2-C7 H15
×10−5
8
6
YC3 H7 - QSSA
0 ×10−5
0 2 4 6 8
YC3 H7
(c) n-C3 H7
Figure 3.22: Scatter plots of species mass fractions computed from the algebraic expression assuming
QSS vs. their actual value in the three-dimensional turbulent flame. A straight line (y = x) is
expected for perfect QSS species.
CaltechMech (a better reaction rate should be implemented), it is not clear if it is always the case.
More importantly, for very large mechanisms such as those developed at the Lawrence-Livermore
National Laboratories (LLNL) such reactions are present. It is impractical to identify each of these
reactions and assess if they can or cannot be removed for the specific reacting flow being simulated
(as was done in Section 3.6.1.1 with CaltechMech). A time-integration method used with such
mechanisms has to be efficient even in the presence of such reactions (an example can be found
in Ref. [116]). This is achieved at the cost of making the preconditioner more complex (and non-
72
1-CH2
105 n-C3 H7

2-C7 H15
100
10−5
10−10
500 1000 1500 2000
Temperature (K)
Figure 3.23: Chemical consumption timescale associated with 1-CH2 , 2-C7 H15 , and n-C3 H7 vs.
temperature in the one-dimensional flame.
diagonal). Therefore, the proposed method is not expected to be efficient for the simulation of
reacting flows with very large chemical mechanisms such as those developed at LLNL. However,
these mechanisms are mainly used for 0D ignition calculations and are too large to be used for the
simulation of turbulent flames.
Third, the temperature equation is integrated explicitly, which may limit the largest stable time
step size in some reacting flow configurations. For the laminar n-heptane/air flame (Section 3.4.1),
the temperature time scale (related to heat release) is of the order of 10−4 s. This temperature time
scales goes down to 10−5 s for the 30 atm 0D ignition case (section 3.6.1.3). For turbulent reacting
flows, the convective CFL limit is generally more restrictive than any of these time scales. Therefore,
explicit treatment of the chemistry in the temperature equation should not affect the performance
of the proposed scheme for the applications it is intended for, i.e. unsteady reacting flows such as
turbulent flames. If the method were to be used to simulate flows in which the temperature time
scale is smaller than the convective CFL limit, then a similar implicit treatment of the temperature
equation as the one proposed for the species equations may be desirable.
Fourth, the proposed approximation of the diagonal of the chemical Jacobian (Eq. 3.5) may,
in some cases, introduce non-negligible deviation from the exact diagonal. For species whose con-
73
sumption rate is mostly due to recombination reactions, the corresponding term in the approximate
diagonal may be up to twice smaller (in magnitude). For instance, for the n-heptane/air flame
tested with the 35-species mechanism, the reaction OH + OH → O + H2 O accounts for most of
the consumption rate of OH. As a consequence, the exact term in the diagonal of the chemical
Jacobian corresponding to OH is about 1.7 times larger than its approximation. However, when the
approximation is replaced by the exact diagonal of the Jacobian, the increase in efficiency of the
method was found to be negligible. More specifically, the stability limit increases by only 25%, and
the convergence rate is unaffected for time step sizes smaller than 2 × 10−6 s. Since computing the
exact diagonal requires additional operations, the proposed implementation is marginally more effi-
cient. Under other circumstances, for instance in wall/flame interactions, where the importance of
the H recombination reaction has been shown [68], or in hypersonic flows, replacing the approximate
diagonal by the exact diagonal may lead to better efficiency.
In summary, use of the proposed preconditioner is particularly relevant to moderately to highly
turbulent (premixed or non-premixed) flames (high Karlovitz numbers for premixed flames) in which
the convective CFL limit is more restrictive than the largest stable time step size (due to the
chemistry) with the proposed time-integration method.
3.7 Summary
A semi-implicit preconditioning strategy, applied to an iterative method, is proposed for the time-
integration of the stiff chemistry in the simulation of unsteady reacting flows, such as turbulent
flames. The preconditioner consists of an approximation of the diagonal of the chemical Jacobian.
It is integrated into the iterative procedure already implemented in the NGA code, in order to
account for the non-linearities of the governing equations. Upon convergence of the sub-iterations,
the fully-implicit Crank-Nicolson method is recovered. Therefore, the stability of the scheme is
dictated by the stability of the sub-iterations.
The performance of the proposed method was numerically tested on two flow configurations:
a one-dimensional unstretched premixed flame and a three-dimensional turbulent premixed flame,

74
both with an unburnt mixture of air and n-heptane. First, the species lifetimes evaluated from the
preconditioned chemical Jacobian represent appropriately the smallest chemical timescales. Second,
a theoretical approximation of the rate of convergence of the sub-iterations was derived and shown
to be in good agreement with numerical results. Third, the stability limit was found to be well
approximated by the theoretical analysis. It was also shown that the stability limit does not depend
on the number of sub-iterations. Fourth, the method was shown to be second-order accurate in
time, even with only four sub-iterations. Increasing the number of sub-iterations led to a reduction
of the magnitude of the errors. With a time step size as large as a third of the stability limit, four
sub-iterations were shown to be sufficient to achieve acceptable accuracy. Fifth, while other methods
using diagonal preconditioned chemical Jacobians have been shown to lack elemental conservation
or were argued to not be time-accurate [85, 28, 27], the proposed method was shown to conserve
properly elements over time thanks to the sub-iterations. Sixth, the computational cost of a single
iteration with the proposed method is similar to that of an explicit time-integration scheme (since
the same number of sub-iterations are used). Therefore, the simulation speed-up achieved with the
proposed method corresponds to the increase in the largest stable time step size. For the three-
dimensional turbulent premixed flame, the simulation could be performed with a convective CFL of
0.8 (optimal, with or without chemistry).
The theoretical analysis for stability and convergence rate is general and is not limited by the
type of fuel, chemical mechanism, or flow configuration. Therefore, it was repeated, in the context of
one-dimensional premixed flames, with several fuels, unburnt conditions, and chemical mechanisms.
It was also performed with non-premixed flamelets using different scalar dissipation rates. The
method provided good convergence rates of the sub-iterations close to the stability limit for all the
chemical mechanisms considered. Consequently, the proposed preconditioning method showed great
potential for the efficient time-integration of turbulent flames. Although not a primary target, the
method was also shown to work for a homogeneous ignition case.
Finally, the proposed method is more suited than other methods for reacting flows in which the
convective timescales are of the order of 10−6 s or less. These correspond to moderately to highly
75
turbulent (non-premixed or premixed) flames (high Karlovitz for premixed flames).

76
Chapter 4
Direct numerical simulations1
Relying on the time-integration scheme introduced in the previous chapter, a series of direct numer-
ical simulations (DNS) of high Karlovitz n-C7 H16 /air premixed turbulent flames are performed. In
Section 4.1, the numerical approach is presented. The chemistry and transport models are validated
in Section 4.2. Grid-independence is verified in Section 4.3. Section 4.4 presents qualitative results
from the turbulent flame simulations.
4.1 Numerical approach
The flow configuration is first introduced, followed by a few additional details on the equations solved
(not mentionned in Chapter 2). Finally, the turbulence forcing method is described.
4.1.1 Flow configuration
Figure 4.1 presents a schematic diagram of the flow configuration. This is the same configuration
as the one briefly presented in Section 3.4.2. A statistically-planar, freely-propagating flame was
chosen in order to isolate the effects of turbulence on the flame from mean flow shear and curvature
effects. Furthermore, since both an inflow and an outflow are present, the simulation can be run for
an unbounded arbitrary time, allowing the turbulent flame to reach a statistically-stationary state.
The unburnt gas is a slightly lean (φ = 0.9) n-C7 H16 /air mixture at standard temperature
1 Most of the work presented in this section is published in B. Savard, B. Bobbitt, and G. Blanquart, Proc. Comb.
Inst. (2015) and in B. Savard and G. Blanquart, Combust. Flame (2015). Brock Bobbitt has contributed to the
choice of the turbulent flame parameters. Unless otherwise mentioned, the author of this thesis has contributed to all
the work presented in this chapter.
77
forcing forcing
y begins flame ends
L
inflow outflow
0
0 0.5L 8L 11L x
Figure 4.1: Schematic diagram of the flow configuration.
(Tu = 298 K) and pressure (P0 = 1 atm). These standard conditions were chosen in order to
match the experimental conditions of most premixed turbulent flames [30]. Two sets of simulations
are considered: one at a higher Karlovitz number, and one at a lower Karlovitz number2 . For
each of these sets, two simulations are performed: one with non-unity Lewis numbers and one
with unity. The parameters for both simulations are presented in Table 4.1. Given this choice of
unburnt conditions, the other parameters are chosen to maximize the Karlovitz number (for the
3
higher-Ka), while keeping a sufficiently large l/lF , where l = u0 / is the integral length scale and
lF = (Tb − Tu ) /|∇T |max is the laminar flame thickness. u0 is the rms velocity fluctuation, is the
dissipation rate, Tb is the temperature in the burnt gas, and |∇T |max is the maximum temperature
gradient throughout the flame. The Karlovitz number is defined as the ratio of the flame time
1/2
scale to the Kolmogorov time scale, i.e. Ka = tF /tη = (lF /SL ) (/ν) , where ν is the kinematic
viscosity in the unburnt mixture. The only difference between the two sets of simulations is u0 . The
Karlovitz numbers chosen are sufficiently high that the first set of flames is expected to fall at the
transition between the thin/broken reaction zones regimes (see Fig. 1.6), while the other set should
lie in the thin reaction zone regime, further away from the broken reaction zones regime. Note that
the parameters for the higher-Ka simulations were chosen optimally such that the computational
cost is constaint by both the chemistry (resolving the flame front) and the turbulent flow (resolving
the Kolmogorov lenght scale).
The turbulent flame speed is not known a priori. A first simulation was performed with tabulated
chemistry [97] (unity Lewis number) to obtain an estimate of the turbulent flame speed. The flame
was allowed to slightly drift for more than 100τ , where τ is the eddy turnover time defined as τ = k/,
2 The lower-Ka set of flames was simulated in collaboration with Simon Lapointe and can be found in S. Lapointe,
B. Savard, and G. Blanquart, Combust. Flame (2015) under review.

78
higher-Ka lower-Ka
unity Le non-unity Lei unity Le non-unity Lei
Domain size L × L × 11L L × L × 11L
L (m) 2.3 × 10−3 2.3 × 10−3
Grid 128 × 128 × 1408 128 × 128 × 1408
∆x (m) 1.8 × 10−5 1.8 × 10−5
η (m) 9 × 10−6 1.5 × 10−5
nF 23 21 23 21
φ 0.9 0.9
SL (m/s) 0.29 0.36 0.29 0.36
lF (mm) 0.43 0.39 0.43 0.39
l/lF 1.0 1.1 1.0 1.1
u0 /sL 21 18 10 9.0
Ka = tF /tη 280 220 91 78
Ret = (u0 l) /ν 190 83
Table 4.1: Parameters of the simulation. ∆x is the grid spacing (uniform), η the Kolmogorov length
scale in the unburnt gas, nF the number of grid points through the laminar flame thickness, ∆t the
time step, φ the equivalence ratio, and Ret the turbulent Reynolds number in the unburnt gas.
with k the turbulent kinetic energy (TKE). From this simulation, the turbulent flame speed was
estimated and the inlet velocity was changed to match this estimate. Then, the finite-rate chemistry
simulations (either with unity or with non-unity Lewis numbers) were started from the statistically
steady (tabulated) simulation. The simulations were run until statistically steady state was reached
(run for 10τ , almost two flame brush through times, which corresponds to the ratio of the turbulent
flame thickness to the mean bulk velocity). The data was collected over at least the next 10τ . The
unity Lewis number flame drifted by less than 0.1L from its initial position (x = 3.5L), and the
non-unity Lewis number flame by less than 0.3L. More details on the drift of the flame front (and
the associated turbulent flame speed) will be provided in Section 6.3.
The simulated flames are characterized by an important velocity ratio u0 /SL (about 20). This
is a direct consequence of the large Karlovitz number. Consequently, a special treatment has to be
applied to the inlet and the outlet in order to avoid negative inflow/outflow velocities (for numerical
stability). The unburnt gas is injected with a low TKE, such that there are no negative inlet
velocities. This inflow is generated from a separate homogenous isotropic turbulence simulation [157,
31]. Velocity field forcing (subsection 2.3) maintains this low TKE over a distance of 0.5L, after
which the forcing magnitude is increased such that the TKE reaches the desired value. This nominal
velocity field forcing is stopped after a distance of 8L, allowing the turbulence to decay sufficiently
79
that there are no negative axial velocities at the outlet. The forcing method used is described in
subsection 2.3.
Note that unless otherwise mentioned, the analysis throughout this thesis is done on the higher-
Ka flames. The lower-Ka flames will be used only as a validation tool for the models presented in
the next chapters (more details in Section 4.4.3).
4.1.2 Governing equations
The governing equations solved are described in Section 2.1, with the exception of the momentum
equation, to which a forcing term is added:
∂
(ρu) + ∇ · (ρu ⊗ u) = −∇p + ∇ · σ + f , (4.1)
∂t
where, f is a forcing term used to maintain the presence of turbulent fluctuations (see subsec-
tion 4.1.3).
The species and enthalpy production terms are taken from the reduced, n-heptane chemical
model which contains 35 species and 217 elementary reactions introduced in Section 3.4.1. This
model corresponds to a slightly reduced version of that found in Ref. [17]. A validation is presented
in Section 4.2.1.
The species viscosities µi are computed by standard kinetic theory [76] and the mixture vis-
cosity µ is calculated using Wilke’s formula [179]. The mixture thermal conductivity is obtained
following Mathur et al. [115], where the species thermal conductivities λi are computed by Eucken’s
formula [60]. In order to reduce the computational cost of the simulation, the species diffusivities
are computed as Di = α/Lei , with the Lewis numbers Lei assumed to be constant throughout the
flame. For the non-unity Lewis number simulation, these species Lewis numbers Lei are extracted
from the simulation of a one-dimensional, laminar premixed flame with full transport properties,
using FlameMaster [146]. The Lewis numbers are evaluated in the burnt gas such that deviations
from the full transport solution are negligible. These are listed in Table 4.2. A validation of the
80
constant Lewis number assumption is presented in Section 4.2.2.
N2 0.99 HCO 1.22 A-C3 H5 1.79

1-CH2 0.94 CH2 O 1.23 n-C3 H7 1.81
3-CH2 0.94 CH3 0.96 C2 H6 1.40
O 0.69 CO2 1.37 P-C3 H4 1.68
H2 0.28 CH4 0.97 A-C3 H4 1.68
H 0.16 C2 H3 1.28 C3 H6 1.80
OH 0.70 C2 H4 1.28 1-C4 H8 1.99
H2 O 0.79 C2 H5 1.39 1-C5 H10 2.27
O2 1.06 C2 H 1.25 1-C5 H11 2.08
HO2 1.07 HCCO 0.86 2-C7 H15 2.84
CH 0.64 C2 H2 1.27 n-C7 H16 2.84
CO 1.07 C3 H3 1.6
Table 4.2: Constant species Lewis numbers used in the turbulent flame simulations (evaluated in
the burnt gas from a one-dimensional flame solution).
The governing equations are solved numerically using the energy conservative, finite difference
code NGA [55], described in Section 2.2. The scheme used is second-order accurate in both space
and time. The semi-implicit Crank-Nicolson time integration described in Chapter 3 is used. The
third-order Bounded QUICK scheme, BQUICK [74], is used as the scalar transport scheme to ensure
the transported species mass fractions and temperature remain within their physical bounds.
4.1.3 Turbulence forcing
As all other similarly high Ka numerical simulations from the literature [10, 9, 151], the present
configuration misses the generation of turbulence due to large scale flow straining (larger than the
domain size). As a result, the turbulence is expected to decay. The decay of the TKE ahead of the
flame (in the unburnt gas) in the absence of velocity field forcing can be estimated theoretically by
analogy to decaying isotropic turbulence, considering dk/dt = −k/τ . With U the mean bulk/inlet
velocity, the characteristic length scale over which the TKE decays is U τ = 0.1 mm, which is too
small compared to the laminar flame thickness (about 0.4 mm, see Table 4.1). As a consequence,
the use of velocity field forcing is necessary.
In previous work, spectral forcing techniques were often used to offset the decay of TKE and
maintain the turbulence characteristics [10, 151]. In the present work, the linear velocity forcing
method [111, 157, 32] was preferred for its more physical nature and good stability properties [31].
81
The linearly forced turbulent field under comparable Reynolds number was analyzed in Ref. [32]
and it was shown that the second- and third-order structure functions and the energy spectrum are
self-consistent and in agreement with experimentally obtained data [126] of decaying grid turbulence.
The linear forcing method mimics the missing large scale straining by appending a source term to
the momentum equation (Eq. 4.1) [111, 157].
The method is adapted to take into account the axially evolving nature of the flow. Consequently,
the forcing term f in Eq. 4.1 takes the following form:
k0
f (x, y, z, t) = A (4.2)
k (x, t)
× (ρ (x, y, z, t) u (x, y, z, t) − ρu (x, t)) ,
1/2 1/2
where A = 0 / (2k0 ) is the forcing coefficient (computed as A = (2/27) k0 /l2 [31]), which
takes the form of the inverse of a time scale, k0 is the desired TKE, 0 is the corresponding desired
dissipation rate, and k is the planar Favre-averaged TKE, defined as

1 ]00 )2 + (v
] 00 )2 + (w
^ 00 )2 .
k= (u (4.3)
2
Planar Favre averages are defined as
e = ρφ ,
φ (4.4)
ρ
with the standard (Reynolds) planar average
Z L Z L
1
φ (x, t) = φ (x, y, z, t) dydz. (4.5)
L2 0 0
With the forcing method used, homogeneous isotropic turbulence is imposed upstream of the
flame. A slight decrease in the TKE through the flame and a relaxation back to the imposed
TKE further downstream was found, as shown in Fig. 4.2. This evolution is consistent with the
experimental results of Cheng et al. [40]. While the trends agree, in both studies the variations in
82
TKE through the flame remain marginal, which has also been observed computationally [160].
1.2
0.8
k/k0
0.6
0.4
0.2
0
0 2 4 6 8 10
x/L
Figure 4.2: Time-averaged (over 10 eddy turnover times), planar Favre-averaged TKE vs. x for the
non-unity Lewis number flame. The average flame position is shown by the dashed line (more details
on the flame position are given in Section 6.3.1).
Figure 4.3 presents the energy spectra (computed as in Ref [152]) in the unburnt gases for both
the lower- and the higher-Ka flames3 . First, it is important to note that, as expected, both profiles
collapse on each other. Second, the presence of an inertial sub-range is very limited. This is to be
expected given the turbulent Reynolds numbers (Table 4.1). Finally, for both cases, the turbulent
kinetic energy is contained over two decades of length scales. This is also to be expected given
the turbulent Reynolds numbers of approximately 100. This range of length scale is limited by the
inherent computational cost.
4.2 Chemistry and transport models validation
In this section, we provide additional justifications and validation for the choice of chemical model
and transport model.

3 The profiles were computed by Simon Lapointe and can be found in S. Lapointe, B. Savard, and G. Blanquart,
Combust. Flame (2015) under review.

83
1000
lower-Ka
higher-Ka
100
-5/3
10
5 1/4
E(κ)/(εν )
1
0.1
0.01
0.001
0.0001
0.01 0.1 1 10
κη
Figure 4.3: Normalized energy and dissipation spectra for the higher- and lower-Ka non-unity Lewis
number flames. The two-dimensional three components spectra taken in a y-z plane in the unburnt
gases (averaged over time). κ = 2kπ/L, for k = 1, 2, ..., Ny /2, with Ny the number of points in the
y- or z-direction, is the wavenumber.
4.2.1 Reduced mechanism performance
The turbulent flame speed is one of the most important quantity that characterizes a turbulent
premixed flame. This overall consumption rate is a strong function of the local flame speed. An
effect of differential diffusion is to modify the local equivalence ratio in locations where the flame
is curved [142, 113]. Since highly turbulent flames are subject to locally large curvatures, the
mechanism should predict adequately the local flame speed over a range of equivalence ratios. In
order to validate the performance of this reduced mechanism (presented in Section 3.4.1), the flame
speed vs. equivalence ratio profile is compared to experimental data in Fig. 4.4. The flame speeds
are obtained from one-dimensional unstretched flames simulated with the same reduced mechanism
with FlameMaster [146]. Two solutions are compared: one using full transport (mixture-averaged
diffusivities and Soret diffusion) and one using constant Lewis numbers (as listed in Table 4.2)
without Soret diffusion. The laminar flame speeds obtained with both transport models are in very
good agreement with experimental data over a wide range of equivalence ratios.
84
45
40
35
30
(cm/s)
25
20
L
0
S
15
10 full transport
no Soret, cst Le
5 Davis 1998
van Lipzig 2011
0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3
φ
Figure 4.4: Laminar flame speed vs. equivalence ratio profiles obtained numerically with the 35-
species mechanism and experimentally [51, 170]. The numerical simulations were performed with
full transport (solid line) and constant Lewis numbers (dashed line).
4.2.2 Transport model validation
The transport model is further tested on a one-dimensional unstretched flame simulated with
FlameMaster. The same unburnt gas parameters (Tu , P0 , φ) as in the DNS are used. The mass
fraction vs. temperature profiles for the fuel (n-C7 H16 ), an intermediate species (CH2 O), and a
product (H2 O), obtained from full transport (mixture-averaged diffusivities and Soret diffusion),
mixture-average transport without Soret diffusion, and constant Lewis numbers without Soret dif-
fusion solutions, are shown in Fig. 4.5(a) to 4.5(c). Similarly, the source term vs. temperature
profiles are shown in Fig. 4.5(d) to 4.12(b) for the same species. It is clear from these figures that
the constant Lewis number assumption introduces virtually no deviation from the mixture-averaged
transport solution. Soret diffusion has a noticeable effect only in the fuel mass fraction vs. temper-
aure profile. This effect would, however, be attenuated in turbulent flames as turbulent transport is
especially important at low temperature in the preheat zone (more details in Chapter 5). The good
agreements shown in Fig. 4.5 are particularly important given the computational cost reduction of
using fixed, constant Lewis numbers without Soret diffusion.
Note that while the simplified transport model gives good agreement with the full transport model
in a one-dimensional laminar setting, it is unclear to what extent it should perform equally well in a
85
0.06 0.0016 0.08

full transport full transport
no Soret 0.0014 no Soret 0.07 full transport
0.05 no Soret, cst Le no Soret, cst Le no Soret
n-C7H16 mass fraction
CH2O mass fraction

0.06 no Soret, cst Le
H2O mass fraction

0.0012
0.04
0.001 0.05
0.03 0.0008 0.04
0.0006 0.03
0.02
0.0004 0.02
0.01
0.0002 0.01
0 0 0
300 700 1100 1500 1900 300 700 1100 1500 1900 300 700 1100 1500 1900
T (K) T (K) T (K)
(a) (b) (c)
0 8 250
full transport
-50 6 no Soret
200 no Soret, cst Le
4
ωC H (kg m-3 s-1)
-100 ωCH O (kg m-3 s-1)
ωH O (kg m-3 s-1)

2 150
-150
0
-200 100
-2
7 16
-250
2
-4 50
2
full transport
-300 -6 no Soret
full transport no Soret, cst Le 0
-350 no Soret -8
no Soret, cst Le
-400 -10 -50
300 700 1100 1500 1900 300 700 1100 1500 1900 300 700 1100 1500 1900
T (K) T (K) T (K)
(d) (e) (f)
Figure 4.5: Comparison of species mass fraction (top row) and source term (bottom row) vs. tem-
perature profiles for n-C7 H16 , CH2 O, and H2 O between the one-dimensional, unstretched, flame
solutions obtained with full transport (solid lines) and constant Lewis numbers (dashed lines). The
unburnt conditions are the same as in Section 4.1.
three-dimensional turbulent setting. In order to partially asses the validity of this transport model
in the turbulent flame, the non-unity simulation was performed with mixture-averaged diffusivities
by Nicholas Burali4 . Given the cost of the simulation, it was performed for 13 eddy turnover times.
Figure 4.6 compares the fuel burning rate statistics for the two simulations. The last 3 turnover
times of the mixture-averaged simulation are considered (the first ten being considered a transient
period). The fuel burning rate is shown as it will be investigated extensively in the next chapters.
This quantity is also chosen because the species source terms are much more sensitive to turbulent
fluctuations than the species mass fractions (more details in the following chapters). Figure 4.6
shows that only marginal differences can be identified between the two flames. This means that the
effect of turbulence is captured independently of the transport model.

4 N. Burali and G. Blanquart, Combust. Flame, in preparation.
86
Finally, the effect of Soret diffusion in the turbulent flame should be analyzed in more details in
future work.
0 0.0035
mix avg mix avg
cst Le 0.003
cst Le
-50
<ω• F|T> (kg m-3 s-1)
0.0025
-100 0.002
PDF
0.0015
-150
0.001
-200 0.0005
0
-250 0 100 200 300 400 500 600 700 800 900
400 600 800 1000 1200 1400 1600 1800 2000 2200
T (K) -ω• F (kg m-3 s-1)
(a) Conditional mean (b) Conditional PDF (T = 1240 K)
Figure 4.6: Comparison of the fuel burning rate statistics in the non-unity Lewis number turbulent
flame with mixture-averaged diffusivities and constant Lewis numbers.
4.3 Grid refinement
The grid spacing is chosen such that κη > 1.5 everywhere in the domain, and is limited by the
turbulence in the unburnt gases, where η is the smallest. In order to verify the quality of the
solution, a non-unity Lewis number simulation with twice the number of grid points per direction
is performed for 13 eddy turnover times (limited by the cost of such a simulation). Figure 4.7
compares the fuel burning rate statistics for the nominal- and refined-grid simulations. Again, the
last 3 turnover times are considered for the refined-grid simulation. Only small differences can be
identified. It will become more clear in Chapter 6 that these differences are marginal for the purpose
of the analysis made in this work. The conclusions made in the next chapters should therefore be
grid independent.
87
0 0.0035
refined refined
nominal 0.003
nominal
-50
<ω• F|T> (kg m-3 s-1)

0.0025
-100 0.002
PDF
0.0015
-150
0.001
-200 0.0005
0
-250 0 200 400 600 800 1000
400 600 800 1000 1200 1400 1600 1800 2000 2200
T (K) -ω• F (kg m-3 s-1)
(a) (b)
Figure 4.7: Comparison of the fuel burning rate statistics in the non-unity Lewis number turbulent
flame with refined and nominal grids.
4.4 Turbulent flame
In this section, qualitative results for the non-unity Lewis number flame are presented. Further
analysis will be performed in Chapters 5 to 6 (using both the non-unity and the unity Lewis number
flames).
4.4.1 Thickened flame
At high Karlovitz numbers, the spatial structure of a premixed flame is expected to depart from that
of a laminar flame, as turbulent structures are sufficiently small enough to penetrate the preheat
zone, and potentially the reaction zone. Figure 4.8 presents contours of vorticity, n-C7 H16 and CH2 O
mass fractions, and temperature through the flame brush on a two-dimensional horizontal slice of
the non-unity Lewis number turbulent flame. This figure is representative of both the unity and the
non-unity Lewis number turbulent flame brushes.
As expected the preheat zone appears largely thickened [142] (approximately 10 times larger
than the laminar flame). It is interesting to note that smaller turbulent structures are observed
upstream of the flame (i.e. in the preheat zone) compared to close to the reaction zone. This is
partially-explained by the fact that the kinematic viscosity increases (by up to a factor of 30) and
88
1/4
the Kolmogorov length scale, η = ν 3 / , increases through the flame by about a factor of 13.
3x105
1.5x105
Vorticity (s-1)
0
0.056
0.028
Mass fraction
n-C7H16
0
1x10-3
5x10-4
Mass fraction
CH2O
0
2100
1850
1200
350
Temperature (K)
300
Figure 4.8: Contours of vorticity, n-C7 H16 and CH2 O mass fractions, and temperature through the
flame brush on a two-dimensional horizontal slice. The laminar flame thickness lF is added for
comparison.
4.4.2 Reaction zone
Contours of the source term of n-C7 H16 and H2 O for the non-unity Lewis number flame are shown
in Fig. 4.9 on the same two-dimensional slice as in Fig. 4.8. For a Karlovitz numbers as large as 220,
scaling arguments suggest that turbulent structures should be sufficiently small enough to penetrate
the reaction zone [142]. However, two observations can be made from Fig. 4.9: 1) the reaction zone
appears to be thin and 2) signs of local extinctions (broken reaction zone) can be observed in the
fuel consumption zone, but cannot be observed in the H2 O production zone.
The first observation could seem inconsistent with the fact that the Kolmogorov length scale
in the unburnt gas is 10 times smaller than the laminar reaction zone thickness. However, as
mentioned above, the Kolmogorov length scale increases through the flame due to the increasing
kinematic viscosity, and becomes as large as the laminar reaction zone thickness as it reaches the
burnt side.
To simplify the following analysis, only the fuel consumption will be further investigated in this
89
thesis. This is first justified because only the consumption rate of the fuel shows local extinction.
Second, the magnitude of the relative fluctuations in the fuel chemical source term is amongst the
largest of all species source terms. Third, besides hydrogen (atomic and molecular), the fuel has the
Lewis number furthest from unity, which makes it a good candidate to analyze differential diffusion
effect, and, in particular, greater-than-unity Lewis number effects.
550
275
(kg m-3 s-1) 0

250
125
(kg m-3 s-1) 0
Figure 4.9: Contours of the source terms of n-C7 H16 (top) and H2 O (bottom) on the same two-
dimensional horizontal slice as in Fig. 4.8. The isoterm T = 1500 K (white) is also shown. The
laminar reaction zone thicknesses (full width at half-height) of n-C7 H16 (δC7 H16 ) and H2 O (δH2 O )
are also shown for comparison.
The qualitative results presented in this section and the previous one are summarized in Fig-
ure 4.10: turbulence is affected as it progresses through the flame (increased viscosity resulting in
an increased Kolmogorov length scale), the flame is largely thickened (increased turbulent mixing
in the preheat zone), and the reaction zone is thin and locally broken.
4.4.3 Lower-Ka flame
Results from the lower-Ka flame are qualitatively similar, but the effects of turbulence identified for
the higher-Ka flame are less pronounced, as illustrated in Fig. 4.11 and 4.12. For this reason, only
the higher-Ka flame will be characterized in Chapters 5 and 6. The lower-Ka flame will be used to
validate the models developed in Chapters 5 and 7, as these should hold through the thin reaction
zones regime.
90
3x105
1.5x105
Vorticity (s-1) 0
1.23
0.695
Density (kg m-3)

0.16
600
300
(kg m-3 s-1)

0
Figure 4.10: Contours of vorticity, density, and fuel burning rate through the flame brush on a
two-dimensional horizontal slice of the non-unity Lewis number flame. The T = 1240 K isocontour
is also shown (red).
Figure 4.11: Two-dimensional slices showing the vorticity and temperature for the non-unity Lewis
number flames. The yellow line indicates the T = 940 K isocontour and the red line indicates the
T = 1540 K isocontour. The vorticity ranges are satureated at [0,8e4] (s−1 ) and [0,1.6e6] (s−1 ). The
temperature ranges are both [298,2200] K.
(a) Ka = 78 (b) Ka = 220
Figure 4.12: Two-dimensional slices showing the fuel consumption rate for the non-unity Lewis
number flames. The fuel consumption rate range is saturated at [0,650] (kg m−3 s−1 ). The T = 1240
K isocontour is also shown (white).
91
Chapter 5
Characterization and modeling of

the flame structure1
The qualitative results from the previous chapter show that the flame is significantly affected (thick-
ened) by turbulence. The objective of this chapter is to characterize the effect of turbulence on
the flame structure and to model the effect of turbulent mixing on this mean structure. The flame
structure is characterized first. Second, an a priori model for the flame structure of non-unity Lewis
number turbulent flames is developed/validated using a set of DNS of turbulent hydrogen/air flames
(simpler chemistry) performed by Aspden et al. [10] for a wide range of Karlovitz numbers. Finally,
the model is tested on the present non-unity n-heptane/air flame. A discussion on the validity and
the limitations of the model for this flame is also provided.
5.1 Structure of the n-heptane flame
In this section, the structure of both the unity and the non-unity Lewis number flames are presented.
The effects of turbulence in the absence of differential diffusion are presented first (unity Lewis
number flame). Second, the differential diffusion effects are highlighted with the non-unity Lewis
number flame.
1 The first two sections of this chapter are published in B. Savard, B. Bobbitt, and G. Blanquart, Proc. Comb.
Inst. (2015) and in B. Savard and G. Blanquart, Combust. Flame (2014), respectively. The author of this thesis has
contributed to all the work presented in this chapter.
92
5.1.1 Turbulent flame structure in the absence of differential diffusion
The effects of turbulence on the flame are illustrated in Fig. 5.1. By comparing this figure to Fig. 4.8,
the unity and the non-unity Lewis number flames look very similar qualitatively. In particular, both
flames look very different from a one-dimensional flame.
Figure 5.1: Contours of vorticity, n-C7 H16 and CH2 O mass fractions, and temperature through the
flame brush on a two-dimensional horizontal slice. The laminar flame thickness lF is added for
comparison.
To properly assess the influence of turbulence on the flame structure, one can analyze the correla-
tion between species and temperature (or any other progress variable). As such, the flame structure
can be adequately compared to that of a one-dimensional laminar flame, which is well represented in
temperature space. Any departure from a one-dimensional laminar flame due to turbulence should
be captured by these species mass fraction profiles.
In this sense, several species mass fractions are plotted against temperature and are compared
to their one-dimensional laminar flame equivalent. Figure 5.2 shows joint probability densities of
n-C7 H16 , C2 H4 , and CO2 mass fraction, vs. temperature. These species correspond to a reactant,
an intermediate species, and a product, respectively. The conditional mean of these species mass
fraction (conditional on temperature) is also shown. This figure is representative of the overall flame
structure as the mass fractions of other species show similar behaviors. These results suggest that
93
0.06 0.008
<Y|T> <Y|T> 0.12
flamelet 0.03 flamelet
C2H4 mass fraction

0.006
0.04
0.02 0.08
0.004
0.02 0.04
0.01
0.002
0 0 0 0
500 1500 2500 500 1500 2500
T (K) T (K)
0.18
<Y|T>
CO2 mass fraction flamelet 0.01
0.12
0.005
0.06
0 0
500 1500 2500
T (K)
Figure 5.2: Joint PDF and conditional mean (solid line) of the n-C7 H16 (top left), C2 H4 (top right),
and CO2 (bottom) mass fraction vs. temperature from the unity Lewis number DNS. The unity
Lewis number flamelet solution is also shown (dashed line).
the influence of turbulence on the flame structure in the absence of differential diffusion is very
limited as the spread of the joint PDF is limited (this has also been observed by Aspden et al. for a
high-Ka CH4 /air flame [9]). More interestingly, the conditional mean profiles of these species follow
very closely the profiles of a one-dimensional, unstretched laminar flame at the same condition. This
result is surprising, as the turbulent flame is clearly not in the flamelet regime and does not look
like a flamelet.
Although the last result may be surprising, an expected first order effect of turbulence on species
transport is the increase in the effective diffusivity through increased mixing. Assuming turbulence
mixes all scalars the same way, the turbulent diffusivities (DT ) may be assumed equal for all of these
94
scalars, and the effective diffusivity becomes
Deff = D + DT . (5.1)
To assess the effect of diffusivity on the flame structure, additional laminar flamelet solutions were
obtained varying this diffusivity. Whereas the resulting laminar flame speeds and flame thicknesses
1/2
were accordingly altered by a factor of (Deff /D) , the flame structure was virtually unaffected, as
shown in Fig. 5.3.
1.0x10-3
D=0.1D0, SL=9cm/s
D=D0, SL=29cm/s
-4 D=10D0, SL=92cm/s
8.0x10
H2 mass fraction
6.0x10-4
5.43x10-4
4.0x10-4
2.0x10-4
-4
5.40x10
1800 1805
0
600 1000 1400 1800 2200
T (K)
Figure 5.3: Flamelet solutions of the H2 mass fraction vs. temperature with varying diffusivity
coefficients.
In summary, while the turbulent flame is thickened by turbulence and is clearly not a thin flame,
its structure is similar to that of a flamelet. This may suggest that the use of a progress variable
with tabulated chemistry [171, 65, 97] would be justified and sufficient even at such high Karlovitz
number.
5.1.2 Turbulent flame structure with differential diffusion
Similarly to Fig. 5.2, Fig. 5.4 presents the structure of C2 H4 through the non-unity Lewis number
flame. The full-transport flamelet solution is also added for comparison. While turbulence has
almost no impact on the structure of the unity Lewis number flame, it has a clear effect on that of a
95
non-unity Lewis numbers flame. In this case, the turbulent flame structure lies between that of a full
transport and a unity Lewis number flamelet. Once again, a first order effect of turbulence on scalar
transport is an increase in the effective diffusivity of each scalar (including species mass fractions
and temperature) through increased mixing. As a result, the effective species Lewis number take
the following from:

α + DT
Lei,eff = . (5.2)
Di + DT
A similar expression for these effective Lewis numbers was first suggested by Peters [142]. More
details on DT and Lei,eff and how they relate to the scalar transport equations will be provided in
Section 5.2.2.3. Equation 5.2 suggests that if the turbulence were sufficiently intense, the non-unity
Lewis number case would behave the same as the unity Lewis number case, i.e. turbulence would
suppress differential diffusion effects.
0.012
<Y|T>
0.12
flamelet - full transport
0.01 flamelet - unity Le
C2H4 mass fraction
0.008
0.08
0.006
0.004
0.04
0.002
0 0
600 1000 1400 1800 2200
T (K)
Figure 5.4: Joint PDF and conditional mean (solid line) of the C2 H4 mass fraction vs. temperature
from the non-unity Lewis number DNS. The non-unity and unity Lewis number flamelet solutions
are also shown (dashed line).
This behavior is observed at low temperatures (i.e. in the preheat zone), where the data from
the DNS collapse perfectly with the unity Lewis number flamelet solution. However, at higher
temperatures, the structure deviates from that of a unity Lewis number flamelet. This can be
attributed to the fact that the Kolmogorov length scale grows through the flame, and therefore the
“local Karlovitz number” is reduced. As a result, preferential diffusion effects may still be present,
96
especially towards the reaction zone. These effects are disscussed in more details in the following
chapter.
5.2 Effective species Lewis numbers
In this section, a model for the effective species Lewis numbers (Eq. 5.2) is provided. Given the
complexity of the n-heptane/air flame studied, another set of turbulent premixed flames is chosen
to verify/develop the model. A series of Direct Numerical Simulations of lean hydrogen/air flames
performed by Aspden et al. [10] is considered. The reason behind this choice of flames is three-
fold: 1) the hydrogen/air flame has far simpler chemistry (similar to one-step chemistry) than the
n-heptane/air flame, 2) the viscosity ratio across the flame is considerably lower in the hydrogen/air
flame (13 vs 30 for the n-heptane/air flame), and 3) the hydrogen/air flames were performed over a
wide range of Karlovitz numbers, such that the transition between the purely laminar to the fully
turbulent flame structure can be studied.
Section 5.2.1 presents the flame structure obtained from the DNS data of Aspden et al. and
compares it to corresponding laminar unstretched flamelets. Section 5.2.2 derives an a priori model
from simplified species and temperature balance equations. Section 5.2.3 compares the model against
the effective Lewis numbers computed from the DNS. Section 5.2.4 discusses different approaches
to derive a model for these effective Lewis numbers. A Reynolds number versus a Karlovitz number
dependency is especially emphasized. Finally, Section 5.2.5, in addition to disscussing the model’s
sensitivity, applicability, and practical use, presents the impacts of the effective Lewis numbers on
the laminar flame speed and the laminar flame thickness, the effective Karlovitz number and the
regime diagram, and the turbulent flame speed models.
The validity and the limitations of the model for the n-heptane/air flame will be discussed in
Section 5.3.
97
5.2.1 Turbulent flame structure
In this section, the flame structure from the series of Direct Numerical Simulations performed by
Aspden et al. [10] is compared to the structure of laminar unstretched flamelets. A schematic
diagram of the flow configuration used for the DNS is presented in Fig. 5.5. The complete set of
parameters describing the DNS cases can be found in Ref. [10] and are summarized in Table 5.1. Note
that Aspden et al. used a different definition for the Karlovitz number than the one presented in
3/2 −1/2
the introduction, i.e. Ka = (u0 /SL ) (l/lF ) . Based on scaling arguments, the two definitions
are equivalent [142]. The one used by Aspden et al. assumes SL lF = ν. The reactants are a
lean (φ = 0.4) hydrogen-air mixture. A reduced version of GRI-MECH 2.11 was used as the
chemical mechanism (9 species, 27 reactions; all carbon-based species and associated reactions were
removed). Soret and Dufour effects as well as radiation were not included in the DNS [10]. As
confirmed in Ref. [10] and shown in Fig. 5.6, the transition from the thin reaction zone to the broken
reaction/distributed burning zone is covered by the simulation cases.
L
Flame propagation
Product
Forced turbulence
zero mean flow
Fuel
Figure 5.5: Schematic diagram of the flow configuration used by Aspden et al. [10]. Diagram taken
from Ref. [10].
The laminar flame counterparts are simulated using FlameMaster [146]. The equivalence ratio
is fixed to 0.4; the same chemical model as in the DNS is used; and Soret and Dufour effects and
98
Figure 5.6: Contours of hydrogen mass fraction and temperature on a two-dimensional vertical slice
of the DNS case D [10].
Case A B C D
Equivalence ratio (φ) 0.4 0.4 0.4 0.4
Laminar flame speed (SL ) (m/s) 0.224 0.224 0.224 0.224
Laminar flame thickness (lF ) (mm) 0.629 0.629 0.629 0.629
Length ratio (l/lF ) 0.5 0.5 0.5 0.5
Velocity ratio (u0 /SL ) 3.69 17.1 32.9 106.8
Turbulent Reynolds number 14.2 65.8 126.1 410.7
based on viscosity of unburnt gases (ReT )
3/2 −1/2
Karlovitz number (Ka = (u0 /SL ) (l/lF ) ) 10 100 266 1526
Table 5.1: Parameters for the series of turbulent premixed hydrogen flame DNS performed in
Ref. [10].
radiation heat losses are ignored.
The conditional means of the mass fractions with respect to temperature < Yi | T > are calcu-
lated for several instantaneous snapshots of the established (statistically steady) propagating flame.
Figure 5.7 shows the conditional mean hydrogen mass fraction as a function of temperature for
cases A through D. The temperature is used here as a progress variable. The profiles for laminar un-
stretched flamelets with full transport [177, 91] and unity Lewis numbers are also shown in Fig. 5.7.
The “full transport” and “unity Lewis numbers” laminar flames correspond to limiting cases of
purely laminar and fully turbulent premixed flames, respectively. A clear trend is observed: the
DNS profiles gradually move from the purely laminar towards the fully turbulent limiting cases. It
99
is important to stress that all DNS data are “bracketed” by these two limiting cases. As mentioned
earlier, one of the objectives of this section is to model this transition between the purely laminar
to the fully turbulent flame structure.
0.012
A
0.01
DNS
{ B
C
D
H2 mass fraction
full transport
0.008
unity Lewis
0.006
0.004
0.002
0
200 400 600 800 1000 1200 1400 1600 1800
T (K)
Figure 5.7: Conditional mean hydrogen mass fraction profiles as a function of temperature for the
DNS cases A through D and unstretched laminar flames with full transport and unity Lewis numbers.
DNS from Aspden et al. [10]
The effects of thermo-diffusive instabilities are revealed by the presence of hot-spots in Fig. 5.7,
i.e. by the extension of the DNS profiles to temperatures higher than the adiabatic flame temperature
(Tad ≈ 1400 K). However, the present analysis is aimed to address the turbulent transport problem
only. The hot spots are in particular not relevant to the n-heptane/air flames considered in this
thesis as they are thermo-diffusively stable (more details in Chapter 6). Unfortunately, in the
hydrogen/air flames, the turbulent transport is influenced by these thermo-diffusive instabilities.
While these instabilities have an important impact on laminar flames [53], they are expected to
be less important as turbulence increases, i.e. as preferential diffusion becomes negligible due to
turbulent mixing. This is observed for the highest Ka DNS case (D). The qualitative behavior of the
turbulent flame structures shown in Fig. 5.7 and, as it will be presented in Section 5.2.3, the good
agreement of the proposed model suggest that these instabilities have limited impact on the statistical
mean structure. As the hot spots and the super-adiabatic temperatures are not a consequence of
turbulent transport (of concern in this chapter) and are entirely due to those instabilities, they are
not considered in this work. Computing the conditional mean as < Yi | T > instead of < T | Yi >
100
Species Lei
N2 1.169
O 0.772
O2 1.223
H 0.202
OH 0.787
H2 0.325
HO2 1.195
H2 O2 1.203
H2 O 0.932
Table 5.2: Lewis numbers used in the constant, non-unity Lewis numbers model. These Lewis
numbers are evaluated at 723 K from a full transport flamelet solution.
allows one to isolate these hot spots.
In order to simplify the theoretical analysis in the following sections, the full transport model will
be replaced by a constant, non-unity Lewis numbers model. Figure 5.8 shows that flamelets with
full transport and constant Lewis numbers have virtually the same hydrogen mass fraction profiles.
This indeed justifies the assumption of constant Lewis numbers through the flame. Regardless of
the temperature at which the constant Lewis numbers are evaluated, the deviation from the full
transport profile is negligible. This deviation was, however, minimized by evaluating the Lewis
numbers at 723 K, i.e. at the beginning of the reaction zone. These constant Lewis numbers are
listed in Table 5.2.

0.012
full transport
constant Lewis
0.01
H2 mass fraction
0.008
0.006
0.004
0.002
0
200 400 600 800 1000 1200 1400 1600
T (K)
Figure 5.8: Hydrogen mass fraction profiles as a function of temperature for flamelets with full
transport and constant Lewis numbers.
101
5.2.2 Proposed model
In this section, we develop a model for the effective Lewis numbers in a turbulent flame.
5.2.2.1 Species transport equation
The species transport equation (Eq. 2.5) presented in Section 2.1 is also relevant to the hydrogen/air
flame:
diffusion
}| { z velocity correction
∂ρYi α z }| {
+ ∇ · (ρuYi ) = ∇ · ρ ∇Yi + ω̇i + ∇ · (ρYi Vc,i ) . (5.3)
∂t | {z } Lei |{z}
advection chemical source
Figure 5.9 shows the contribution from each term in Eq. 2.5 for H2 , in the case of the constant,
non-unity Lewis numbers (Table 5.2) unstretched laminar flamelet (same unburnt conditions as in
Section 5.2.1). The unsteady term is obviously absent. The conclusion is that the diffusion term
associated with the correction velocity is negligible in comparison to the other terms.
1
advection
0.8 diffusion
normalized magnitude
chemical source
0.6
velocity correction
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
200 400 600 800 1000 1200 1400 1600
T (K)
Figure 5.9: Contribution from each term in the species transport equation for H2 as a function of
temperature, obtained for an unstretched laminar flamelet with constant, non-unity Lewis numbers.
Each contribution is normalized by the largest absolute value of all contributions.
102
5.2.2.2 Temperature equation
The temperature equations (Eq. 2.4) presented in Section 2.1 can also be written in the following
form:
others
diffusion z }| {
∂ρT z }| { ω̇T ρα X cp,i α
+∇ · (ρuT ) = ∇ · (ρα∇T ) + + ∇cp · ∇T + ρ ∇Yi + Yi Vc,i · ∇T .
∂t | {z } cp cp i
cp Lei
advection |{z}
chemical source
(5.4)
Figure 5.10 shows the contribution from each term in Eq. 5.4 calculated for the same flamelet as in
subsection 5.2.2.1. As shown, the sum of the last two terms on the right hand side of Eq. 5.4 has
negligible contribution.
1.2
advection
1 diffusion
normalized magnitude
0.8 chemical source

others
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
200 400 600 800 1000 1200 1400 1600
T (K)
Figure 5.10: Contribution from each term in the temperature equation as a function of temperature,
obtained for an unstretched laminar flamelet. Each contribution is normalized by the largest absolute
value of all contributions.
5.2.2.3 Effective Lewis numbers
Any model should at least capture the mean of the turbulent flame (in a statistical sense). Reynolds-
averaged transport equations provide information about the ensemble/statistical average of the
transported quantity. Hence, the choice of a RANS formalism is made. Omitting the terms found
to be negligible in subsections 5.2.2.1 and 5.2.2.2, the Reynolds-averaged species and temperature
103
balance equations can be written as
∂ e h i
ρYi + ∇ · ρũYei = ∇ · ρ (Di + DT ) ∇Yei + ω̇i (5.5)
∂t

∂ e h i ω˙T
ρT + ∇ · ρũTe = ∇ · ρ (α + DT ) ∇Te + , (5.6)
∂t cp
where ¯· denotes a Reynolds average and e· a Favre average. DT is the eddy diffusivity and
is assumed to be identical for all species and temperature. This result is a direct consequence of
assuming turbulence mixes all scalars the same way. A similar assumption is made in transported
PDF methods [174, 152]. From Eq. 5.5 and 5.6, an effective Lewis number can be obtained for each
species:
Dheat α + DT 1 + DαT
Lei,eff = = = 1 DT
. (5.7)
Dmass Di + DT Lei + α
As previously stated, a similar expression for the effective/turbulent Lewis number was first proposed
by Peters [142].
5.2.2.4 Turbulent model
Using a k- model for the eddy viscosity νT [178], the eddy diffusivity can be written as
νT Cµ k 2
DT = = , (5.8)
P rT P rT
where is the turbulent dissipation rate, k is the turbulent kinetic energy, Cµ is the k- model
coefficient, and P rT is the turbulent Prandtl number. Using the same definition for the integral
length scale l as in Ref. [10],

3
u0
l= (5.9)

and the turbulent kinetic energy

3 02
k= u , (5.10)
2
104
one obtains
9 Cµ u0 l
DT = , (5.11)
4 P rT
with u0 the root-mean-square velocity fluctuation. Using the definition of the mixture Prandtl
number P r, the ratio of the eddy to the thermal diffusivities can be written as
DT 9 Cµ u0 lP r 9 Pr
= = Cµ ReT , (5.12)
α 4 νP rT 4 P rT
where ν is the mixture kinematic viscosity and ReT = u0 l/ν is the turbulent Reynolds number. The
species effective Lewis number becomes
1 + aRANS ReT
Lei,eff = 1 RANS Re
, (5.13)
Lei + a T
where
9 Pr
aRANS = Cµ . (5.14)
4 P rT
In Eq. 5.13, Lei and ReT are the inputs. Lei is known for all species and can be obtained from a
one-dimensional flame simulation (Table 5.2) and ReT is a known quantity from the DNS parameters.
The parameter aRANS is a function of several quantities. An exact a priori evaluation of this
parameter is difficult. However, an estimate can be provided. First, we set Cµ = 0.09 from a
standard k- model [178]. Second, P r is a function of temperature and mixture composition and
can be obtained from a one-dimensional flame simulation. It is bounded by its value in the unburnt
mixture (P ru = 0.54) and its value in the burnt mixture (P rb = 0.70). Third, the turbulent
(eddy) Prandtl number P rT has to be estimated. Since the thermal and species eddy diffusivities
are assumed equal, it is of the same order of approximation to assume that the eddy viscosity is
also equal to the eddy diffusivities. It is therefore reasonable to assume P rT to be unity. A similar
assumption is made in transported PDF methods [174, 152]. However, other values for the turbulent
Prandtl number have been reported in the literature: in the case of free shear flow, a constant value
of 0.7 for the turbulent Prandtl number was found to give better agreement between experiments
105
and analytical predictions [152]. We therefore assume that P rT is bounded by 0.7 and 1. With these
parameters, an estimated range for the parameter aRANS is obtained:
aRANS ≈ 0.1–0.2. (5.15)
However, it is important to note that a large uncertainty exists on this parameter. Regardless,
this gives the order of magnitude one should obtain for aRANS . Note that this parameter is used as
a fitting coefficient in the following section.
5.2.3 Evaluation of the effective Lewis numbers and model validation
In this section, we evaluate the effective Lewis numbers from the DNS data and compare the results
with the model proposed in the previous section. In order to identify an adequate method for the
evaluation of these effective species Lewis numbers from the DNS data, a sensitivity analysis of the
flamelet structure to the species Lewis numbers is first performed.
5.2.3.1 Sensitivity analysis to species Lewis numbers
The laminar unstretched flamelet profiles are not equally sensitive to all species’ effective Lewis
number. For instance, O2 has a “laminar” Lewis number already close to unity (Lei ≈ 1.2 for
O2 ) and therefore as its effective Lewis number progresses towards unity, very small changes in the
laminar flame profile should be expected. On the other hand, the effective Lewis number of H2 will
change from around 0.3 to unity. Therefore, it is useful to perform a sensitivity analysis first.
One species Lewis number at a time is modified, the other ones being fixed to the species Lewis
numbers obtained for a laminar flame with full transport. Multiple flamelets are simulated for
each modified species, their Lewis number varying from their “laminar” value towards unity. The
resulting H2 and OH mass fractions versus temperature profiles are compared with the DNS profiles.
Figure 5.11 shows these profiles for H2 (top) and OH (bottom) mass fractions for modified Lewis
numbers of H2 , H, and all species but H2 and H, respectively, from left to right. Note the presence
106
0.012 0.012 0.012
flamelets flamelets flamelets
A A A
H2 mass fraction
0.01
{ 0.01
{ 0.01
H2 mass fraction
H2 mass fraction
DNS B DNS B DNS B
0.008 C 0.008 C 0.008 C
D D D
0.006 0.006 0.006
0.004 0.004 0.004
0.002 0.002 0.002
0 0 0
600 1000 1400 600 1000 1400 600 1000 1400
T (K) T (K) T (K)
0.006 0.006 0.006
flamelets flamelets flamelets
0.005 A 0.005 A 0.005 A
{ { {
OH mass fraction
OH mass fraction
OH mass fraction
B B B
DNS DNS DNS
0.004 C 0.004 C 0.004 C
D D D
0.003 0.003 0.003
0.002 0.002 0.002
0.001 0.001 0.001
0 0 0
600 1000 1400 600 1000 1400 600 1000 1400
T (K) T (K) T (K)
Figure 5.11: Comparison of H2 and OH mass fraction profiles between DNS conditional means and
flamelets with modified H2 (left), H (center), and all species but H2 and H (right) Lewis numbers.
DNS from Aspden et al. [10]
of hot spots in Fig. 5.11 (for T > 1400 K). As mentionned in Section 5.2.1, these are beyond the
scope of this thesis and will not be further considered.
Modifying the Lewis number of H2 has by far the most influence, whereas modifying the Lewis
number of H has small effect on H2 mass fraction and small but non negligible effect on OH mass
fraction. Modifying the other species’ Lewis number has negligible influence on the flame structure.
It is not surprising that the Lewis numbers of H2 and H have the most influence. Two explana-
tions are suggested and cannot be differentiated in this study: 1) H2 and H are the species with
Lewis numbers the furthest away from unity and 2) H2 being the limiting reactant and H the most
important radical, their diffusion controls the flame structure.
5.2.3.2 Computing the effective Lewis numbers
Based on the sensitivity analysis performed in the previous subsection, the unstretched laminar
flame structure is very sensitive to the H2 Lewis numbers. Therefore, one should be able to identify
the H2 Lewis number that leads to the best flame structure when compared to each DNS data set.
107
This “optimized” Lewis number is referred to as the effective Lewis number. One such effective
Lewis number can be evaluated for each DNS data set. Unfortunately, the unstretched laminar
flame structure is almost insensitive to the other species Lewis numbers (except H to some extent),
which means that the effective Lewis numbers for these species cannot be identified directly.
As a first step, only the Lewis number of H2 is modified. The L2-norm of the error between
the conditional mean profiles from DNS and the laminar profiles is minimized to obtain the best
estimate of the hydrogen effective Lewis number, i.e., for each DNS case (A through D):
N 2
LeH2 ,eff = arg min N1 YHlam (Le∗H2 , Ti ) − YHDN S
P
2 2
(Ti )
Le∗ i=1
H 2 (5.16)
subject to Le∗H2 ∈ [LeH2 , 1] ,
where YHDN
2
S
(Ti ) corresponds to the interpolated value of < YH2 | T > (from the DNS) at T = Ti ,
and YHlam
2
(Le∗H2 , Ti ) corresponds to the value of YH2 obtained from the laminar unstretched flamelet
simulation with LeH2 = Le∗H2 , interpolated at T = Ti . The temperature is discretized uniformly
from the minimum temperature in the domain Tu to the adiabatic flame temperature Tad such that
i
Ti = Tu + N +1 (Tad − Tu ). Once again, the hot spots are not considered.
Table 5.3 presents the effective Lewis numbers obtained and Fig. 5.12 compares them against the
model presented in Section 5.2.2. The parameter aRANS in Eq. 5.13 was used as a fitting coefficient
and was set to 0.05 to obtain a better agreement with the data. Note that this value is only a factor
of two away from the estimate given by Eq. 5.15. In addition, the value of aRANS does not change
the slope in the semilog plot of Fig. 5.12. With this slight adjustment, the results are in very good
agreement with the model. More specifically, the model captures the evolution of the Lewis number
over the range of Reynolds numbers from its laminar value towards unity.
The proposed model does not make a difference between a species or another, i.e. all species
Lewis numbers should be adjusted. Recognizing that all the parameters in Eq. 5.13 are the same
108
Case A B C D
LeH2 ,eff obtained through Eq. 5.16 (only LeH2 modified) 0.436 0.638 0.795 0.897
LeH2 ,eff obtained through Eq. 5.18 (all Lei modified) 0.446 0.682 0.871 0.995
Table 5.3: Effective Lewis numbers obtained from the DNS (average from three instantaneous snap-
shots separated by several eddy turnover times).
for each species, each Lewis number is modified according to
1+γ
Lei,eff = 1 , (5.17)
Lei + γ
where γ = aRANS ReT . Defining Leeff as the vector containing all species effective Lewis numbers,
this vector is obtained, equivalently to Eq. 5.16, as
N 2
Leeff = arg min N1 YHlam (Le∗ , Ti ) − YHDN S
P
2 2
(Ti )
Le∗
 i=1
 Lej = 1+γ

 ∗
1 , (5.18)
Lej +γ
subject to

 γ ∈ [0, +∞) .

The hydrogen effective Lewis numbers obtained are presented in Table 5.3. They are also compared
against the model in Fig. 5.12. The effective Lewis numbers are virtually the same as obtained
with changing only H2 for the low Reynolds number cases (A and B). Only a small difference is
registered for the higher turbulent cases (C and D). Regardless, the agreement with the model
remains relatively good. It is interesting to note that the slope predicted by the model seems to be
slightly off, which would suggest that the power dependence on ReT in Eq. 5.13 may be incorrect,
or equivalently the Reynolds number may not be the best variable to use in Eq. 5.13.
5.2.4 Model dependence on Reynolds versus Karlovitz number
The model proposed in Section 5.2.2 suggests that the effective Lewis numbers depend on the tur-
bulent Reynolds number. The eddy diffusivity that appears in the equation for the effective Lewis
numbers (Eq. 5.7) was approximated using a k- turbulent model and a model based on ReT was
obtained (Eq. 5.13). One could however argue that the effective Lewis numbers are mostly relevant
109
1
0.9
0.8
2,eff
0.7
LeH
0.6
0.5
0.4 ReT-based model

DNS - only LeH
2
0.3 DNS - all Lei
0 1 2 3
10 10 10 10
ReT
Figure 5.12: Effective Lewis number for H2 versus turbulent Reynolds number obtained by L2-
norm minimization from the DNS where only H2 (black circles) or all Lewis numbers (red triangles)
are modified in the flamelet simulations (three instantaneous snapshots separated by several eddy
turnover times for each case).
within the flame, where large scalar gradients are present. Based on this argument, the flame char-
acteristics should appear in the scaling for the eddy diffusivity. The following subsections consider
and compare different approaches to obtain this scaling and, as a result, a possible dependence of
the proposed model on the Karlovitz number instead of the turbulent Reynolds number is suggested.
5.2.4.1 RANS-based approach
The RANS-based approach used to derive the model (in Section 5.2.2) assumes that the eddy
diffusivity scales as the product of the integral length scale and the velocity fluctuation, i.e.
DT ∝ lu0 . (5.19)
Consequently, using Eq. 5.7, the model for the effective Lewis numbers can be expressed as
1 + aRANS ReT
LeRANS
i,eff = 1 RANS Re
, (5.20)
Lei + a T
110
where a proportionality constant aRANS is used. With this approach, the effective Lewis number
does not depend on any flame characteristic. This result is rather surprising as such a dependency
should be expected.
5.2.4.2 Flame thickness-based approach
Another approach is to assume that the eddy diffusivity is controlled by eddies of the size of the
flame. This leads to the following scaling:
DT ∝ lF uF , (5.21)
where uF is the turnover velocity of an eddy of size lF . This velocity is obtained by [142]
1/3
uF = (lF ) . (5.22)
Using Eq. 5.9, 5.21, and 5.22, one obtains
4/3
lF 0
DT ∝ u (5.23)
l1/3
or
DT
∝ Ka2/3 . (5.24)
α
A model for the effective Lewis numbers follows
1 + alF Ka2/3
Leli,eff
F
= 1 lF 2/3
, (5.25)
Lei + a Ka
where alF is a proportionality constant.

111
5.2.4.3 Time scale-based approach
A third approach is to consider the response time of the flame to unsteady perturbations. This
response time is expected to scale like the flame time tF . Eddies of turnover time corresponding to
the flame time should then be used in the scaling for the eddy diffusivity, i.e.
DT ∝ l(tF )u(tF ), (5.26)
where l(tF ) and u(tF ) are respectively the size and eddy turnover velocity of an eddy of turnover
time tF . Peters [142] derived a mixing length scale based on the quench time and showed that this
quench time is of the order of the flame time scale [138]. Therefore, the scaling results derived here
are equivalent to Peters’ approach.
The relevant eddy size is obtained by
1/2
l(tF ) = t3F . (5.27)
Recall that the flame time scales as

lF2
tF ∝ . (5.28)
α
The turnover velocity u(tF ) is obtained as
1/2
u(tF ) = (tF ) . (5.29)
Using Eq. 5.9, 5.27, and 5.29 in Eq. 5.26, one obtains
3
u0 lF4
DT ∝ . (5.30)
l α2
Equivalently,
DT
∝ Ka2 . (5.31)
α
112
As a consequence, one can obtain a model for the effective Lewis numbers as
1 + atF Ka2
Leti,eff
F
= 1 tF 2
, (5.32)
Lei + a Ka
where a proportionality constant atF is used.
5.2.4.4 Integral length scale limit
Subsections 5.2.4.2 and 5.2.4.3 made use of eddies of specific sizes (lF and l(tF )). However, the
largest size these eddies can be is the integral length scale. Consequently, the length scale lF used
in subsections 5.2.4.2 should be replaced by the smallest value between lF and l, i.e. min (lF , l).
Equivalently, l(tF ) used in subsection 5.2.4.3 should be replaced by min (l(tF ), l). In the limit where
lF > l and l(tF ) > l, the original RANS-based model is recovered for both cases. Nevertheless, it is
important to remember that expressions 5.24 and 5.31 are only scaling arguments and a (unknown)
proportionality coefficient should be used.
For the DNS simulations used in this section, the integral length scale is smaller than both the
flame thickness lF and the length scale l(tF ) introduced in subsection 5.2.4.3. They are, however,
of the same order of magnitude. As a result, it is difficult to assess which length scale is really the
limiting quantity.
Another particularity of the series of simulations performed in Ref. [10] is that the l/lF ratio is
fixed. In other words, we have the following scaling
ReT ∝ Ka2/3 . (5.33)
Consequently, the effects of Ka cannot be differentiated from those of ReT . Using Eq. 5.33, one
could have replaced ReT by Ka2/3 in the ReT -based model (Eq. 5.13 or equivalently Eq. 5.20). Note
that the resulting model would correspond to Eq. 5.25.

113
5.2.4.5 Comparison with DNS results
Based on the observation made in subsection 5.2.3.2, the exponent of ReT in the ReT -based model
was found to be slightly off. Furthermore, it seems reasonable that the effective Lewis numbers
should depend on laminar flame characteristics, as discussed in subsections 5.2.4.1-5.2.4.3. Towards
that end, we have proposed two new scalings based on the Karlovitz number. The tF approach
predicts an exponent of 2 to the Karlovitz number, whereas the lF approach predicts an exponent
of 2/3 (for fixed l/lF the RANS approach also predicts an exponent of 2/3). The predictions with
these models are compared against the effective Lewis numbers presented in Table 5.3 (obtained
from Eq. 5.18), and are depicted in Fig. 5.13. An additional empirical model is considered in which
the ReT is simply replaced by Ka in Eq. 5.13:
1 + aKa Ka
LeKa
i,eff = 1 Ka Ka
. (5.34)
Lei + a
For each of these models, the proportionality coefficient is adjusted to best fit the data (aKa = 0.05).
The best agreement is found to be with an exponent of 1 (Eq. 5.34), which is bounded by the 2/3
and 2 exponents. The agreement with the DNS values is very good and in particular better than for
the ReT -based model. This better agreement with the empirical Ka-based model raises interesting
questions.
However, as mentioned in subsection 5.2.4.4, since the l/lF ratio is fixed in the DNS used, the
Reynolds and Karlovitz numbers are related to each other (Eq. 5.33). As such, the dependence of
the model to the Reynolds or the Karlovitz number cannot be verified. Nevertheless, the Ka-based
model is still prefered due to its more physical nature and better agreement with the DNS data.
5.2.5 Discussion
In this section, the empirical Ka-based model (Eq. 5.34) is considered. As discussed in Section 5.2.4,
this Ka-based model was found to agree the best with the DNS results.
114
1
0.9
0.8
2,eff
0.7
LeH
0.6
0.5
Ka2/3-based model
0.4
Ka2-based model
Ka-based model
0.3 DNS
0 1 2 3 4
10 10 10 10 10
Ka
Figure 5.13: Comparison of the effective Lewis number for H2 versus Karlovitz number for three
different models.
5.2.5.1 Model’s sensitivity to unburnt pressure, temperature, and equivalence ratio
Only four DNS datasets were used to assess the model’s validity. Unfortunately, the number of
available datasets suitable for the present study is limited. Nevertheless, one might wonder what
the model’s sensitivity is to the unburnt equivalence ratio, pressure, and temperature, and in what
context the model could be used. Three points are discussed in this subsection: 1) the form of the
model, 2) the parameter’s sensitivity to the unburnt conditions, and 3) the range of applicability of
the model.
First, the model for the effective species Lewis numbers maps a pair of species Lewis number-
Karlovitz number to an effective species Lewis number. In other words, one can define a function f
such that
Lei,eff = f (Lei , Ka) , (5.35)
where
1 + aKa ξ2
f (ξ1 , ξ2 ) = 1 Ka ξ
. (5.36)
ξ1 + a 2
The function f is derived from ensemble-averaged equations (Eq. 5.5 and 5.6). Its form is therefore
independent from the unburnt equivalence ratio, pressure, and temperature.
Second, while the form of the function is fixed, the parameter aKa in the function f is expected to
115
vary somewhat with the unburnt equivalence ratio, pressure, and temperature. Its sensitivity to the
unburnt conditions is expected to be similar to that of aRANS appearing in the ReT -based model.
This parameter’s sensivity is itself expected to follow closely that of P r (Eq. 5.14), as the other
two parameters, namely P rT and Cµ , come from turbulent models that are independent of these
unburnt conditions. The Prandtl number is the ratio of the thermal diffusivity and the kinematic
viscosity. These two quantities scale the same way with respect to temperature and pressure. More
specifically, we have the following scaling:
ν, α ∝ T 3/2 p−1 . (5.37)
Since the Prandtl number is the ratio of these two quantities, its sensitivity to the unburnt temper-
ature and pressure is expected to be low. Its sensitivity to the unburnt equivalence ratio appears
through the mixing rules for the thermal diffusivity and the kinematic viscosity of the mixture [177].
P r is therefore bounded by the Prandtl number of air and that of the fuel. In summary, we expect
a weak sensitivity to the unburnt equivalence ratio, pressure, and temperature for the parameters
aRANS and aKa .
Third, the model was derived and validated considering only statistically flat, unstretched flames,
and therefore might not be applicable outside this context. However, as mentioned above, the model
should remain valid independently from the unburnt conditions. Furthermore, the model should be
applicable to any range of Karlovitz numbers, as the effective Lewis numbers predicted are bounded
by their laminar values and unity.
5.2.5.2 Practical use of the model
While the model presented in this work was developed from the RANS equations, it may be applied
equally well to LES and DNS.
The results of the previous sections show that the structure of a turbulent flame resembles that
of a laminar flame (from a statistical point of view) with an appropriate choice of effective Lewis
numbers. In other words, all chemical quantities may be expressed as a function of a single progress
116
variable (such as temperature or product mass fraction). Such methodology is often refered to as
tabulated chemistry. The model presented in this chapter should be used in the generation of such
a chemistry table.
Lewis numbers appear at two places in the simulation of premixed flames with tabulated chem-
istry (either RANS, LES, or DNS): 1) in the transport equations for the species in the one-
dimensional, unstretched laminar flame simulations (used to tabulate the chemistry), and 2) in
the transport equation for the progress variable in the computational fluid dynamics (CFD) code.
The model presented in this chapter is only relevant to the Lewis numbers that appear at the first
place (1D flame simulations for tabulation).
Note that additional impacts of the effective species Lewis number model are detailed in Ap-
pendix A.
5.3 Use of the effective species Lewis number model with the
n-heptane flame
In this section, the effective species Lewis number model presented in the previous section is tested
on the non-unity n-heptane/air flame. First, the Ka-based model (Eq. 5.34) is used to compute a set
of effective species Lewis numbers. The same proportionality coefficient as used in Section 5.2.4.5,
aKa = 0.05, is considered. Note that given the definition of the Karlovitz number used in Section 5.2,
a Karlovitz number of 73 is obtained for the n-heptane/air flame. As such, the effective Lewis number
for the fuel is 1.16. The solution of an unstretched laminar flame with this set of effective species
Lewis numbers is obtained with FlameMaster. The species mass fraction vs. temperature profiles
are compared, in Fig. 5.14, to the corresponding conditional means from the DNS for the fuel, C2 H4 ,
and CO2 . Although differences are noticeable, especially for the product, the effective Lewis number
model captures fairly well the effect of turbulent mixing on the mean flame structure. It is important
to highlight the fact that the model was proposed/validated with a series of lean hydrogen/air flames,
117
0.06 0.012
<Y|T> <Y|T>
flamelet - non-unity Le flamelet - non-unity Le
0.05 0.01

flamelet - unity Le flamelet - unity Le
C2H4 mass fraction

flamelet - Leeff flamelet - Leeff
0.04 0.008
0.03 0.006
0.02 0.004
0.01 0.002
0 0
400 600 800 1000 1200 1400 1600 600 1000 1400 1800 2200
T (K) T (K)
0.16
<Y|T>
0.14 flamelet - non-unity Le
flamelet - unity Le
CO2 mass fraction
0.12 flamelet - Leeff
0.1
0.08
0.06
0.04
0.02
0
600 1000 1400 1800
T (K)
Figure 5.14: Conditional mean (solid line) of the n-C7 H16 (top left), C2 H4 (top right), and CO2
(bottom) mass fraction vs. temperature. The non-unity, the unity, and the effective Lewis number
flamelet solutions are also shown (dashed line).
but is also relevant to the present n-heptane/air flame, without any modification needed.
Second, the procedure described in Section 5.2.3.2 is followed, i.e. Eq. 5.18 is solved considering
the n-heptane vs. temperature profile. The results are compared in Fig. 5.15 to those obtained for
the hydrogen/air flames. Since both fuels have very different Lewis numbers, the effective diffusivity
is presented instead (corresponding to γ = DT /α in Eq. 5.18). This diffusivity is also computed
for the lower-Ka flame (introduced in Chapter 4). Independently of the flame, the data is in fairly
good agreement with the model, at least for moderate Ka, corresponding to the thin reaction zones
regime.
118
10000
H2
1000 n-C7H16
0.05Ka
100
DT/α
10
0.1
10 100 1000 10000
Ka
Figure 5.15: Normalized effective diffusivity obtained by solving Eq. 5.18 for the respective fuels
3/2 −1/2
(hydrogen and n-heptane) vs. Karlovitz number (Ka = (u0 /SL ) (l/lF ) to be consistent with
Ka
the analysis in Section 5.2). The solid line represents DT /α = a Ka (corresponding to Eq. 5.34)
with aKa = 0.05.
5.4 Limitations of the model
As mentioned in Section 5.1.2, a single set of effective species Lewis numbers is not sufficient to
entirely capture the mean structure of the non-unity Lewis number n-heptane/air flame, as shown
in Fig. 5.14. We argued that this is due to the evolving nature of turbulence through the flame.
As such, the “local DT ” should vary accross the flame. Again, it is clear from Fig. 5.14 that the
flame structure is more unity Lewis number-like at lower temperatures than higher temperatures.
Modeling this “local DT ” with scaling arguments (as was done is Section 5.2.4) is not an easy task
since relevant local scales need to be identified.
Finally, it was argued in Section 5.1.1 that the use of a progress variable with tabulated chemistry
could be justified in the unity Lewis number flame, since the species mass fractions can be mapped
unto a single progress variable. Similarly, it was discussed in Section 5.2.5.2 that a progress variable
approach could be used for the hydrogen/air flames analyzed in Section 5.2, as long as the proper
effective species Lewis numbers are considered. However, it is unclear how this would apply to the
non-unity Lewis number n-heptane/air flame, in which a single set of effective species Lewis numbers
cannot properly characterize both the preheat and the reaction zones. This will be discussed in more
details in Chapter 7.
119
Chapter 6
Characterization of the reaction

zone1
In this chapter, a characterization of the reaction zone is provided. The effects of turbulence on the
reaction zone are assessed in the first section. The mechanism by which stretching affects the fuel
consumption rate is then investigated. In the third section, the turbulent flame speed is evaluated
and the mechanism through which turbulence affects this flame speed is described. Finally, a short
discussion is provided.
All throughout this chapter, the effects of differential diffusion are highlighted by comparing
the results between the unity and the non-unity Lewis number flames. Note that the results from
the previous chapter indicate that effects of differential diffusion should not be fully suppressed by
turbulent mixing at the reaction zone.
6.1 Turbulent reaction zone
The first subsection recalls some of the observations made in Section 4.4.2 on the turbulent reaction
zone and provides quantitative information on the fuel source term fluctuations, and the effect of
differential diffusion is highlighted. In the second subsection, the structure of the reaction zone in
temperature space is analyzed.

1 The work presented in this chapter is published in B. Savard, B. Bobbitt, and G. Blanquart, Proc. Comb. Inst.
(2015) and in B. Savard and G. Blanquart, Combust. Flame (2015). The author of this thesis has contributed to all
the work presented in this chapter.
120
6.1.1 Source term fluctuations and differential diffusion effect
Figure 6.1(a) shows contours of the source term of n-C7 H16 for the unity Lewis number flame.
Similarly, Fig. 6.1(b) shows the same plot but for the non-unity Lewis number flame. For both
flames, the turbulent reaction zone is locally of the same thickness as that of a one-dimensional
flat flame (as already shown in Section 4.4.2 for the non-unity flame). These plots suggest that
the flames belong to the thin reaction zones regime. As mentioned already, the Kolmogorov length
scale increases through the flame due to the increasing kinematic viscosity, and becomes as large
as the laminar reaction zone thickness as it reaches the burnt side. Consequently, turbulence has
limited impact on the reaction zone thickness. The location in temperature space of the reaction
zone is also virtually unaffected. This is evidenced in Fig. 6.1(a) and 6.1(b) by the solid black lines
which are isocontours of the respective temperatures of peak source term in a one-dimensional flame
(more details in the next section). Nevertheless, the species source terms fluctuate by up to 100%
around the laminar value on these isocontours, as highlighted in Fig. 6.1(c). This figure shows both
source terms of n-C7 H16 , normalized by their peak value in a corresponding laminar flame, extracted
along the isocontour corresponding to the temperature of that peak source term, Tpeak . Note that
Tpeak = 1364 K for the unity Lewis number flame, and Tpeak = 1240 K for the non-unity Lewis
number flame. These fluctuations are suspected to be a consequence of stretching, as observed in
Ref. [71, 22] for lower Karlovitz number flames. This will be assessed in Section 6.2.
Interestingly, while little deviations from the laminar flame structure were found (Fig. 5.2) in
the unity Lewis number flame, the source terms fluctuate by up to 100% around the laminar value
(Fig. 6.1(c)). Although the species mass fractions correlate very well with temperature, small devia-
tions from the laminar profiles are present (see Fig. 5.2), especially around the peak fuel consumption
temperature (T = 1300 K). These small absolute deviations from the laminar profiles can result in
large relative fluctuations and hence large fluctuations in the chemical source terms.
Although source term fluctuations are present for both flames, extinction events seem to be more
present with differential diffusion (Fig. 6.1(b) and Figure 6.1(c)). This is confirmed by Fig. 6.2,
which presents the distribution of the normalized fuel consumption rate on the isocontour T = Tpeak
121
2.5
unity Le
7H16,lam
2 non-unity Le
1.5
/ωC
1
7H16
0.5
ωC
0
0 2 4 6 8
arc length (mm)
(a) ω̇C7 H16 ; Tpeak = 1364 K (b) ω̇C7 H16 ; Tpeak = 1240 K (c) ω̇C7 H16 at Tpeak
Figure 6.1: (a,b) Contours of n-C7 H16 source term normalized by its peak laminar value on a two-
dimensional horizontal slice. Also shown are three temperature isocontours: 600 K (white, left of the
reaction zone), temperature of peak source term Tpeak (black), and 1850 K (white, right of reaction
zone). The laminar reaction zone thicknesses of n-C7 H16 , δC7 H16 is also shown for comparison. (c)
Normalized n-C7 H16 source term vs. distance along the isocontour T = Tpeak .
(surface in three dimensions) for both flames. The distributions are surface-area weighted and
averaged over 10 eddy turnover times. While both distributions have similar shapes, a clear shift to
smaller fuel consumption rates is observed with the presence of differential diffusion. Local extinction
events are found in the non-unity Lewis number flame, but are absent in the unity Lewis number
flame. This is further evidence that the effective diffusivity is not sufficiently high at the reaction
zone to suppress the effects of differential diffusion.
1
unity Le
0.9 non-unity Le
Probability density
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5 3 3.5
Source term ωC /ωC
7H16 7H16,lam
Figure 6.2: Probability density, at T = Tpeak , of n-C7 H16 source term, ω̇C7 H16 , normalized by its
peak laminar value for both the non-unity and the unity Lewis number cases.
It is important to note that the reaction zone is thin and broken at several locations, but not
distributed. It may be argued that the Karlovitz number is not sufficiently high to lead to a
122
distributed reaction zone. Interestingly, Aspden et al. [10] have shown that a lean H2 /air flame
could transition from a thin reaction zone to a distributed reaction zone behavior (as the Karlovitz
number is increased) without showing signs of a broken reaction zone. These different observations
may be due to the fact that H2 has a Lewis number less than unity (LeH2 ≈ 0.3), whereas n-C7 H16
has a greater than unity Lewis number (LeC7 H16 = 2.8). A more diffusive flame may be expected to
be more resistant to turbulent effects/perturbations.
6.1.2 Fuel consumption at Tpeak
Although the extinction events are linked to differential diffusion, it is not clear through which
mechanism. Many authours have previously suggested that extinctions (and, more generally, source
term fluctuations) can be a consequence of stretching [71, 22]. These possible correlations (on
curvature and strain rate) will be investigated in Section 6.2.3. Before performing such analysis, it
is practical to reduce the analysis of the reaction rate from a volumetric (3D) analysis to a surface
(2D) analysis.
The integrated fuel consumption speed has often been used in the literature and would be ideal
to identify such extinction events [53]. This method requires the integration of the fuel consumption
term in a direction normal to the isoterm corresponding to T = Tpeak . Unfortunately, since the
reaction zone is highly curved and sometimes “folded” on itself, the fuel consumption term cannot
be computed with both significance and robustness for the present flames. As a consequence, a more
local quantity would be preferred.
Figure 6.3(a) shows the joint probability density of the normalized fuel consumption rate as a
function of temperature for the unity Lewis number flame. The mean of the source term conditional
to temperature is also shown. This profile is compared to its unstretched one-dimensional flame
equivalent. While the mean profile is very close to that of the one-dimensional flame, fluctuations
around this mean are relatively large. Alternatively, Fig. 6.3(b) presents the joint probability density
of the fuel consumption term away from the flame front (at local temperature T ) normalized by the
consumption term of the closest point on the flame front (i.e. normaly projected on the iso-surface
123
0.0003 0.01
1.4 <ω|T> 1.4 flamelet
flamelet
ωC H /ωC H ,lam(Tpeak)
1.2 1.2 0.008
ωC H /ωC H (Tpeak)
1 0.0002 1
0.006
7 16
7 16
0.8 0.8
0.6 0.6 0.004
7 16
7 16 0.0001
0.4 0.4
0.002
0.2 0.2
0 0 0 0
600 1000 1400 1800 2200 600 1000 1400 1800
T (K) T (K)
(a) Le = 1 (b) Le = 1; normalized by closest point on

isoterm T = Tpeak
1.8 0.01
flamelet - non-unity Le
1.6 flamelet - unity Le
0.008
ωC H /ωC H (Tpeak)
1.4
1.2
0.006
7 16
0.8
0.004
7 16
0.6
0.4 0.002
0.2
0 0
600 1000 1400 1800
T (K)
(c) Le 6= 1; normalized by closest point on

isoterm T = Tpeak
Figure 6.3: (a)Joint PDF and conditional mean (solid line) of n-C7 H16 source term, ω̇C7 H16 , normal-
ized by its peak laminar value vs. temperature from the unity Lewis number simulation. Joint PDF
of n-C7 H16 source term, ω̇C7 H16 , normalized, in the direction normal to the isoterm T = Tpeak , by
its value on this isoterm vs. temperature for the unity Lewis number (b) and the non-unity Lewis
number DNS (c). The non-unity and unity Lewis number flamelet solutions are also shown (dashed
line).
T = Tpeak ). As can be observed, the fluctuations are significantly reduced now. Figure 6.3(c) shows
the same normalized source term for the non-unity Lewis number case. These plots suggest that
fuel consumption locally scales like its value at Tpeak , for both the unity and the non-unity Lewis
number flames, i.e.

ω̇C7 H16 ,lam (T )
ω̇C7 H16 (T ) ≈ ω̇C7 H16 (Tpeak ) . (6.1)
ω̇C7 H16 ,lam (Tpeak )
Therefore, only the fuel conssumption term at T = Tpeak will be further considered. Figure 6.4
124
presents three-dimensional views of this reaction surface, i.e. the isoterm corresponding to T = Tpeak ,
colored by the normalized fuel consumption rate.
(a) Le = 1; unburnt side (b) Le = 1; burnt side
(c) Le 6= 1; unburnt side (d) Le 6= 1; burnt side
Figure 6.4: Contours of ω̇C7 H16 normalized by its laminar value on the isosurface corresponding to
T = Tpeak . (a,b) Le = 1; (c,d) Le 6= 1. (a,c) View from unburnt side; (b,d) view from burnt side.
6.2 Stretching
In order to assess if the fuel consumption rate fluctuations and the local extinctions in the presence
of differential diffusion are caused by stretching, it is interesting to first analyze the distribution of
curvature and strain rate at the reaction zone. Then, the correlation between fuel consumption rate
and strain rate or curvature will be tested.

125
6.2.1 Curvature and strain rate distributions
The mean curvature is defined as κ = κ1 + κ2 , where κ1 and κ2 are the principal curvatures of the
reaction surface. The curvature is defined to be positive when the center of curvature is oriented
towards the burnt gas. It is expressed as
κ = ∇ · n, (6.2)
where n = −∇T /|∇T | is the normal to the reaction zone surface, and is positively oriented towards
the unburnt gas. As previously performed in Ref. [53], the normal is computed everywhere in the
domain and is interpolated on the reaction surface. The probability density distribution of curvature
(surface area weighted and averaged over 10 eddy turnover times) is presented in Fig. 6.5(a) for
both the unity and the non-unity Lewis number cases. Note that the mean curvature is normalized
by the laminar fuel reaction zone thickness δC7 H16 (full width at half-height). For both cases,
δC7 H16 = 0.07 mm lF .
First, it is important to note that the magnitude of curvature is very large. More specifically,
the radius of curvature is (with relatively large probability) much smaller than the laminar flame
thickness and even as small as the reaction zone thickness. Such intense curvatures are difficult to
observe in laminar flames. Second, consistent with the observation that the turbulent reaction zone
remains thin for both flames, Fig. 6.5(a) suggests that only velocity perturbations larger than the
laminar reaction zone actually reach this reaction zone. Third and most importantly, no differential
diffusion effect is visible in Fig. 6.5(a), suggesting that the geometry of the reaction surface is only
influenced by turbulence. Note that this might not remain valid for all fuels and all Karlovitz
numbers, as thermo-diffusive and Darrius-Landau instabilities might have an effect on the geometry
of the reaction surface [53].
The strain rate tangential to the reaction surface is computed as, similarly to Refs. [71, 34],
at = ∇ · u − n · ∇u · n, (6.3)
126
1.6 0.12
unity Le unity Le
1.4 non-unity Le non-unity Le
0.1
Probability density
Probability density
1.2
0.08
1
0.8 0.06
0.6
0.04
0.4
0.02
0.2
0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -10 -5 0 5 10 15 20
Mean curvature κδC H
7 16
Strain rate atδC H /S0L
7 16
(a) Curvature (b) Strain rate
Figure 6.5: Probability density function, on the reaction surface, of κ, normalized by the laminar
reaction zone thickness (top), and at , normalized by the ratio of the laminar reaction zone thickness
to the laminar flame speed (bottom).
where u is the velocity. Note that in the limit of an infinitely thin flame, the strain rate is computed
in the unburnt gas, where the flow is divergence free, and the tangential strain rate is equal to
the normal strain rate. Figure 6.5(b) shows the probability density of the normalized strain rate
tangential to the reaction surface. The magnitude of the strain rate, up to 70,000 s−1 , is very large
(the s-curve turning point of a methane/air premixed flame, with φ = 0.7, Tu = 800 K, and P0 = 1
atm in a back-to-back counterflow configuration was found numerically to be approximately 10,000
s−1 [96]). Again, the strain rate applied to both the unity Lewis number and the non-unity Lewis
number reaction surface are similar. The strain rate for the non-unity Lewis number flame is only
marginaly smaller than for the unity Lewis number case. Consistent with values previously-reported
by several authors for methane/air and hydrogen/air flames in the thin reaction zone [34, 73, 33, 82],
a positive mean tangential strain rate is observed.
With the information presented in this section, it is interesting to go back to Eq. 1.16, which can
be rewritten as

SL L at δC7 H16
=1− κδC7 H16 + . (6.4)
SL0 δC7 H16 SL0
Under similar unburnt conditions (but with an unburnt temperature of 353K), L ≈ 1mm [92],
which means L/δC7 H16 ∼ O(10). Figure 6.5 shows that κδC7 H16 ∼ O(0.1)–O(1),and at δC7 H16 /SL ∼
127
O(1)–O(10), which means that L (κ + at /SL ) 1 and Eq. 6.4 does not hold anymore.
6.2.2 Propagating vs. material surface
The results from the last two sections suggest that differential diffusion virtually does not affect the
reaction zone geometry. This is a consequence of the high Karlovitz number and goes along the
results from Bradley et al. [22], who observed experimentally a suppresion of thermodiffusive effects
at high turbulence levels. This can be explained by an analogy to the work of Yeung et al. [185]
who analyzed theoretically and numerically the geometry of material and propagating surfaces in
the presence of homogeneous isotropic turbulence.
Material surfaces are composed of convected fluid particles. Propagating surfaces have an intrin-
sic local propagation speed w. For a material surface, w is identically zero. In Yeung et al. [185], a
premixed flame is seen as a propagating surface, with w = SL . Note that their work is different from
that of Ashurst et al. [7], who studied the orientation of passive scalar gradients in homogeneous
isotropic turbulence (and shear layers). An iso-surface of a passive scalar is neither a material nor
a propagating surface.
Yeung et al. [185] found that a material surface orientates preferentially with the strain rate
tensor. They identified a universal distribution of strain rate on the material surface. In particular,
they observed that 80% of the strain rate is positive. They further argued that a propagating surface
has the same strain rate distribution as a material surface, independently of the turbulent Reynolds
number, under the condition w/vη 1, where vη is the Kolmogorov velocity.
For the present flames analyzed, 81% and 82% of the strain rate was found to be positive for the
unity and the non-unity Lewis number flames, respectively. These results are consistent with the
ratios SL0 /vη as listed in Table 6.1.
In order to further illustrate the difference between propagating and material surfaces, the distri-
butions of tangential strain rate on the reaction surface in the series of DNS of lean H2 /air turbulent
flames conducted by Aspden et al. [10] are presented in Fig. 6.6. As shown in Table 6.1, the Karlovitz
number ranges from 28 to 4200 (consistently with the definition provided in the introduction). The
128
associated SL0 /vη ratios are also listed in Table 6.1. It can be seen that, even though the gas mix-
tures are different and exhibit Markstein lenghts of opposite signs, the highest Ka H2 /air flame and
the present n-C7 H16 /air flame show more similar distributions than between the two H2 /air flames.
This is consistent with the values of SL0 /vη (shown in Table 6.1) and is due to the intense turbulence
ahead of the reaction zone.

0.6
n-C7H16 (Ka=220)
H2 (Ka=28)
0.5
Probability density
H2 (Ka=4200)
0.4
0.3
0.2
0.1
0
-2 -1 0 1 2 3 4
Normalized strain rate
Figure 6.6: Probability density function, on the isoterm T = Tpeak , of at , normalized by the standard
deviation of the distributions. Distributions from the H2 /air flame of Aspden et al. [10] with Ka = 28
and Ka = 4200 and from the non-unity Lewis number n-C7 H16 /air presented in this thesis.
0
Fuel Ka SL /vη % at > 0
n-C7 H16 /air (Le = 1) 280 0.08 81
6 1)
n-C7 H16 /air (Le = 220 0.11 82
H2 /air 28 0.30 80
H2 /air 4200 0.02 80
Table 6.1: SL0 /vη ratios (see Ref. [185]) and curvature and strain rate statistics for the flames
presented in this thesis and for two lean H2 /air (φ = 0.4) flames presented in [10]. vη is evaluated
at Tpeak .
The overall conclusion of this subsection and the previous one is that the reaction zone surface
behaves as a material surface under the intense turbulent field. Even if the local consumption rates
are affected, they are too weak to alter the shape of the flame front. As such, the distributions of
curvature and strain rate are independent from differential diffusion effects. Note that curvature
and strain rate are also found to be very weakly dependent on one another in both flames (see
Appendix B).
129
6.2.3 Fuel consumption rate vs. curvature and strain rate
Figures 6.7(a) and 6.7(b) present the joint probability densities of the normalized fuel consumption
term vs. strain rate on the isoterm T = Tpeak , for both the unity Lewis number and the non-unity
Lewis number flames (again, the data is averaged over 10 eddy turnover times). Pearson’s correlation
coefficient and the distance correlation (these two metrics are detailed in Appendix B) are listed in
Table 6.2 for these two figures. A stronger correlation (yet still weak) with strain rate is found for
the unity Lewis number case. For the non-unity Lewis number case, the source term appears to be
more independent from the local strain rate. As a result, strain rate can hardly explain the presence
of local extinctions in the non-unity Lewis number case as it can be observed in Fig. 6.7(b).
The sign of the correlation found for both cases (non-unity and unity Lewis number) does not
correspond to what would be expected from Eq. 1.16 given the positive sign of the Markstein length.
Similarly, Hawkes and Chen [71] have found that, for a lean methane/air mixture, the dependence of
the local flame consumption speed (integrated fuel consumption rate normal to the reaction surface)
on tangential strain rate is of opposite signs between a laminar flame (exposed to low strain rates)
and a high Karlovitz (Ka ≈ 300, two-dimensional) flame. A similar behavior was also reported
by Haworth and Poinsot, albeit for a low Karlovitz number, two-dimensional flame with one-step
chemistry [73].
Figures 6.8(a) and 6.8(b) present, similarly to Fig. 6.7(a) and 6.7(b), the joint probability den-
sities of the normalized fuel consumption term vs. curvature on the isoterm T = Tpeak . Again, the
correlation coefficients can be found in Table 6.2. This time, there is a much stronger correlation
between the fuel consumption rate and the curvature with the presence of differential diffusion. Note
that the use of the distance correlation is particularily relevant for this case, as the relation between
the fuel consumption rate and curvature is clearly non-linear. Although Lκ > 1, which implies that
Eq. 6.4 is not valid throughout the range of curvatures found, the sign of the correlation found is
consistent with Eq. 6.4.
A correlation between the local flame consumption speed and curvature has also been reported
previously in turbulent lean hydrogen/air flames [53]. Consistent with the present results, the
130
2.5 0.12 2.5 0.12

2.5 0.12
7H16,lam
,lam
16,lam
0.1 0.1
77 16
2 0.1 2
/ωCC HH
2
/ωC
0.08
term ωωCC HH /ω 0.08 0.08
7H16
16
77 16 1.5
1.5 1.5
0.06
Source term ωC
0.06 0.06
1
1 1
Source term
0.04
0.04 0.04
Source
0.5 0.5
0.02 0.02
0 0 0 0
-10 -5 0 5 10 15 20 -10 -5 0 5 10 15 20
Strain a δ H /SL 0 /S0L
tδ7C 16
Strain rate taC H /S L
Strain rate atδC
7 16 7H16
(a) Le = 1 (b) Le 6= 1
Figure 6.7: Joint probability density function, on the isoterm T = Tpeak , of n-C7 H16 source term,
ω̇C7 H16 , normalized by its peak laminar value vs. the normalized tangential strain rate. (a) Le = 1;
(b) Le 6= 1.
correlation was found to be positive, since such flames are associated with a negative Markstein
length.
The fact that the signs of the correlations between fuel consumption rate and curvature are
consistent between laminar and turbulent flames is, however, not trivial. More specifically, it has
been previously reported that, at sufficiently high Karlovitz numbers, differential diffusion effects
are suppressed (see Ref. [142] and Chapter 5). However, the turbulence intensity is not sufficiently
high to suppress the differential diffusion effects in the reaction zone for the present flame. This was
already evidenced by the characterization of the flame structure (Chapter 5).
r dCorn
Strain / Le = 1 0.61 0.57
Strain / Le 6= 1 0.49 0.45
Curvature / Le = 1 -0.12 0.33
Curvature / Le 6= 1 -0.53 0.65
Table 6.2: Pearson’s correlation coefficient, r, and distance correlation, dCorn , between strain rate
and fuel consumption, and between curvature and fuel consumption at T = Tpeak for both flames
presented the unity and the non-unity Lewis number flames.
It is not surprising that the correlations listed in Table 6.2 are relatively weak. First, as mentioned
earlier, Eq. 1.16 was derived for much weaker strain rates and curvatures. Second, unsteady effects
131
2.5 2 2
7H16,lam
7H16,lam
1.4
2 1.2
1.5 1.5
/ωC
/ωC
1
7H16
7H16
1.5
Source term ωC 0.8
Source term ωC
1 1
1 0.6
0.4 0.5 0.5

0.5
0.2
0 0 0 0
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
Mean curvature κδC Mean curvature κδC
7H16 7H16
(a) Le = 1 (b) Le 6= 1
Figure 6.8: Joint probability density function, on the isoterm T = Tpeak , of n-C7 H16 source term,
ω̇C7 H16 , normalized by its peak laminar value vs. the normalized mean curvature. (a) Le = 1; (b)
Le 6= 1.
may also have a strong impact. For instance, considering much weaker stretch, linear theory [84, 42]
and results from unsteady counterflow flames [81] have shown an attenuation of Markstein number
with the frequency of strain rate fluctuations. Unsteady effects in the response of chemistry to stretch
have also been observed experimentally and computationally in flame-vortex interaction [120, 127].
Hence, a loss of correlation with instantaneous stretch is expected for highly turbulent flames [35, 36].
To summarize, it was found that the fluctuations in fuel consumption rate are, to some extent,
attributed to strain rate for the unity Lewis number flame, and curvature for the non-unity Lewis
number flame. These observations (for premixed flames at the edge of the broken reaction zones
regime) are consistent with the results of Haworth and Poinsot (for premixed flames at the edge
of the corrugated flamelets regime) [73]. The local extinctions were attributed to high curvatures
through differential diffusion for the non-unity Lewis number flame. However, the correlations found
may be too weak to be useful for modeling purposes (more details in Chapter 7).
6.2.4 Fuel consumption rate limiting species vs. curvature and strain rate
Given the chemical mechanism used in the simulations, the fuel is consumed through six reactions.
The following two account for 85% (unity Le) and 90% (non-unity Le) of the overall fuel consumption
132
rate:
n–C7 H16 + H → 2–C7 H15 + H2 , (6.5)
n–C7 H16 + OH → 2–C7 H15 + H2 O.
Consequently, at the reaction zone (fixing T = Tpeak ), the burning rate is mostly a function of YC7 H16 ,
YH , and YOH . It is interesting to identify which of these species are responsible for the correlations
found in the previous section. Their respective correlations with strain rate and curvature are
presented in Table 6.3.
In the unity Lewis number case, all three species mass fractions show similar correlations with
strain rate and curvature as those obtained for the fuel burning rate (Table 6.2). This is not
surprising. In the absence of differential diffusion, all chemical species react to strain and curvature
the same way.
In the non-unity Lewis number case, the fuel mass fraction shows the largest correlation with
curvature (the joint probability density function is shown in Fig. 6.9) and becomes decorrelated with
strain rate. This is qualitatively consistent with the defocusing/focusing differential diffusion effects
in laminar flames, which have been widely studied [133, 24, 43]. On the other hand, both radicals
show only little correlation with curvature (more for H than OH) and retain their correlation with
strain rate.
In Ref. [58], for a two-dimensional stoichiometric methane/air premixed flame in the thin re-
action zones regime, Echekki and Chen found a stronger correlation of the hydrogen atom mass
fraction with curvature than with strain rate. They attributed this dependence on curvature to
focusing/defocusing differential diffusion effect (H has a very low Lewis number). In a similar way,
for the present flames, the correlation of hydrogen atom mass fraction with curvature increases from
the unity Lewis to the non-unity Lewis number flame (as opposed to OH mass fraction, which has a
Lewis number much closer to unity). However, this increase in correlation is overshadowed by that
of the fuel mass fraction with curvature, which was not observed in Ref. [58]. It is important to
133
remark that the Lewis numbers of the reactants in Ref. [58] are very close to unity, while that of
the limiting reactant in the present flame is much greater than unity (LeC7 H16 = 2.84). Important
qualitative differences are expected between such flames [36].
Le = 1 Le 6= 1
r dCorn r dCorn
YC7 H16 vs. strain 0.50 0.49 0.16 0.22
YH vs. strain 0.59 0.56 0.52 0.51
YOH vs. strain 0.60 0.57 0.53 0.49
YC7 H16 vs. curvature -0.10 0.34 -0.69 0.74
YH vs. curvature -0.15 0.35 -0.33 0.45
YOH vs. curvature -0.16 0.36 -0.14 0.31
and species mass fractions, and between curvature and species mass fractions at T = Tpeak for both
(higher-Ka) flames presented in this thesis.
2 3
7H16,lam
2.5
/YC
1.5
2
7H16
Mass fraction YC
1 1.5
1
0.5
0.5
0 0
-1.5 -1 -0.5 0 0.5 1 1.5
Curvature κ δC
7H16
Figure 6.9: Joint probability density function, on the isoterm T = Tpeak , of n-C7 H16 mass fraction,
YC7 H16 , normalized by its peak laminar value vs. the normalized mean curvature for the non-unity
Lewis number flame.
6.3 Turbulent flame speed
It was shown in the last two sections that the fuel consumption rate was affected by differential
diffusion, whereas the curvature distribution on the reaction surface, characteristic of the geometry
of this reaction surface, was virtually unaffected by it. In the present section, a link between these
two results and the turbulent flame speeds of both flames is discussed.
134
6.3.1 Computing the turbulent flame speed
The turbulent flame speed may be expressed as
Z t2
1
ST = −ṁF (t) dt, (6.6)
ρu YC7 H16 ,u L2 (t2 − t1 ) t1
where the subscript u is used for unburnt conditions, t2 − t1 is the data collection time, and ṁF is
the instantaneous fuel consumption rate, obtained through the following:
Z
ṁF (t) = ω̇C7 H16 (x, y, z, t) dV, (6.7)
Ω
with dV an element of volume, and Ω the whole domain. The values obtained are presented in
Table 6.4.
As mentioned in Section 4.1.1, these turbulent flame speeds do not correspond exactly to the inlet
bulk velocity (set to 1 m/s). As a result, the flame drifts from its initially position with a relative
speed corresponding to the difference between the inlet bulk velocity and the effective turbulent
flame speed. This computed drift (linear in time) is compared, in Fig. 6.10, to the instaneous flame
drift computed as
∆xf = xf − x0 , (6.8)
where x0 is the position of the flame at time t = 0 and xf is the instantaneous position of the flame,
computed as
Z
1
xf (t) = 11L − c (x, y, z, t) dV, (6.9)
L2 cb Ω
where c = YH2 + YH2 O + YCO + YCO2 is a progress variable, which takes the value cb in the burnt
gas. The agreement between the drifts computed by both method is very good, suggesting that the
turbulent flame speeds presented in Table 6.4 are good estimates over the data collection period.
The difference in the ST /SL ratios (between unity and non-unity Lewis numbers) highlights an
important differential diffusion effect (the only difference between the flames). The unity Lewis
number flame encounters a significantly larger increase in flame speed than the non-unity Lewis
135
0.25
0.2 unity Le
non-unity Le
0.15
ST=0.7 m/s
0.1
∆xf/L
0.05
0 ST=1.1 m/s
-0.05
-0.1
0 2 4 6 8 10 12 14 16
t/τ
Figure 6.10: Flame drift over time. Recall that the x-axis is positively oriented downstream. A
constant inflow of 1 m/s was imposed at the inlet. The straight black lines correspond to the
turbulent flame speeds obtained from Eq. 6.6.
number flame, although both flames are subjected to the same incoming turbulent flow. As observed
by many authors [172, 99, 94, 100, 50, 69, 21, 125], this further suggests that differential diffusion
has to be taken into account in turbulent flame speed models (e.g. in the form of a Lewis number
dependence). More details can be found in Appendix A.3.
6.3.2 Link between differential diffusion effects and turbulent flame speed
Although the latter observation is interesting in itself, a better understanding of where this differ-
ence (in turbulent flame speed) comes from is necessary for modeling purposes. Using continuity,
Damköhler proposed that, for a (thin) corrugated flame, the turbulent flame speed is proportional
to the turbulent flame surface area AT , or ST /SL = AT /A, with A the cross-section area [48]. In
this regime, it is often assumed that the local flame speed is equal to the laminar flame speed. In
contrast, for the thin reaction zones regime, Damköhler suggested that the turbulent flame speed
should, instead, be proportional to the turbulent diffusivity [48].
Even if the current flames are not in the (thin) corrugated flamelet regime, the reaction zone
remains thin and is only weakly corrugated. As a consequence, a reaction surface (isocontour
T = Tpeak , as introduced in Section 6.1.2), as opposed to a flame surface, and a local consumption
speed can still be defined. For the present analysis, a similar approach to that used by Damköhler
136
for the corrugated flamelet regime is taken. This time, the local consumption speed is not fixed to
the laminar value.
First, because the reaction zone is only weakly corrugated, one can rewrite exactly Eq. 6.7 as an
integral along the isocontour of T = Tpeak (surface integral) and an integral in the normal direction,
i.e.
Z Z
ṁF (t) = ω̇C7 H16 dn dA, (6.10)
T =Tpeak
R
with ω̇C7 H16 dn being the local consumption mass flow rate. Recalling that n = −∇T /|∇T |, the
latter can be rewritten as

Z Z
dT
ω̇C7 H16 dn = ω̇C7 H16 . (6.11)
|∇T |
Second, as previously noted, the reaction zone remains thin, which means that the temperature
gradients at the reaction zone are approximately equal to those of a laminar flame, i.e. |∇T | ≈
|∇Tlam | and
Z Z
dT
ω̇C7 H16 dn ≈ ω̇C7 H16 . (6.12)
|∇Tlam |
Following Damköhler’s argument, the reaction zone thickness scales with the square root of the
effective diffusivity (molecular plus turbulent). Therefore, the previous observation of a thin reaction
zone is consistent with a negligible turbulent diffusivity at the reaction zone (with respect to the
characteristic scales of the reaction zone). As discussed in Section 4.4.1, the intensity of the turbulent
flow field decreases dramatically from the preheat zone to the reaction zone.
Third, as shown in Fig. 6.3(b) and 6.3(c), the profiles of ω̇C7 H16 in the direction normal to the
reaction surface (isocontour of T = Tpeak ) remain fairly unchanged when scaled by their value at
T = Tpeak (Eq. 6.1). Hence, using Eq. 6.1, 6.10, and 6.12 in Eq. 6.6, the turbulent flame speed can
be approximated as

AT ω̇C7 H16
ST ≈ SL0 , (6.13)
A ω̇C7 H16 ,lam Tpeak
where hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak is surface-weighted and averaged in time. It corresponds exactly to
the means of the distributions presented in Fig. 6.2. AT is defined as the surface area of the reaction
137
zone surface (i.e. the isoterm T = Tpeak ) as shown in Fig. 6.11.
Reaction
zone
Figure 6.11: Schematic drawing to illustrate Eq. 6.13. SFeff = hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak · SL0 is used.
The values for hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak , ST /SL0 , and AT /A are presented in Table 6.4 for the
present flames. First, consistent with the results of Section 6.2.1, differential diffusion has a limited
effect on the turbulent surface area. The flame acts as a material surface. Second, the values
presented are in very good agreement with Eq. 6.13. Third, the main contribution to the differences
in turbulent flame speed due to differential diffusion appears in the hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak ratio.
Note that the reduction in this ratio (from the unity to the non-unity Lewis number flames) is not
solely due to local flame extinctions. By choosing a threshhold of 5% of the peak laminar burning
rate to identify regions of extinctions (similar to Ref. [53]), 3% of the overall surface area undergoes
extinction (as opposed to none in the unity Lewis number flame) as shown in Table 6.4. This cannot
account for the 30% reduction in the mean local burning rate. However, the conditional PDF of
the burning rate at Tpeak (Fig. 6.2) shows a clear shift towards smaller values from the unity to the
non-unity Lewis number flame. For instance, for 77% of the overall surface area, the burning rate is
smaller than its peak laminar value in the non-unity Lewis number flame vs. 56% in the unity Lewis
number flame (see Table 6.4). This shift in the conditional PDF of the burning rate towards smaller
values results in the reduction in the mean burning rate, with hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak corresponding
to its first moment, i.e.
Z
ω̇C7 H16 ω̇C7 H16 ω̇C7 H16 ω̇C7 H16
= P d . (6.14)
ω̇C7 H16 ,lam Tpeak ω̇C7 H16 ,lam ω̇C7 H16 ,lam ω̇C7 H16 ,lam
138
unity Lei non-unity
Lei
ST (m/s) 1.06 0.69
0
ST /SL 3.7 1.9
AT /A
D E 3.5 ± 1.0 2.9 ± 0.8
ω̇C7 H16
ω̇C7 H16 ,lam
1.00 0.66
D Tpeak
E
AT ω̇C7 H16
A ω̇C7 H16 ,lam
3.5 1.9
Tpeak
A5% /AT 0% 2.9%
A100% /AT 56.4% 76.6%
Table 6.4: Ratios relevant to the turbulent flame speed (see Eq. 6.13). Confidence intervals corre-
spond to plus or minus one standard deviation.
If the fuel consumption rate at the reaction surface were perfectly correlated with strain rate
and/or curvature, the ratio hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak could be expressed as
Z
ω̇C7 H16 ω̇C7 H16
= (κ) P (κ) dκ, (6.15)
ω̇C7 H16 ,lam Tpeak ω̇C7 H16 ,lam
or
Z
ω̇C7 H16 ω̇C7 H16
= (at ) P (at ) dat , (6.16)
ω̇C7 H16 ,lam Tpeak ω̇C7 H16 ,lam
where P (κ) and P (at ) are the probability density functions of curvature and strain rate at the
reaction surface, respectively (shown in Fig. 6.5). While the correlations found in Section 6.2.3 are
not perfect, these equations can help explain the values obtained for hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak .
For the unity Lewis number case, only strain rate has an effect on the fuel consumption rate
(see Section 6.2.3). This effect is approximately symmetric with respect to the mean strain rate (see
Fig. 6.7(a)) and this strain rate is symmetrically distributed with respect to its mean (see Fig. 6.5(b)).
Therefore, a close-to-unity hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak ratio is expected. This is consistent with results
previously reported for turbulent flames with a close-to-unity fuel Lewis number [71, 73, 58].
For the non-unity Lewis number case, the fuel consumption rate is correlated with curvature.
While curvature is symmetrically distributed with respect to its mean (see Fig. 6.5(a)), its effect
on the fuel consumption term is highly non-linear and is not symmetric with respect to its mean.
Negative values of curvature have a marginal effect on the fuel consumption rate, while positive
curvature has a very strong effect (reduced consumption rate). Consequently, the overall effect is a
139
reduction in the average fuel consumption rate, i.e. hω̇C7 H16 /ω̇C7 H16 ,lam iTpeak < 1.
6.4 Summary and discussion
The flow chart presented in Fig. 6.12 illustrates the mechanism through which turbulence and
differential diffusion affect the overall turbulent flame speed and summarizes the three main results
of this chapter.
Turb.
Figure 6.12: Flow chart illustrating the mechanism through which turbulence affects the flame speed
for the flames presented in this thesis. SFeff = hω̇F /ω̇F ,lam iTpeak SL0 is used.
First, differential diffusion was found to have limited effect on the strain rate and curvature
distribution at the reaction zone. This is a consequence of the high turbulence level (SL0 /vη 1),
the reaction surface behaving more like a material (i.e. passive) surface than a propagating surface.
Second, a correlation was found between the fuel consumption rate and curvature for the non-
unity Lewis number flame, whereas a similar correlation was found between tangential strain rate
and fuel consumption rate for the unity Lewis number case.
Finally, local extinctions were shown to have a strong effect on the turbulent flame speed by
altering the local consumption rate of the fuel.
This chapter brought a qualitative understanding of the mechanism behind local extinctions and
their effect on the turbulent flame speed, highlighting the importance of considering differential
diffusion in models. However, quantities that better correlate with source term fluctuations would
be required to adequately develop such models. The next chapter introduces such quantities and
presents a model for the source term fluctuations.

140
Chapter 7
Modeling of the reaction zone1
As previously identified in Chapters 4 to 6, the n-C7 H16 /air premixed turbulent flames considered
are found to have a considerably broadened preheat zone, but a thin reaction zone. While this
reaction zone is thin, large burning rate fluctuations were observed. More specifically, the following
results were identified:
1. largely thickened preheat zone (see Fig. 4.10),
2. thin reaction zone (see Fig. 4.9 and 4.10),
3. large burning rate fluctuations around the mean profile vs. progress variable (as previously
mentioned, the temperature is a progress variable) (see Fig. 6.1(c) and 6.3(a)),
4. for unity Lewis numbers:
< ω̇F |c >≈ ω̇F,lam (c) , (7.1)
5. for non-unity Lewis numbers:
< ω̇F |c >6= ω̇F,lam (c) , (7.2)
6. for both cases:
ω̇F (x, t) 6= ω̇F (c (x, t)) . (7.3)
Note that a discussion on the choice of the progress variable (c) for modeling purposes is provided
in the first section.

1 The author of this thesis has contributed to all the work presented in this chapter.
141
The work in this chapter is directly related to point 6, i.e. the first objective is to identify a
set of variables ψ such that ω̇F (x, t) can be approximated accurately by ω̇F (c (x, t) , ψ (x, t)), while
keeping the dimensionality of ψ small. The second objective is to generate the function ω̇F (c, ψ).
The last objective is to provide/discuss a tabulation approach with this reduced set of variables.
7.1 Progress variable
While temperature is a natural choice for the progress variable, this choice is not unique. For
modeling purposes, a progress variable with a simpler transport equation than the temperature
equation is desired. Typically, a linear combination of species mass fractions is considered:
N
X
c= bj Yj , (7.4)
j=i
where bj is a real coefficient. It is associated with the following transport equation
 
N
∂c α X
ρ + ρu · ∇c = ∇ · ρ ∇c + ∇ · ρ bj Yj Vc,j  + ω̇c , (7.5)
∂t Lec j=1
with Lec such that

N
X 1 1
bj ∇Yj = ∇c (7.6)
j=1
Lej Lec
and
N
X
ω̇c = bj ω̇j . (7.7)
j=1
In this chapter, all the derivations will be done with the progress variable defined by Eq. 7.4 (more
general), but all the figures and calculations will be done with c = YH2 + YH2 O + YCO + YCO2 . The
later is typically used in n-heptane/air flames [118, 79, 80]. Figure 7.1 shows the joint probability
density function of this progress variable vs. temperature. Figure 7.2 shows the fuel burning rate
as a function of temperature for both the higher-Ka and lower-Ka flames. As one can expect
from the strong correlelation between this progress variable and temperature (Fig. 7.1), the results
142
0.25 10000 0.25 10000

flamelet
0.2 8000 0.2 8000
0.15 6000 0.15 6000

c
c
0.1 4000 0.1 4000
0.05 2000 0.05 2000

flamelet - unity Le
flamelet - non-unity Le
0 0 0 0
600 1000 1400 1800 2200 600 1000 1400 1800 2200
T (K) T (K)
(a) Le = 1 (b) Le 6= 1
Figure 7.1: Joint PDF of the progress variable c = YH2 + YH2 O + YCO + YCO2 vs. temperature in
the higher-Ka flame. One-dimensional unstretched flame profiles are shown by dashed lines.
concerning the fuel burning rate obtained in the previous chapters are similar whether temperature
or c = YH2 + YH2 O + YCO + YCO2 is considered as the independent variable. Again, for the non-
unity Lewis number flames, we want to highlight the fact that differential diffusion effects are still
important at the reaction zone (as discussed in Chapter 6) and a single set of effective Lewis number
cannot fully characterize the flame structure throughout the flame (as discussed in Chapter 5).
7.2 Coordinate transformation
In order to identify the set of variables ψ, a similar approach to the one used in Ref. [184] for
non-premixed flames is taken in this section. The following coordinate transformation is proposed
(more details can be found in Ref. [184]):
(x1 , x2 , x3 , t) → (c (x1 , x2 , x3 , t) , c2 (x1 , x2 , x3 , t) , c3 (x1 , x2 , x3 , t) , τ ) , (7.8)
with
∇c · ∇c2 = 0, and ∇c · ∇c3 = 0, (7.9)

143
0.003 0.003
1.4 <ωF|c> 1.4 <ωF|c>
flamelet flamelet
0.0025 0.0025
1.2 1.2
ωF/ωF,lam(cpeak)
ωF/ωF,lam(cpeak)
1 0.002 1 0.002
0.8 0.8
0.0015 0.0015
0.6 0.6
0.001 0.001
0.4 0.4
0.0005 0.0005
0.2 0.2
0 0 0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(a) higher-Ka Le = 1 (c) lower-Ka Le = 1
0.003 0.003
1.4
<ωF|c> 1.4
<ωF|c>
flamelet - non-unity Le 0.0025 flamelet - non-unity Le 0.0025
1.2 1.2
ωF/ωF,lam(cpeak)
ωF/ωF,lam(cpeak)
1 0.002 1 0.002
0.8 0.8
0.0015 0.0015
0.6 0.6
0.001 0.001
0.4 0.4
0.0005 0.0005
0.2 0.2
0 0 0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(b) higher-Ka Le 6= 1 (d) lower-Ka Le 6= 1
Figure 7.2: Joint PDF and conditional mean (solid line) of the fuel burning rate, ω̇F , normalized by
its peak laminar value vs. progress variable c for both the higher-Ka and the lower-Ka flames. The
unity, the non-unity, and the effective Lewis number flamelet solutions are also shown (dashed line).
i.e. the variables c2 and c3 lie in the surface of constant c. Note that these variables can be
curvilinear. The following derivatives result from the coordinate transformation:
∂ ∂c ∂ ∂c2 ∂ ∂c3 ∂ ∂
= + + + , (7.10)
∂t ∂t ∂c ∂t ∂c2 ∂t ∂c3 ∂τ
∂ ∂c ∂ ∂c2 ∂ ∂c3 ∂
= + + ,
∂x1 ∂x1 ∂c ∂x1 ∂c2 ∂x1 ∂c3
∂ ∂c ∂ ∂c2 ∂ ∂c3 ∂
= + + ,
∂x2 ∂x2 ∂c ∂x2 ∂c2 ∂x2 ∂c3
∂ ∂c ∂ ∂c2 ∂ ∂c3 ∂
= + + .
∂x3 ∂x3 ∂c ∂x3 ∂c2 ∂x3 ∂c3
144
The only assumption made in this section is that this transformation exists, i.e. the Jacobian of the
transformation is not singular. Unfortunately, this is not the case everywhere in the flame. Inside
pockets of unburnt fuel, or where the reaction zone is closed on itself the Jacobian will be singular
point-wise. However, only a discrete number of singularities are found and this local transformation
is valid everywhere else.
7.3 Transformed transport equations
Applying the coordinate transformation (Section 7.2) to Eq. 2.5 and 2.4, and using Eq. 7.5, the
following transformed equations for the species mass fractions and temperature are obtained:
   
N
 1 1 X  ∂Y
i
∇ · ρα − ∇c + ∇ · ρ bj Yj Vc,j  + ω̇c (7.11)
 Lec Lei j=1
 ∂c
| {z }
normal convection
ρχ ∂ 2 Yi
− − ω̇i − ∇ · (ρYi Vc,i )
2Lei ∂c2 |{z} | {z }
| {z } chemical source velocity correction
normal diffusion
3
∂Yi X ∂Yi ∂ck
= −ρ −ρ + u · ∇ck
∂τ ∂ck ∂t
k=2
| {z }
Lagrangian transport
3 2
X ρχk ∂ Yi 2ρα ∂ 2 Yi
+ + (∇c 2 · ∇c3 )
2Lei ∂c2k Lei ∂c2 ∂c3
k=2
| {z }
tangential diffusion

ρα ∂Yi ρα ∂Yi
+∇· ∇c2 +∇· ∇c3 ,
Lei ∂c2 Lei ∂c3
| {z }
tangential convection
145
   
N
 1 X ρχ ∂cp  ∂T
cp ∇ · ρα − 1 ∇c + ∇ · ρ bj Yj Vc,j  + ω̇c − (7.12)
 Lec j=1
2cp ∂c  ∂c
| {z }
normal convection 1
N N
!
X ρcp,i χ ∂Yi X ∂T
+ + ρcp,i Yi Vc,i · ∇c
i=1
2Lei ∂c i=1
∂c
| {z }
normal convection 2
ρcp χ ∂ 2 T
− 2
− ω̇T
| 2 {z∂c }
|{z}
chemical source
normal diffusion
3
∂T X ∂T ∂ck
= −ρcp − ρcp + u · ∇ck
∂τ ∂ck ∂t
k=2
| {z }
3 2
X ρcp χk ∂ T ∂2T
+ + 2ρcp α (∇c2 · ∇c3 )
2 ∂c2k ∂c2 ∂c3
k=2
| {z }
∂T ∂T
+ ∇ · (ρcp α∇c2 ) + ∇ · (ρcp α∇c3 )
∂c2 ∂c3
| {z }
tangential convection 1
3 N N
!
X X ρcp,i χk ∂Yi ∂T X ∂T
+ + ρcp,i Yi Vc,i · ∇ck
i=1
2Lei ∂ck ∂ck i=1 ∂ck
k=2
| {z }
N
X ρcp,i α ∂Yi ∂T ∂Yi ∂T
+ + (∇c2 · ∇c3 ),
i=1
Lei ∂c2 ∂c3 ∂c3 ∂c2
| {z }
where χ = 2α|∇c|2 and χk = 2α|∇ck |2 for k = 2, 3 are the dissipation rates of c, c2 , and c3 , respec-
tively. An additional step in the transformation can be found in Appendix C. In the above equations,
the Lagrangian transport corresponds to the convection of the scalar, in the c2 -c3 -direction, i.e. in
the surface of iso-c, induced by the Lagrangian transport of these c2 and c3 coordinates. The normal
terms correspond to convection or diffusion of the scalar in the c-direction, whereas the tangential
terms correspond to convection or diffusion of the scalar in the c2 -c3 -direction.

146
7.4 Unity Lewis number limit
As a first step, the unity Lewis number limit is considered. As mentioned throughout this thesis,
in the absence of differential diffusion, the effect of turbulence on the chemistry can be directly
investigated. In this section, the unity Lewis number flamelet equations are presented first and the
extent of their validity in the present turbulent flames is discussed. Then, these equations are used
to provide a physical explanation to the fuel burning rate fluctuations. Finally, approaches to model
the fuel burning rate are presented.
7.4.1 Flamelet equations
Recall that, in the regime considered, the reaction zone is considered to be thin. It is therefore
assumed that, in the vicinity of the reaction zone, ∂/∂ck << ∂/∂c, for k = 2, 3. With this assumption
and in the absence of differential diffusion (i.e. by setting all Lewis numbers to unity), Eq. 7.11
and 7.12 reduce to

∂Yi ∂Yi ρχ ∂ 2 Yi
ρ + ω̇c = + ω̇i . (7.13)
∂τ ∂c 2 ∂c2
N
∂T ∂T ρcp χ ∂ 2 T ρχ ∂cp ∂T X ρcp,i χ ∂Yi ∂T
ρ + cp ω̇c = 2
+ ω̇T + + . (7.14)
∂τ ∂c 2 ∂c 2 ∂c ∂c i=1
2 ∂c ∂c
Note that Eq. 7.11 and 7.12 under unity Lewis number assumption can be found in Appendix D.
It is further assumed that the dependence in τ in the transformed coordinate system is negligible.
Note that the coordinate c is still a function of time t. With this assumption, the flamelet equations
are obtained:
dYi ρχ d2 Yi
ω̇c = 2
+ ω̇i (7.15)
| {zdc} |2 {zdc } chemical source
|{z}
convection diffusion
N
dT ρcp χ d2 T ρχ dcp dT X ρcp,i χ dYi dT
cp ω̇c = 2
+ ω̇T + + .
| {zdc} | 2 {zdc } chemical source |2 dc
{z dc} 2 dc dc
|{z}
i=1
convection diffusion grad cp -induced convection
| {z }
species mass flux-induced convection
(7.16)
These equations form a system of ordinary differential equations in which the Yi and T are
the unknowns, c is the variable, and the dissipation rate χ(c) is the parameter. Given the two
147
assumptions made in this section, each of the terms in the above equations are only a function of c
and χ.
In order to test if these assumptions (∂/∂ck << ∂/∂c, for k = 2, 3, and ∂/∂τ = 0) are valid, each
of the terms in Eq. 7.15 (for several species) and Eq. 7.16 are evaluated point-wise in the turbulent
flame. Then, the conditional mean and the standard deviation of these terms are computed and
the results are presented in Fig. 7.3 (species) and 7.4 (temperature) for both the higher-Ka and the
lower-Ka unity Lewis number flames. The species flamelet equations for n-C7 H16 (a reactant), CO
(an intermediate that contributes to the progress variable), and H2 O (a product that contributes
to the progress variable) are considered. Note that due to the large number of grid points covering
the reaction zone and the large number of data files considered (spanning over 50 turnover times),
the statistics are smooth and well converged. First and foremost, it is clear that the residual of
the respective equations (the sum of the neglected terms corresponding to the right hand side of
Eq. 7.11 and 7.12) is small in comparison to the other terms in these equations. These results
indicate that the assumptions made in this section are valid. This means that, although the flames
are highly three-dimensional and unsteady (in physical space), they can be well described by the
one-dimensional (in phase space) flamelet equations (Eq. 7.15 and 7.16) parameterized by a single
parameter χ(c).
The unity Lewis number flamelet equations are particularly interesting to study the reaction
zone. In contrast with the one-dimensional transport equations in physical space (e.g. Fig. 5.9 for
hydrogen), the terms in the flamelet equations are only large in the reaction zone. Considering the
species transport equations, the convection term and the chemical source term can only be non-zero
in the reaction zone. On the other hand, the dissipation term would be non-zero if the tangential and
the τ -dependent terms are non-negligible. However, the species mass fraction vs. progress variable
profiles are linear outside the reaction zone, even in the turbulent flames (see Fig. 5.2). Therefore,
the dissipation term is close to zero, and so is the sum of the neglected terms.
Note that although the residuals are small for both flames, they are larger in the higher-Ka flame.
It is also interesting to note that these residuals are only large where the diffusion term is large.
148
1.5 1.5
Conv Conv
Diff Diff
1 1
Src Src
Relative magnitude
Relative magnitude
Res Res
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-1.5 -1.5
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(a) higher-Ka, YC7 H16 (d) lower-Ka, YC7 H16
1.5 1.5
Conv Conv
Diff Diff
1 1
Src Src
Relative magnitude
Relative magnitude
Res Res
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-1.5 -1.5
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(b) higher-Ka, YCO (e) lower-Ka, YCO
1.5 1.5
Conv Conv
Diff Diff
1 Src 1 Src
Relative magnitude
Relative magnitude
Res Res
0.5 0.5
0 0
-0.5 -0.5
-1 -1
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(c) higher-Ka, YH2 O (f) lower-Ka, YH2 O
Figure 7.3: Mean of the terms in Eq 7.15 (for three different species) conditioned on c for both
unity Lewis number flames, < term|c >. The bars correspond the standard deviation of these terms
conditioned on c,< (term− < term|c >)2 |c >1/2 . The red dashed line corresponds to the residual of
the equation (right hand side of Eq. 7.11 in the unity Lewis number limit). Each term is normalized
by the absolute value of the largest term (corresponding to the same equation) in a unity Lewis
number one-dimensional unstretched flame.
149
1 1
Relative magnitude
Relative magnitude
0.5 0.5
0 0
Conv Conv
-0.5
Diff -0.5
Diff
Src Src
grad c conv grad c conv
mass flux mass flux
-1 Res -1 Res
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(a) higher-Ka (b) lower-Ka
Figure 7.4: Mean of the terms in Eq 7.16 conditioned on c for both unity Lewis number flames,
< term|c >. The bars correspond the standard deviation of these terms conditioned on c,< (term− <
term|c >)2 |c >1/2 . The red dashed line corresponds to the residual of the equation (right hand side
of Eq. 7.12 in the unity Lewis number limit). Each term is normalized by the absolute value of
the largest term (corresponding to the same equation) in a unity Lewis number one-dimensional
unstretched flame.
It is hypothesized that the terms left out in Eq. 7.11 contribute to an effective diffusivity, which
would result in an effective dissipation rate. Unfortunately, appropriately identifying this effective
diffusivity is not a straight forward task, as this effective diffusivity has to be consistent with the
(c,c2 ,c3 )-coordinate system. In addition, as discussed in Chapter 5, it is expected to vary with c,
since turbulent mixing is stronger in the unburnt gas. To verify this hypothesis, a framework that
can be used to identify locally the effective diffusivity in this coordinate system would be required.
This should be done in future work.
7.4.2 Fuel burning rate
In this section, the flamelet equations are used to model the fuel burning rate in the two unity
Lewis number turbulent n-heptane flames considered in this thesis (the higher-Ka and the lower-Ka
flames).
150
7.4.2.1 Dependence of the burning rate on c and χ
Given the validity of the flamelet equations (as shown in the previous section), the fuel burning
rate should solely be a function of the progress variable c and its dissipation rate χ. In contrast
with Fig. 6.7 and 6.8 shown in the previous chapter, Fig. 7.5 presents the joint probability density
function of the fuel consumption rate vs. the dissipation rate of the progress variable, both of which
are evaluated at cpeak , with cpeak being the location in c-space of the maximum fuel consumption
rate. The correlation coefficients are listed in Table 7.1 and are compared to those obtained with
strain rate and curvature (Section 6.2.3). It is clear from Fig. 7.5 and Table 7.1 that the fuel burning
rate is far more correlated with χ at cpeak . Physically, this means that turbulence affects the fuel
burning rate by compressing or extending the isosurfaces of the progress variable. It also means
that stretching the flow field is not equivalent to stretching these isosurfaces in the context of high
Karlovitz premixed flames.
7
2.5
6
2
ωF/ωF,lam,peak
1.5 4
3
1
2
0.5
1
0 0
0 0.5 1 1.5 2 2.5 3
χ/χlam,peak
Figure 7.5: Joint PDF of the fuel burning rate vs. the dissipation rate of the progress variable,
normalized by their peak laminar values, on the isosurface of c = cpeak in the higher-Ka unity Lewis
number flame.
This result suggests that there should exist a function f (c, χ) that estimates accurately the fuel
burning rate. Such a function can be easily identified a posteriori, i.e. using the data from the DNS.
We first define the prediction error made by the function f by the following L2-norm over the whole
151
r dCorn
Strain rate 0.61 0.57
Curvature -0.12 0.33
Dissipation rate 0.996 0.997
and fuel consumption rate, curvature and fuel consumption rate, and dissipation rate and fuel
consumption rate at c = cpeak for the higher-Ka unity Lewis number flame.
domain:
Z t2 Z
1 2
F
f = (f − ω̇F ) dV dt. (7.17)
V (t2 − t1 ) t1 Ω
A similar approach was used to quantify the error made by LES filters on chemical source terms [123]
and by scalar dissipation rate sub-grid-scale models [13]. The function that minimizes this error,
given that f is a function of c and χ only, also called the optimal estimator [119, 13], is the mean of
the fuel burning rate conditional on c and χ, i.e.
arg minF
f =< ω̇F |c, χ > . (7.18)
f (c,χ)
This result is well described in Ref. [119] and a brief discussion is provided in this paragraph.
An optimal estimator Ω defined on an L2-norm, for a set of variables π, and a quantity to estimate,
2
γ, minimizes ||Ω (π) − γ||2 =< (Ω (π) − γ) > with < · > the statistical mean (or expectation). The
quadratic estimation error made from an estimator g (π) satisfies the following orthogonal relation
(see proof in Ref. [119]):
2 2 2
< (γ − g (π)) >=< (γ− < γ|π >) > + < (< γ|π > −g (π)) > . (7.19)
The first term on the right hand side is refered to as the irreducible error, whereas it can be shown
the the second term is null if and only if g (π) =< γ|π >. Therefore, the optimal estimator Ω (π) is
< γ|π >.
The optimal estimator given by Eq. 7.18 is evaluated for the two turbulent flames. The prediction
errors are listed in Table 7.2. They are compared to the prediction errors made by the following
152
functions: < ω̇F |c > (optimal estimator given that f is a function of the progress variable only),
< ω̇F |c, κ > (optimal estimator given that f is a function of the progress variable and the curvature
only), and < ω̇F |c, at > (optimal estimator given that f is a function of the progress variable and the
tangential strain rate only). The last function, ω̇F,FGM (c, χ), will be discussed in the next section.
It is clear from this table that, first, allowing a dependence on c and χ instead of c only improves
dramatically the prediction of the fuel burning rate. This is illustrated in Fig. 7.6 which shows the
probability density function of the point-wise comparison between the predicted and the actual fuel
burning rate, for both these functions (higher-Ka flame). Second, the choice of the second variable
has a strong effect on the prediction error (Table 7.2). Consistently with the results of Table 7.1,
the fuel burning rate appears to be much more correlated with the progress variable dissipation rate
than with either curvature or tangential strain rate.
F 1/2
F

Estimator f f /0
higher-Ka lower-Ka
< ω̇F |c > 0.606 0.426
< ω̇F |c, κ > 0.489 0.337
< ω̇F |c, at > 0.459 0.321
< ω̇F |c, χ > 0.084 0.042
ω̇F,FGM (c, χ) 0.109 0.067
Table 7.2: Prediction error (Eq. 7.17) for various estimators of the fuel burning rate normalized by
the prediction error for f = 0 in both unity Lewis number flames.
2 0.1 2 0.1
<ωF|c,χ>/ωF,lam,peak
<ωF|c>/ωF,lam,peak
0.08 0.08
1.5 1.5
0.06 0.06
1 1
0.04 0.04
0.5 0.5
0.02 0.02
0 0 0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
ωF/ωF,lam,peak ωF/ωF,lam,peak
(a) < ω̇F |c > vs. ω̇F (x, t) (b) < ω̇F |c, χ > vs. ω̇F (x, t)
Figure 7.6: Joint PDF of the comparison between the predicted fuel burning rate and the actual
burning rate in the higher-Ka unity Lewis number flame.
The optimal estimator < ω̇F |c, χ > provides a very good estimation of the fuel burning rate. The
153
results form this section show that ω̇F (x, t), which lives in a 35-dimensional phase-space (function
of N − 1 species and temperature), collapses very nicely on a two-dimensional manifold. While
the optimal estimator < ω̇F |c, χ > provides such an a posteriori representation of this manifold,
the flamelet equations can also be used to generate a (c, χ)-manifold a priori. The quality of the
prediction provided by this a priori manifold is assessed in the next section.
7.4.2.2 Flamelet-generated manifold
Equations 7.15 and 7.16 are solved numerically (using a modified version of FlameMaster) for a
range of β such that χ (c) = βχ0 (c), with χ0 (c) the dissipation rate profile in a one-dimensional
unstretched flame, to generate a (c, χ)-manifold. Figure 7.7 shows that the solution from the flamelet
equation, for β = 1, is identical to the solution obtained for an unstretched flame solved in physical
space (Eq. 2.2 to 2.5). Dirichlet boundary conditions corresponding to the unburnt and the burnt
mixture (in a one-dimensional unstretched flame) are used to solve the flamelet equations.
1.2
phys. space
1
c space
ωF/ωF,lam,peak
0.8
0.6
0.4
0.2
0
0 0.05 0.1 0.15 0.2 0.25
c
Figure 7.7: Comparison of the fuel burning rate vs. progress variable profile between the 1D flame
solved in physical space and the flamelet solved in c-space (Eq. 7.15 and 7.16 with χ (c) = χphys (c)).
A model for the fuel burning rate, ω̇F,FGM (c, χ) is extracted from the manifold. Figure 7.8
presents the functional dependence of this function on χ at cpeak . It is compared to that of the
optimal estimator and the joint PDF previously presented in Fig. 7.5. The FGM agrees very well
with the conditional mean at cpeak . The prediction error (Eq. 7.17) made by ω̇F,FGM (c, χ) is listed
154
7
2.5
ωF,FGM
<ωF|c,χ> 6
2
ωF/ωF,lam,peak
5
1.5 4
3
1
2
0.5
1
0 0
0 0.5 1 1.5 2 2.5 3
χ/χlam,peak
Figure 7.8: Comparison between the fuel burning rate predicted by ω̇F,FGM (c, χ) and < ω̇F | (c, χ) >
as a function of the dissipation rate, normalized by their peak laminar values, given c = cpeak . The
joint PDF on the isosurface of c = cpeak in the higher-Ka unity Lewis number flame is also shown
for comparison.
in Table 7.2. This error is very close to that made by the optimal estimator, meaning that the a
priori manifold generated by the flamelet equations is very close to the optimal (i.e. a posteriori)
manifold given c and χ as variables.
To summarize, ω̇F,FGM (c, χ) provides a very good prediction of the local and instantaneous fuel
burning rate.
7.5 Non-unity Lewis numbers
The previous section provided a framework that was used to physically explain the effect of the
turbulence on the fuel burning rate, in the absence of differential diffusion, and to model this fuel
burning rate. In the present section, effects of differential diffusion are included in a similar analysis.
7.5.1 Flamelet equations
Similar to Section 7.4.1, we make the same assumptions of local one-dimensionality in phase space,
i.e. ∂/∂ck << ∂/∂c, for k = 2, 3, and steady state in phase space, i.e. ∂/∂τ = 0, but no assumption
155
is made on the species Lewis numbers. Accordingly, Eq. 7.11 and 7.12 simplify to
   
N
 1 1 X  ∂Y
i
∇ · ρα − ∇c + ∇ · ρ bj Yj Vc,j  + ω̇c (7.20)
 Lec Lei j=1
 ∂c
ρχ ∂ 2 Yi
= + ω̇i + ∇ · (ρYi Vc,i ) ,
2Lei ∂c2
and
 
X N
 1  ∂T
cp ∇ · ρα − 1 ∇c + bj ∇ · (ρYj Vc,j ) + ω̇c (7.21)
 Lec j=1
 ∂c
N N
ρcp χ ∂ 2 T ρχ ∂cp ∂T X ρcp,i χ ∂Yi ∂T X
= + ω̇ T + + + ρcp,i Yi Vc,i · ∇T.
2 ∂c2 2 ∂c ∂c i=1
2Lei ∂c ∂c i=1
Let ξ = ∇ · (ρα∇c). Then, Eq. 7.20 and 7.21 can be expended to obtain the non-unity Lewis number
flamelet equations:
dYi ρχ d2 Yi
ω̇c = + ω̇i (7.22)
| {zdc} 2Lei dc2 |{z}
| {z } chemical source
convection diffusion

1 1 ρχ d 1 1 dYi
+ ξ − + −
Lei Lec 2 dc Lei Lec dc
| {z }
differential diffusion-induced convection
N
X dYi
+ ∇ · (ρYi Vc,i ) − bj ∇ · (ρYj Vc,j ) ,
j=1
dc
| {z }
velocity correction
156
with

1 Yi ρχ d Yi dW
∇ · (ρYi Vc,i ) = ξ + (7.23)
Lei W 2 dc W dc
  
N N
Yi
X Yj ρχ d Y i
X Yj  dW
− ξ + 
W j=1 Lej 2 dc W j=1 Lej dc
N
ρχ Yi d2 W

ρχ dYi X 1 dYj
− ξYi + +
2 dc j=1 Lej dc 2Lei W dc2
 
N N
ρχ Yi X Yj d2 W ρχ d2 X Yj 
− − Yi
2 W j=1 Lej dc2 2 dc2 j=1 Lej
N
ρχ Yi d 1 dW ρχ X d 1 dYj
+ + Yi
2 W dc Lei dc 2 j=1
dc Lej dc
N
ρχ dYi X d 1
+ ξYi + Yj ,
2 dc j=1 dc Lej
and
N
dT ρcp χ d2 T ρχ dcp dT X ρcp,i χ dYi dT
cp ω̇c = 2
+ ω̇T + +
| {zdc} | 2 {zdc } chemical source |2 dc
{z dc} 2Lei dc dc
|{z}
i=1
convection diffusion grad cp -induced convection
| {z }
species mass flux-induced convection
(7.24)

1 ρχ d 1 dT
+ cp ξ 1 − −
Lec 2 dc Lec dc
| {z }
differential diffusion-induced convection
N N
X X dT
+ ρcp,i Yi Vc,i · ∇T − cp bj ∇ · (ρYj Vc,j ) ,
i=1 j=1
dc
| {z }
velocity correction
with
N
"N N
#
X ρχ X cp dYi X cp − cp,i Yi dW dT
ρcp,i Yi Vc,i · ∇T = − + . (7.25)
i=1
2 i=1 Lei dc i=1
Lei W dc dc
1
PN bj dYj
Note that the inverse of the Lewis number of the progress variable simplifies two Lec = j=1 Lej dc .
In this set of equations, a second parameter, ξ (c), is involved. It corresponds to the diffusion of c in
the original coordinate system. Note that, as one would expect, this parameter includes curvature
157
effects
χα 1/2
1 d χ d
ξ= (ρχ) + (ρα) − ρκ , (7.26)
4 dc α dc 2
with κ = ∇ · (−∇c/|∇c|) (similar to Eq. 6.2 in which T was considered as the progress variable). A
detailed derivation of Eq. 7.26 can be found in Ref. [184].
Similar to Section 7.4.1, to verify the validity of the two assumptions made in this section
(∂/∂ck << ∂/∂c, for k = 2, 3, and ∂/∂τ = 0), the budget of Eq. 7.22 and 7.24 from the DNS is shown
in Fig. 7.9 and 7.10. In contrast with the unity Lewis number limit, non-zero terms are found
outside the reaction zone for n-C7 H16 and H2 O (diffusion term and differential diffusion-induced
convective term). These are associated with relatively large residuals (the sum of the neglected
terms corresponding to the right hand side of Eq. 7.11 and 7.12). Recalling that the preheat zone
is largely thickened in the turbulent flames, it is not expected that the assumptions made in this
section are valid throughout this preheat zone. Nevertheless, the residuals are sufficiently small in
the reaction zone to expect some dependence of the source terms on c, χ, and ξ, which will be
evaluated in the next section and Section 7.6.2.
As discussed in Section 7.4.1, it is hypothesized that the terms left out in Eq. 7.11 contribute to
effective diffusivities, which would result in an effective dissipation rate and effective species Lewis
numbers. The fact that the residuals are larger where the diffusion term and the differential diffusion-
induced convective term are large is consistent with this hypothesis. For the reasons mentionned in
Section 7.4.1 this hypothesis should be validated in future work. In particular, a framework that
can be used to identify locally the effective Lewis numbers in the transformed coordinate system
would be required.
Consistent with the unity Lewis number case, the residuals are larger in the higher-Ka flame.
The terms associated with the velocity correction are negligible for all species considered. The
velocity correction terms will therefore be neglected in the rest of the analysis. This is an important
result for Section 7.6.2.

158
0.8 0.8
0.6 0.6
0.4 0.4
Relative magnitude
Relative magnitude
0.2 0.2
0 0
-0.2 -0.2
-0.4 Conv -0.4 Conv
Diff Diff
-0.6 Src -0.6 Src
-0.8 Diff-Conv -0.8 Diff-Conv
Vel cor Vel cor
-1 Res -1 Res
-1.2 -1.2
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(a) higher-Ka, YC7 H16 (d) lower-Ka, YC7 H16
1.2 1.2
1 Conv 1 Conv
Diff Diff
0.8 0.8
Src Src
Relative magnitude
Relative magnitude
0.6 Diff-Conv 0.6 Diff-Conv
0.4
Vel cor 0.4
Vel cor
Res Res
0.2 0.2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1 -1
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(b) higher-Ka, YCO (e) lower-Ka, YCO
1.2 1.2
1 Conv 1 Conv
Diff Diff
0.8 Src 0.8 Src
Relative magnitude
Relative magnitude
0.6 Diff-Conv 0.6 Diff-Conv

Vel cor Vel cor
0.4 Res 0.4 Res
0.2 0.2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(c) higher-Ka, YH2 O (f) lower-Ka, YH2 O
Figure 7.9: Mean of the terms in Eq 7.22 (for three different species) conditioned on c for both non-
unity Lewis number flames, < term|c >. The bars correspond the standard deviation of these terms
conditioned on c,< (term− < term|c >)2 |c >1/2 . The red dashed line corresponds to the residual
of the equation (right hand side of Eq. 7.11). Each term is normalized by the absolute value of the
largest term (corresponding to the same equation) in a non-unity Lewis number one-dimensional
unstretched flame.
159
1.5 1.5
Conv Conv
Diff Diff
1 Src 1 Src
Relative magnitude
Relative magnitude
diff. conv diff. conv
grad c conv grad c conv
0.5 mass flux 0.5 mass flux
vel. cor. vel. cor.
Res Res
0 0
-0.5 -0.5
-1 -1
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
Figure 7.10: Mean of the terms in Eq 7.24 conditioned on c for both non-unity Lewis number
flames, < term|c >. The bars correspond the standard deviation of these terms conditioned on
c,< (term− < term|c >)2 |c >1/2 . The red dashed line corresponds to the residual of the equation
(right hand side of Eq. 7.12). Each term is normalized by the absolute value of the largest term
(corresponding to the same equation) in a non-unity Lewis number one-dimensional unstretched
flame.
7.5.2 Fuel burning rate
The results from the previous section are used to model a posteriori the fuel burning rate in the
non-unity Lewis number flames. An approach to obtain an a priori model is discussed.
7.5.2.1 Dependence of the burning rate on c, χ, and ξ
The joint probability density function of the fuel burning rate vs. the dissipation rate at cpeak is
presented in Fig. 7.11. Note that cpeak corresponds to the location in c-space of the peak mean (con-
ditional on c) burning rate in the DNS. In contrast, clam,peak is the equivalent for a one-dimensional
unstretched flame. Recall that, as shown in Fig. 7.2(b) and 7.2(d), these are different. The cor-
relation coefficients associated with Fig. 7.11 are listed in Table 7.3. They are compared to those
obtained in Chapter 6 when considering the strain rate or the curvature instead of the dissipation
rate. The fuel burning rate correlates better with the dissipation rate than with curvature or strain
rate. However, the correlation found in the unity Lewis number flames is significantly stronger than
in the non-unity Lewis number case (wide spread in Fig. 7.11).
Given the results of the last section, it is expected that the spread in Fig. 7.11 will be reduced by
160
allowing a dependence on ξ. This can indeed be observed in Fig. 7.12, where the joint probability
density function of the fuel burning rate vs. the dissipation rate is presented at cpeak , but for three
different intervals of ξ. The associated correlation coefficients are listed in Table 7.3. Intervals of
ξ are used instead of single values in order to have a sufficient number points from the DNS to
obtain statistically relevant correlations. Note that the width of the intervals (in ξ-space) is five
times smaller than the difference between the center of these intervals. It is clear from Fig. 7.12 and
Table 7.3 that the correlation between the fuel burning rate and the dissipation rate is significantly
improved when ξ is set to a fixed value.

1.6
4.5
1.4
4
1.2 3.5
ωF/ωF,lam,peak
1 3
0.8 2.5
2
0.6
1.5
0.4
1
0.2 0.5
0 0
0 0.5 1 1.5 2
χ/χlam,peak
normalized by their peak laminar values (at c = clam,peak ), on the isosurface of c = cpeak in the
higher-Ka non-unity Lewis number flame.
r dCorn
Strain rate 0.49 0.45
Curvature -0.53 0.65
Dissipation rate 0.75 0.72
Dissipation rate |ξ ∈ Ξlow 0.94 0.94
Dissipation rate |ξ ∈ Ξmid 0.92 0.92
Dissipation rate |ξ ∈ Ξhigh 0.87 0.86
Table 7.3: Pearson’s correlation coefficient, r, and distance correlation, dCorn , between fuel burning
rate and various variables at c = cpeak for the higher-Ka non-unity Lewis number flame.
This result suggests that the optimal estimator given, c, χ, and ξ should predict the fuel burning
rate significantly better than the optimal estimator given, c only or c, χ. Similar to Fig. 7.6, Fig. 7.13
shows the probability density function of the point-wise comparison in the higher-Ka flame between
161
1.6 8 1.6 8
1.4 7 1.4 7
ωF/ωF,lam,peak
ωF/ωF,lam,peak
1.2 6 1.2 6
1 5 1 5
0.8 4 0.8 4
0.6 3 0.6 3
0.4 2 0.4 2
0.2 1 0.2 1
0 0 0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
χ/χlam,peak χ/χlam,peak
(a) ξ ∈ Ξlow (b) ξ ∈ Ξmid
1.6 8
1.4 7
ωF/ωF,lam,peak
1.2 6
1 5
0.8 4
0.6 3
0.4 2
0.2 1
0 0
0 0.5 1 1.5 2
χ/χlam,peak
(c) ξ ∈ Ξhigh
normalized by their peak laminar values (at c = clam,peak ), on the isosurface of c = cpeak in the
higher-Ka
4 non-unity Lewis number flame, conditional on ξ ∈ Ξk , with k = low, mid, high. Ξlow =
− 3 ξlam,peak , − 23 ξlam,peak , Ξmid = 23 ξlam,peak , 43 ξlam,peak , Ξhigh = 83 ξlam,peak , 10

3 ξlam,peak .
the predicted and the actual fuel burning rate, for three optimal estimators: < ω̇F |c >, < ω̇F |c, χ >,
and < ω̇F |c, χ, ξ >. The prediction errors for these functions are listed in Table 7.4 for both flames.
For the lower-Ka flame, < ω̇F |c, χ, ξ > provides a good estimation of the fuel burning rate. In
particular, the prediction error is six times smaller than for < ω̇F |c >. For the higher-Ka, the
error associated with < ω̇F |c, χ, ξ > is three times smaller than that for < ω̇F |c >. Although this
error corresponds to a significant improvement with respect to the prediction error associated with
< ω̇F |c >, it is twice larger than for the lower-Ka flame. This shows that, as mentioned in the
162
1.5 0.1 1.5 0.1
<ωF|c,χ>/ωF,lam,peak
<ωF|c>/ωF,lam,peak
1.25 0.08 1.25 0.08
1 1
0.06 0.06
0.75 0.75
0.04 0.04
0.5 0.5
0.25 0.02 0.25 0.02
0 0 0 0
0 0.25 0.5 0.75 1 1.25 1.5 0 0.25 0.5 0.75 1 1.25 1.5
ωF/ωF,lam,peak ωF/ωF,lam,peak
(a) < ω̇F |c > vs. ω̇F (x, t) (b) < ω̇F |c, χ > vs. ω̇F (x, t)
1.5 0.1
<ωF|c,χ,ξ>/ωF,lam,peak
1.25 0.08
1
0.06
0.75
0.04
0.5
0.25 0.02
0 0
0 0.25 0.5 0.75 1 1.25 1.5
ωF/ωF,lam,peak
(c) < ω̇F |c, χ, ξ > vs. ω̇F (x, t)
Figure 7.13: Joint PDF of the comparison between the predicted fuel burning rate and the actual
burning rate in the higher-Ka non-unity Lewis number flame.
previous section, the tangential terms are non-negligible in the higher-Ka flame. To reduce this
error, further modeling efforts would be required.
7.5.2.2 Flamelet-generated manifold
In a manner similar to what was done in Section 7.4.2.2 for the unity Lewis number cases, Eq. 7.22
and 7.24 can be solved for ranges of β1 and β2 , such that χ (c) = β1 χ0 (c), and ξ (c) = β2 ξ 0 (c), with
ξ 0 (c) the diffusion term of the progress variable (in the original coordinate system) profile in a one-
dimensional unstretched flame, in order to generate an a priori (c,χ,ξ)-manifold. While theoretically
possible, this flamelet-generated manifold would not adequately represent the fuel burning in the
163
F 1/2
F

Estimator f f /0
higher-Ka lower-Ka
< ω̇F |c > 0.765 0.680
< ω̇F |c, χ > 0.515 0.341
< ω̇F |c, χ, ξ > 0.256 0.112
Table 7.4: Prediction error (Eq. 7.17) for various estimators of the fuel burning rate normalized by
the prediction error for f = 0 in both non-unity Lewis number flames.
DNS. Indeed, the residual terms in the preheat zone would need to be modeled (Fig. 7.9). It
was suggested in Section 7.5.1 that these could possibly be modeled by effective Lewis numbers.
The position of the peak fuel burning rate in Fig. 7.2(b) and 7.2(d) also suggests that effective
Lewis numbers are necessary to capture this position in c-space. Similar to the argument made in
Chapter 5, since turbulent mixing (and the resulting tangential terms in the flamelet equations)
varies accross the flame, a single set of Lewis numbers cannot be considered over the whole range of
c. As mentioned, in Sections 7.5.1 and 5.4, to identify such Lewis numbers is not a simple task and
this should be done in future work.
Within the framework of this thesis, an a posteriori DNS-generated manifold (from which <
ω̇F |c, χ, ξ > is extracted) is presented (previous section), but an a priori flamelet-generated manifold
cannot be provided for the non-unity Lewis number flames.
7.6 Tabulation approach with the reduced-dimension mani-
fold
The results from the previous sections suggest that the 35-dimensional phase space can be replaced
by a two-dimensional (c,χ)-manifold in the unity Lewis number case, and, to some extent, by a three-
dimensional (c,χ,ξ)-manifold in the non-unity Lewis number case. As mentioned in the introduction,
one of the objectives of this thesis is to reduce the dimensionality of the chemistry. With these
reduced-chemistry manifolds, the thermodynamic state is given by c and χ in the unity Lewis
number case and by c, χ, and ξ in the non-unity Lewis number case. During a simulation of
premixed turbulent flames, all chemistry-related quantities can be tabulated as a function of these
164
controlling variables (tabulated chemistry approach [171, 165, 63, 64, 98]). For the momentum and
the continuity equations, the density and the viscosity need to be tabulated
Given the definition of χ and ξ, these can be determined by the field of c, if α is tabulated.
However, c needs to be transported and additional properties/terms need to be tabulated.
In this section, the transport equation for the progress variable are provided in these two lower-
dimension manifolds and closure for the chemistry-related quantities is provided/discussed in light
of the results obtained in this chapter and Chapter 5.
7.6.1 Unity Lewis number limit
The progress variable transport equation in the unity Lewis number limit reads
∂c
ρ + ρu · ∇c = ∇ · (ρα∇c) + ω̇c . (7.27)
∂t
In this equation, ρ, α, and ω̇c need to be closed (tabulated). The viscosity, ν, also requires closure
for the momentum equation.
First, as expected from the results presented in Section 5.1.1, and as shown in Fig. 7.14, ρ ≈
ρFGM (c), ν ≈ νFGM (c) and α ≈ αFGM (c), and these properties do not present modeling challenges.
Second, the progress variable production rate needs to be modeled. Since the reaction zone
is thin in the flames studied, both the DNS-generated manifold and the FGM presented in the
previous sections should also provide a very good estimator for the progress variable production
rate. This is shown in Table 7.5. Therefore, the progress variable source term can be modeled
as ω̇c =< ω̇c |c, χ >DNS (tabulation from a DNS-generated manifold) or ω̇c = ω̇c,FGM (c, χ). As
discussed in Section 7.4.2.1, the DNS-generated manifold provides the optimal predictor given the
choice of variables. Unfortunately, this manifold is specific to the mixture and the unburnt conditions
considered in the DNS. In contrast, the FGM can be easily generated for any mixtures and unburnt
conditions.
165
1.4 10 0.0005 10000

flamelet
flamelet
1.2
8 0.0004 8000
1
ρ (kg m-3)
ν (m2 s-1)
6 0.0003 6000
0.8
0.6
4 0.0002 4000
0.4
2 0.0001 2000
0.2
0 0 0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(a) density (b) viscosity
0.0006 10000
flamelet
0.0005
8000
0.0004
α (m2 s-1)
6000
0.0003
4000
0.0002
2000
0.0001
0 0
0 0.05 0.1 0.15 0.2 0.25
c
(c) diffusivity
Figure 7.14: Joint PDF of the thermodynamic properties vs. progress variable for the higher-Ka
unity Lewis number flame. The 1D flame profiles are also shown (dashed lines).
Note that the joint probability density function (along with the correlation coefficients) of ω̇c vs.
χ at cpeak is provided in Appendix E. The DNSGM and the FGM predictions are also provided.
7.6.2 Non-unity Lewis numbers
With differential diffusion, the following single scalar equation is considered in the reduced (c,χ,ξ)-
manifold:

∂c α
ρ + ρu · ∇c = ∇ · ρ ∇c + ω̇c , (7.28)
∂t Lec
166
1/2
Estimator f cf /c0
higher-Ka lower-Ka
< ω̇c |c > 0.222 0.159
< ω̇c |c, κ > 0.183 0.130
< ω̇c |c, at > 0.182 0.132
< ω̇c |c, χ > 0.049 0.031
ω̇c,FGM (c, χ) 0.111 0.068
Table 7.5: Prediction error (Eq. 7.17) for various estimators of the progress variable source term
normalized by the prediction error for f = 0 in both unity Lewis number flames.
which was already provided in Section 7.1 (Eq. 7.5). Recall that the term associated with the velocity
correction was found to be negligible in the DNS (see Section 7.5.1) and hence it is not shown in
Eq. 7.28. In the non-unity Lewis number case, ρ, ν, α (to evaluate χ and ξ from the field of c),
α/Lec , and ω̇c need to be closed (tabulated).
First, as observed in Fig 7.15, ρ, ν, and α are very well predicted by the flamelet with effective
Lewis numbers (as computed in Section 5.3). Second, a similar closure for the progress variable
diffusivity can also be obtained. Figure 7.16 shows its joint probability density function vs. c in
both non-unity Lewis number turbulent flames. For values of c between 0 and about 0.22 (beyond
which the terms in the flamelet equation are close to zero), little deviation from the conditional
α
mean < Lec |c >DNS is observed. Again, the flamelet solution with effective Lewis numbers (same
α
as in Section 5.3 and 7.1) is very close to < Lec |c >DNS . Therefore, the thermodynamic properties
α α
can be modeled as ρ = ρFGM (c; Leeff ), α = αFGM (c; Leeff ) , and Lec = Lec FGM (c; Leeff ), with
3/2 −1/2
Lei,eff = (1 + 0.05Ka) / (1/Lei + 0.05Ka), where Ka = (u0 /SL ) (l/lF ) , as introduced in
Chapter 5. Third, the closure for the chemical source term of the progress variable can only be
provided a posteriori (from the DNS) within the framework of this thesis. The prediction errors
made with < ω̇c |c, χ, ξ >DNS are listed in Table 7.6. They are comparable to the errors made for
the fuel burning rate (Section 7.5.2.1). Note that the joint probability density function (along with
the correlation coefficients) of ω̇c vs. χ at cpeak is provided in Appendix E for three intervals of ξ.
167
1.4 4 0.0005 10000

flamelet - non-unity Le 3.5 flamelet - non-unity Le
1.2 flamelet - Leeff
0.0004 flamelet - Leeff 8000
3
1
ν (m2 s-1)
2.5
ρ (kg m-3)
0.0003 6000
0.8
2
0.6
0.0002 4000
1.5
0.4
1
0.0001 2000
0.2 0.5
0 0 0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
(a) density (b) viscosity
0.0006 10000
flamelet - unity Le
0.0005 flamelet - non-unity Le
flamelet - Leeff 8000
0.0004
α (m2 s-1)
6000
0.0003
4000
0.0002
2000
0.0001
0 0
0 0.05 0.1 0.15 0.2 0.25
c
(c) diffusivity
Figure 7.15: Joint PDF of the thermodynamic properties vs. progress variable for the higher-Ka
non-unity Lewis number flame. The unity, non-unity, and effective Lewis number 1D flame profiles
are also shown (dashed lines).
7.7 Summary and discussion
While the flames considered in this work have a thin reaction zone, large fluctuations around the
mean burning rate (conditional on c) were observed (Chapter 6). In order to model these fluctuations,
a coordinate transformation was performed in this chapter. The following results were identified:
1. for the unity Lewis number flames,
(a) the turbulent flames can be well represented by a set of one-dimensional (in c-space)
flamelet equations parameterized by χ (c),

168
0.0006 10000 0.0006 10000
<α/Lec|c> <α/Lec|c>
0.0005 flamelet - unity Le 0.0005 flamelet - unity Le
8000 8000
flamelet - non-unity Le flamelet - non-unity Le
α/Lec (m2 s-1)
α/Lec (m2 s-1)

0.0004 0.0004
6000 6000
0.0003 0.0003
4000 4000
0.0002 0.0002
2000 2000
0.0001 0.0001
0 0 0 0
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
c c
Figure 7.16: Joint PDF and conditional mean (solid line) of the progress variable diffusivity vs.
progress variable for both non-unity Lewis number flames. The unity, non-unity, and effective Lewis
number 1D flame profiles are also shown (dashed lines).
1/2
Estimator f cf /c0
higher-Ka lower-Ka
< ω̇c |c > 0.483 0.572
< ω̇c |c, χ > 0.401 0.293
< ω̇c |c, χ, ξ > 0.264 0.223
Table 7.6: Prediction error (Eq. 7.17) for various estimators of the progress variable source term
normalized by the prediction error for f = 0 in both non-unity Lewis number flames.
(b) the fuel burning rate ω̇F (x, t) can be approximated accurately by ω̇F (c (x, t) , χ (x, t)),
(c) the fuel burning rate is well predicted a priori by the solution to the set of flamelet
equations,
(d) all the chemistry-related quantities relevant to a turbulent flame simulation with this
reduced chemistry ((c,χ)-manifold) can be tabulated from an a priori FGM;
2. for the non-unity Lewis number flames,
(a) in the vicinity of the reaction zone, the flame can, to a lesser extent (due to non-negligible
tangential terms), be represented by a set of one-dimensional flamelet equations (in c)
parameterized by χ (c) and ξ (c),
(b) the fuel burning rate ω̇F (x, t) can be approximated accurately (only to some extent in
the higher-Ka flame) by ω̇F (c (x, t) , χ (x, t) , ξ (x, t)),

169
(c) the fuel burning rate cannot be predicted a priori by the solution to the set of flamelet
equations, unless a model for the effective diffusivities as a function of the progress variable
is considered (future work),
(d) the thermodynamic properties relevant to a turbulent flame simulation can be tabulated
from an a priori 1D-FGM considering the effective Lewis numbers provided in Chapter 5,
(e) the source term of the progress variable (relevant to a turbulent flame simulation with
this reduced chemistry) can be closed by an a posteriori DNS-generated manifold.
Whether the models provided in this chapter are over-accurate, sufficiently accurate, or insuf-
ficiently accurate for practical applications remains unclear. As discussed in the introduction, the
simulation of reacting flows in practical devices requires an LES framework. How the accuracy of
these models affects that of the LES closure models has yet to be determined. Nevertheless, an
important contribution of the analysis made in this chapter is the quantification of the predictive
accuracy associated with the proposed models.
Finally, it is important to note that the assumption of a thin reaction zone was made to obtain
the premixed flamelet equations. While this assumption is valid in the thin reaction zone regime,
it may not be the case for higher turbulence intensities corresponding to the distributed burning
regime.
170
Chapter 8
Conclusions
A time-integration scheme has been proposed for the simulation of stiff reacting flows. Using this
scheme, a series of direct numerial simulations of high Karlovitz number, n-C7 H16 , turbulent pre-
mixed flames have been performed. It was found that the flame structure of these turbulent flames
can be well captured by one-dimensional flames accounting for the effective species Lewis numbers.
The reaction zone was found to remain thin, yet large fluctuations in the fuel burning rate were
identified. Extinctions were observed only in the presence of differential diffusion, and these events
were correlated with high curvature regions. A model to capture the burning fluctuations was pro-
posed using a new flamelet approach. Finally, a reduced-chemistry modeling approach using flamelet
equations was presented.
8.1 Time-integration
A semi-implicit preconditioning strategy, applied to an iterative method, was proposed for the time-
integration of the stiff chemistry in the simulation of unsteady reacting flows, such as turbulent
flames. The preconditioner consists of an approximation of the diagonal of the chemical Jacobian.
It is integrated into the iterative procedure already implemented in the NGA code, in order to
account for the non-linearities of the governing equations.
The proposed semi-implicit preconditioning, in combination with the iterative method, is far less
computationally expensive than a fully-implicit method and was shown to be as inexpensive or less
expensive than operator-splitting methods, while being more accurate. It was also observed that
171
the quasi steady-state (QSS) assumption may not be used for conventional species in the turbulent
flame presently studied. As such, the proposed method is more suited than alternative methods for
the type of flow studied, i.e. high Karlovitz flames.
8.2 Direct numerical simulations of high Karlovitz n-heptane/air
flames
Relying on the new time integration scheme, a series of direct numerical simulations of high Karlovitz
number, n-C7 H16 , turbulent premixed flames have been performed. These flames fall in the upper
part of the thin reaction zone regime, close to the transition to the broken/distributed reaction zone
regime.
For both the higher-Ka and the lower-Ka flames, the effects of turbulence on the flame in
the absence of differential diffusion have been assessed through a first simulation with unity Lewis
number, whereas the effects of turbulence on differential diffusion have been analyzed through a
second simulation with non-unity Lewis numbers. Qualitative results indicated that the preheat
zone is largely thickened by turbulent mixing, while the reaction zone remained thin.
8.3 Flame structure and differential diffusion
The flame thickness of the unity Lewis number flame was shown to be largely affected by turbulence.
However, its structure (defined as the dependence of species mass fractions on temperature) is very
similar to that of a one-dimensional, flat flame, suggesting that turbulence has a very limited impact
on the flame in temperature space, in the absence of differential diffusion. On the other hand,
the structure of the non-unity Lewis number flame is affected more substantially by turbulence. It
was argued that turbulence affects the flame structure through an effective Lewis number. At high
turbulence levels (i.e. high Karlovitz number) turbulence reduces differential diffusion effects.
A model for the effective species Lewis numbers was provided and validated first with a series
of DNS of turbulent premixed lean hydrogen flames (φ = 0.4) [10]. The structure of these flames
172
was found to vary considerably with Karlovitz number. However, this structure was found to be
bounded by the structure of two limiting cases: unstretched laminar flamelets with 1) non-unity
Lewis numbers, and 2) unity Lewis numbers. Moreover, the turbulent flames were shown to have the
same structure in average as unstretched laminar flames with appropriate effective Lewis numbers.
An a priori model for those effective Lewis numbers was derived from RANS transport equations.
The model’s dependence on a Reynolds versus Karlovitz number was investigated and a dependence
on the Karlovitz number was found to be more suitable. While the parameters in the model were
estimated with large uncertainties, the model was found to be in very good agreement with the DNS
data.
The model, which was validated with the series of hydrogen/air flames, was directly used (with
the same parameter) for the n-heptane/air flames. The unstretched laminar flamelets computed
with these effective Lewis numbers are in good agreement with the mean turbulent flame structure.
While the effective Lewis number model captures adequately this mean structure, it does not account
for variations in turbulent mixing across the flame (associated with the transformation of turbulence
due to the increase in kinematic viscosity). In other words, differential diffusion effects are almost
suppressed in the preheat zone, but are still present close to the reaction zone.
8.4 Reaction zone
The reaction zone of both the unity and the non-unity Lewis numbers flames was shown to be
thin. However, large source term fluctuations were present in both flames and in particular, local
extinction events were found in the flame with non-unity Lewis numbers. This showed that a reaction
zone can at the same time be thin and broken. The characterization of the reaction zone provided
three main results.
First, differential diffusion was found to have limited effect on the srain rate and curvature
distribution at the reaction zone. This is a consequence of the high turbulence level (SL0 /vη 1),
the reaction surface behaving more like a material (i.e. passive) surface than a propagating surface.
Second, a correlation was found between the fuel consumption rate and curvature for the non-
173
unity Lewis number flame, whereas a similar correlation was found between tangential strain rate
and fuel consumption rate for the unity Lewis number case.
Finally, local extinctions were shown to have a strong effect on the turbulent flame speed by
altering the local consumption rate of the fuel.
These results provide a qualitative understanding of the mechanism behind local extinctions and
their effects on the turbulent flame speed, highlighting the importance of considering differential
diffusion in models. However, a quantity that better correlates with source term fluctuations is
required to adequately develop such models.
To identify such a quantity, a different approach was used. A local, flame-based coordinate
transformation was applied to the scalar transport equations. For the unity Lewis number flames
(both higher-Ka and lower-Ka), a budget analysis showed that the terms in the progress variable (c)
direction are far more important than the tangential terms. In other words, around the reaction zone,
the flame is very well described by a set of one-dimensional (in progress variable space) equations,
or flamelets. These equations are parameterized by the dissipation rate of the progress variable,
χ. The fuel burning rate was shown to correlate strongly with this dissipation rate. Two models
for the fuel burning rate were provided: one corresponding to the conditional mean of the burning
rate in the DNS, < ω̇F |c, χ >, and the other being obtained by solving the set of one-dimensional
equations (FGM), varying the parameter χ, ω̇F,FGM (c, χ). Both models estimate accurately the
local fuel burning rate.
The same approach was used for the non-unity Lewis number flames. In these flames, differential
diffusion has a strong effect which results in non-negligible tangential terms, especially in the preheat
zone. Nevertheless, these terms are sufficiently small in the reaction zone to consider the flamelet
equations (normal terms) for modeling purposes. With differential diffusion, a second parameter
appears in these flamelet equations, ξ = ∇ · (ρα∇c). A model for the fuel burning rate was proposed,
< ω̇F |c, χ, ξ >, which provides relatively good accuracy.
An important contribution is the identification of these two controlling variables, χ and ξ, and
the quantification of the best accuracy that can be achieved by models considering these variables
174
(optimal estimator).
8.5 Reduced-chemistry models
As mentioned in the introduction, an important aspect in the development of LES models is the
reduction of the chemistry. It was shown in this thesis that to represent the flame structure of the
turbulent unity Lewis number flames, a single progress variable is required. To capture the fuel
burning rate, an additional variable is needed, namely the dissipation rate of this progress variable.
This means that the 35-dimensional manifold representing the chemistry can be reduced down to
a two-dimensional (c,χ)-manifold. The scalar transport equations in this dimensionnally-reduced
manifold simplifies to a single transport equation for the progress variable. The thermodynamic
properties can be properly closed with a single flamelet solution, while the source term in the equation
is closed by the DNS-generated manifold, or the FGM, i.e. < ω̇c |c, χ >DNS or ω̇c,FGM (c, χ).
For the non-unity Lewis number flames, the mean flame structure was shown to be well repre-
sented by a single flamelet solved with the appropriate effective species Lewis numbers. To capture
the fuel burning rate, two additional variables are required, χ and ξ. Similarly, a single transport
equation is required in the reduced (c,χ,ξ)-manifold. Closure for the thermodynamic properties is
obtained by a single flamelet solved with the appropriate effective species Lewis numbers, while the
source term in the equation is closed by the DNS-generated manifold < ω̇c |c, χ, ξ >DNS . Future work
should consider the extension of the model to an FGM with appropriate effective Lewis numbers.
It should be noted that the present results (i.e. low dimensional manifold) are expected to be
independent of the size of the chemical model used.
8.6 Limitations and suggestions for future work
The model proposed for the effective species Lewis numbers (Chapter 5) should be seen as a first
attempt to describe the transition from “laminar” Lewis numbers to unity effective Lewis numbers.
Quantification of uncertainties has yet to be performed, and more data points are needed to fully
175
validate the model. In addition, a framework to identify the effective species Lewis numbers as a
function of the progress variable, especially in the transformed coordinates (Chapter 7), would open
doors for the generation of FGMs to be used in turbulent non-unity Lewis number flames.
Quantitative prediction errors have been provided for the chemistry-reduction closure models.
However, it is unclear what level of precision is required for sub-grid-scale closure models in LES. Fu-
ture work should be extended to the development of such models that are consistent with the present
(c,χ)- and (c,χ,ξ)-manifolds. Depending on the accuracy of the LES, better chemistry-reduction
models may be necessary, i.e. the tangential terms in the transformed equations (Chapter 7) may
need to be modeled.
The analysis in this thesis was done on flames with small integral length scales (relatively low
turbulent Reynolds number), due to the expensive computational nature of the DNS. With the
increasing capabilities of super-computers, it could be possible to increase the turbulent Reynolds
number by close to an order of magnitude, while keeping the Karlovitz number constant, in a
relatively close future. This could confirm/disconfirm if the results obtained in this thesis are
Karlovitz number dependent, but Reynolds number independent.
In addition to the integral length scale, the domain width considered in the present flames is
relatively small. This constrains the mode shapes the flame can exhibit. With the current forcing
method, a simulation with a larger integral length scale would also require a larger domain width.
With such a simulation, the effects of the additional mode shapes on the results presented in this
thesis could be assessed.
The results in this thesis are argued to be valid in the thin reaction zone regime. This is
particularly important for the results of Chapter 7, which rely on the fact that the gradients in the
progress variable direction are much larger than those in the tangential directions. This assumption
is not expected to be valid in the distributed burning regime. It would be interesting to perform
simulations in the broken/distributed reaction zone regime to assess the extent of the validity of the
results presented in this thesis.
Finally, the effect of the chemical mechanism, the nature of the large hydrocarbon fuel, and the
176
unburnt temperature on the present results should also be investigated in future work.
177
Appendix A
Impacts of the effective Lewis

number model
A.1 Impact on SL and lF
As the effective Lewis numbers of the different species change with the Karlovitz number, the asso-
ciated laminar flame speed and flame thickness are expected to change as well. This is particularly
important as both of these quantities are used in modeling the turbulent flame speed [147, 141] and
defining a regime diagram [142].
Figure A.1 shows, in solid lines, the effective laminar flame speed (SL,eff ) and flame thickness
(lF,eff ) obtained from laminar unstretched flamelet simulations, in which the Lewis numbers were
modified according to the Ka-based model. These results suggest that these effective Lewis numbers
have an important impact on the laminar flame speed and flame thickness. More specifically, both
the laminar flame speed and flame thickness vary by almost a factor of 2 between the “laminar”
Lewis numbers and the unity Lewis numbers flamelets. Note that this factor is expected to be case
dependent, i.e. it is expected to be different for different fuels and unburnt conditions.
These strong effects may be understood from simple asymptotic flame theory. For a one-step
reaction, matched asymptotic expansion with only the fuel Lewis number different than unity, the
following ratios are obtained [140]:

0.5
SL,2 LeF,2
= , (A.1)
SL,1 LeF,1
178
0.5
lF,2 LeF,1
= , (A.2)
lF,1 LeF,2
where SL,i and lF,i are the laminar flame speed and flame thickness of a flame with corresponding
fuel Lewis number LeF,i . Note that the only different parameter between flames 1 and 2 is the
fuel Lewis number. Using the Ka-based model (Eq. 5.34), the effective laminar flame speed and
flame thickness obtained from Eq. A.1 and A.2 are plotted as a function of the Karlovitz number in
Fig. A.1. The above analytic expressions (Eq. A.1 and A.2) are compared to the flamelet calculation
results using the same effective Lewis numbers. The agreement between the curves is relatively
good. The discrepancies can be due to the numerous simplifying assumptions made in the matched
asymptotic expansion. Note that a 0.6 exponent instead of 0.5 in Eq. A.1 gives better agreement
for the effective laminar flame speed with the flamelet calculations. Nevertheless, it is interesting
to point out that the complex chemistry mechanism used in the flamelet simulations provides the
same behavior as a single-step mechanism.
50 0.65
45 0.6
0.55
40
sL,eff (cm/s)
lF,eff (mm)
0.5
35
0.45
30
0.4
25 0.35
20 0.3
10-1 100 101 102 103 104
Ka
Figure A.1: Effective laminar flame speed and flame thickness versus Karlovitz number from flamelet
simulations (solid lines) and analytical expressions (dotted lines).
179
A.2 Impact on the regime diagram
The result from subsection A.1 suggests that a new effective Karlovitz number should be defined
that takes into account the effective flame thickness:
2
lF,eff
Kaeff = . (A.3)
η2
Using Eq. A.2, this effective Karlovitz number can be related to the “traditional” Karlovitz number
as follows:
LeF
Kaeff = Ka. (A.4)
LeF,eff
Recall the relation derived from Peters [142] from which the iso-Ka lines are obtained in the
premixed regime diagram:

1/3
u0

l
= Ka2/3 . (A.5)
SL lF
Note that SL lF = ν is assumed in Eq. A.5.
Following the results shown in Chapter 5, the relevant quantity to differentiate whether the
smallest eddies will penetrate the preheat zone or the reaction zone is the effective Karlovitz number
(respectively Kaeff = 1 and Kaeff = 100). Using the Ka-based model (Eq. 5.34) in Eq. A.4, Ka can
be expressed as a function of Kaeff :

1
aKa Kaeff − 1

Ka = Ka
2a LeF
h 2 i0.5
Ka Ka
+ a Kaeff − 1 + 4a LeF Kaeff . (A.6)
Fixing Kaeff to its limiting values (1 or 100), the resulting Ka values can be used in Eq. A.5 to
obtain a modified regime diagram as shown in Fig. A.2. Note that the u0 /SL = 1 line is modified
to be the u0 /SL,eff = 1 line. The ReT = 1 line is not modified by the fuel effective Lewis number.
The lines separating the wrinkled/corrugated flamelet regimes and the thin reaction zone regimes
are unchanged. This result is not surprising as, in the wrinkled/corrugated flamelet regimes, eddies
180
4
10
w/o effective LeF
w effective LeF
DNS broken reaction zones
103
102
u’/sL
thin reaction zones
1
10
corrugated flamelets
100
laminar
10-1
-1 0 1 2 3 4
10 10 10 10 10 10
l/lF
Figure A.2: Regime diagram taking into account the effective Karlovitz number, considering the
Ka-based model.
do not penetrate the preheat zone and hence do not lead to an effective Lewis number. On the
other hand, the separation between thin reaction and broken reaction zones is affected. As eddies
penetrate in the preheat zone, the resulting changes in the effective Lewis numbers lead to an increase
in the effective laminar flame speed and a decrease in the effective flame thickness (Fig. A.1). Both
of these effects tend to “counteract” the increase in the Karlovitz number (Eq. A.5).
Aspden et al. found in Ref. [10] that the two lowest Ka (10 and 100) DNS simulations showed
a thin reaction zone-like behavior, whereas the highest Ka (1526) simulation was clearly in the
broken reaction or distributed burning zone. The intermediate simulation (Ka = 266) was labeled
as a transition case, whereas scaling arguments [142] suggest that the transition should be around
Ka = 100. These observations from Aspden et al. agree very well with Fig. A.2, even though the
present authors do not claim to explain the transition Ka found in Ref. [10].
A.3 Impact on ST
Finally, the effective Lewis numbers have an impact on the predicted turbulent flame speed. Nu-
merous experiments have showed a turbulent flame speed dependency on the Lewis number of the
fuel [172, 99, 94, 100, 50, 69]. The effects of Reynolds or Karlovitz number on the flame charac-
teristics through the effective Lewis number might explain these dependencies. Consider the most
181
widely used model for the turbulent flame speed [142]:
n
u0

ST
=1+C . (A.7)
SL SL
Using the idea that the effective laminar flame speed is the relevant quantity rather than the laminar
flame speed, a new model can be derived. More precisely, we propose
ST ST SL,eff
= ·
SL SL,eff SL
0.6 " 0 n 0.6n #
Leeff u Le
= 1+C (A.8)
Le SL Leeff
where Leeff follows Eq. 5.34. The reason for a 0.6 exponent in Eq. A.8 is explained in subsection A.1.
Both a Lewis number and a Karlovtiz number dependency is hence introduced.
In 1992, Bradley et al. [21] found the turbulent flame speed to be Lewis number and Reynolds
0 2
number dependent. After identifying the relevant KLe grouping, where K = 0.157 SuL Re−0.5T is
the Karlovitz stretch factor, they obtained the following correlation:
ST u0 −0.3
= 0.88 (KLe) . (A.9)
SL SL
More recently, Muppala et al. [125] developed a model that also depends on the Lewis number and
the Reynolds number:

0.3
u0

ST A
=1+ Re0.25 , (A.10)
SL exp (Le − 1) T SL
where A is a fitting parameter.
These four models and correlations are compared against one another in Fig. A.3. The parameters
C, n, and A are adjusted such that all models fall on each others for unity Lewis number. The integral
length scale and the viscosity are taken to be the same as in Section 5.2.1. The Lewis number is
then modified to give the solid lines (Le = 0.3) and the dotted lines (Le = 3). The model proposed
in Chapter 5 (C-n with Leeff ) agrees very well with the correlation from Bradley et al. These results
182
45
C-n
40 C-n with Leeff
Bradley
35 Muppala
Le = 0.3
30
sT/sL
25
20
15
10
Le = 3
5
0
0 20 40 60 80 100
u’/sL
Figure A.3: Normalized turbulent flame speed profiles for the models given by Eq. A.7-A.10 for
Le =0.3 (solid lines), 1 (symbols), and 3 (dotted lines).
further suggest that the effective laminar flame speed (SL,eff ) is the relevant quantity rather than
the laminar unstretched flame speed. A similar analysis has been performed with the ReT -based
model. The results were very similar to those shown in Fig. A.3.
A.4 Summary
The effective laminar flame speed and flame thickness have to be considered. Indeed, those quantities
vary by almost a factor of two between the “laminar“ Lewis numbers and the unity Lewis numbers
for premixed hydrogen flamelet (φ = 0.4) simulations.
In light of these results, an effective Karlovitz number Kaeff was defined and should be used
in developing turbulent combustion models. The regime diagram proposed by Peters [142] should
therefore be adapted to the fuel considered whenever the fuel has a Lewis number far from unity.
Finally, a turbulent flame speed model that takes into account the Lewis number and the Karlovitz
number was derived.

183
Appendix B
Correlations
To better understand the effect of strain rate and curvature on the reaction zone, it is important
to assess if they are correlated. Figures B.1(a) and B.1(b) present the joint density distributions
for these two quantities, in both flames. The Pearson correlation coefficient [132] and the distance
correlation [168] are shown in Table B.1. These measures of correlation will be used further and are
described in the following.
The Pearson correlation coefficient, a well established statistical measure, is computed as
Pn
i=1Xi − X̄ Yi − Ȳ
r = qP 2 qPn 2 , (B.1)
n
i=1 Xi − X̄ i=1 Yi − Ȳ
for a sample pair of data (Xi , Yi ), where ¯· denotes the sample mean, and n is the sample size. It is
important to note that r = 0 does not imply independence of the random variables X and Y . Since
a measure of dependence is sought, the distance correlation was also computed. This statistical
measure was introduced in the previous decade to overcome this issue [168]. Moreover, it is more
relevant than Pearson’s coefficient for non-linear correlations. To compute the distance correlation,
184
for a sample pair of data (Xi , Yi ) (assuming X, Y ∈ R to simplify), let us first introduce
n
1X
aij = |Xi − Xj | āi? = aij (B.2)
n j=1
n n
1X 1 X
ā?j = aij ā?? = aij
n i=1 n2 i,j=1
Aij = aij − āi? − ā?j + ā?? .
Similarly, define bij = |Yi − Yj | and Bij = bij − b̄i? − b̄?j + b̄?? . The sample distance covariance is
then defined as v
uX n
1u
dCovn (X, Y ) = t Aij Bij . (B.3)
n i,j=1
Finally, the distance correlation, is computed as
dCovn (X, Y )
dCorn (X, Y ) = p , (B.4)
dVarn (X)dVarn (Y )
with dVarn (X) = dCovn (X, X).
Given the values listed in Table B.1, curvature and strain rate show low dependency, as previously
observed [71], and, again, differential diffusion has virtually no effect.
20 0.2 20 0.2
/S0L
/S0L
15 15
0.15 0.15
7H16
7H16
10 10
Strain rate atδC
Strain rate atδC
5 0.1 5 0.1
0 0
0.05 0.05
-5 -5
-10 0 -10 0
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
Mean curvature κδC Mean curvature κδC
7H16 7H16
(a) Le = 1 (b) Le 6= 1
Figure B.1: Joint probability density function, on the isoterm T = Tpeak , of the normalized strain
rate vs. the normalized mean curvature. (a) Le = 1; (b) Le 6= 1.
185
r dCorn
Le = 1 -0.26 0.33
Le 6= 1 -0.25 0.35
Table B.1: Pearson’s correlation coefficient, r, and distance correlation, dCorn , between strain rate
and curvature at T = Tpeak for both (higher-Ka) flames presented in this thesis.
186
Appendix C
Intermediate step in the

transformation of the transport
equations
C.1 Species equations
Applying the coordinate transformation (Section 7.2) to Eq. 2.5 yields
3
∂Yi ∂Yi ∂c X ∂Yi ∂ck
ρ +ρ +ρ + u · ∇ck (C.1)
∂τ ∂c ∂t ∂ck ∂t
k=2

∂Yi ∂Yi ρα
+ρ u · ∇c − ∇· ∇c
∂c ∂c Lei
3
ρχ ∂ 2 Yi X ρχk ∂ 2 Yi 2ρα ∂ 2 Yi
= + ω̇i + + (∇c 2 · ∇c3 )
2Lei ∂c2 2Lei ∂c2k Lei ∂c2 ∂c3
k=2

ρα ∂Yi ρα ∂Yi
+∇· ∇c2 +∇· ∇c3 + ∇ · (ρYi Vc,i ) .
Lei ∂c2 Lei ∂c3
Using Eq. 7.5, Eq. 7.11 is obtained.
C.2 Temperature equation
Applying the coordinate transformation (Section 7.2) to Eq. 2.4 yields

187
3
∂T ∂T ∂c X ∂T ∂ck
ρcp + ρcp + ρcp + u · ∇ck (C.2)
∂τ ∂c ∂t ∂ck ∂t
k=2
∂T ∂T
+ ρcp u · ∇c − ∇ · (ρcp α∇c)
∂c ∂c
N N
!
X ρcp,i χ ∂Yi X ∂T
+ + ρcp,i Yi Vc,i · ∇c
i=1
2Lei ∂c i=1
∂c
3
ρcp χ ∂ 2 T X ρcp χk ∂ 2 T ∂2T
= + ω̇ T + + 2ρcp α (∇c2 · ∇c 3 )
2 ∂c2 2 ∂c2k ∂c2 ∂c3
k=2
∂T ∂T
+ ∇ · (ρcp α∇c2 ) + ∇ · (ρcp α∇c3 )
∂c2 ∂c3
3 N N
!
X X ρcp,i χk ∂Yi ∂T X ∂T
+ + ρcp,i Yi Vc,i · ∇ck
2Lei ∂ck ∂ck i=1 ∂ck
k=2 i=1
N
X ρcp,i α ∂Yi ∂T ∂Yi ∂T
+ + (∇c2 · ∇c3 ) ,
i=1
Lei ∂c2 ∂c3 ∂c3 ∂c2
Using Eq. 7.5, Eq. 7.12 is obtained.

188
Appendix D
Transformed transport equations

in the unity Lewis number limit
D.1 Species equation
In the unity Lewis number limit, Eq. 7.11 reads
3
∂Yi X ∂Yi ∂ck
ρ +ρ + u · ∇ck (D.1)
∂τ ∂ck ∂t
k=2
| {z }
∂Yi
+ ω̇c
| {z∂c}
normal convection
3
ρχ ∂ 2 Yi X ρχk ∂ 2 Yi ∂ 2 Yi
= 2
+ ω̇i + 2 + 2ρα (∇c2 · ∇c3 )
|2 {z∂c } 2 ∂ck ∂c2 ∂c3
|{z}
chemical source k=2
normal diffusion | {z }
∂Yi ∂Yi
+ ∇ · (ρα∇c2 ) + ∇ · (ρα∇c3 ) .
∂c2 ∂c3
| {z }
tangential convection
189
D.2 Temperature equation
In the unity Lewis number limit, Eq. 7.12 reads
3
∂T X ∂T ∂ck
ρcp + ρcp + u · ∇ck (D.2)
∂τ ∂ck ∂t
k=2
| {z }

ρχ ∂cp ∂T
+ cp ω̇c −
2cp ∂c ∂c
| {z }
normal convection 1
N
X ρcp,i χ ∂Yi ∂T
+
i=1
2 ∂c ∂c
| {z }
normal convection 2
3
ρcp χ ∂ 2 T X ρcp χk ∂ 2 T ∂2T
= + ω̇T + + 2ρcp α (∇c 2 · ∇c3 )
| 2 {z∂c }
2 |{z} 2 ∂c2k ∂c2 ∂c3
chemical source k=2
normal diffusion | {z }
∂T ∂T
+ ∇ · (ρcp α∇c2 ) + ∇ · (ρcp α∇c3 )
∂c2 ∂c3
| {z }
3 X
N
X ρcp,i χk ∂Yi ∂T
+
2 ∂ck ∂ck
k=2 i=1
| {z }
N
X ∂Yi ∂T ∂Yi ∂T
+ ρcp,i α + (∇c2 · ∇c3 ) .
i=1
∂c2 ∂c3 ∂c3 ∂c2
| {z }
190
Appendix E
Joint probability density function

of ω̇c vs. χ at cpeak
E.1 Unity Lewis number limit
Figure E.1 shows the joint probablity density function of ω̇c vs. χ at cpeak for both unity Lewis
number flames. The DNSGM and the FGM solutions are shown for comparison. The correlation
coefficients for this figure are listed in Table. E.1.
1.4 1.4
30 30
ωc,FGM ωc,FGM
1.3 1.3
<ωc|c,χ> <ωc|c,χ>
25 25
1.2 1.2
1.1 20 1.1 20
ωc
ωc
1 1
15 15
0.9 0.9
10 10
0.8 0.8
5 5
0.7 0.7
0.6 0 0.6 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
χ χ
Figure E.1: Comparison between the fuel burning rate predicted by ω̇c,FGM (c, χ) and < ω̇c | (c, χ) >
as a function of the dissipation rate, normalized by their peak laminar values, given c = cpeak . The
joint PDF on the isosurface of c = cpeak in the turbulent flames is also shown for comparison.
191
r dCorn
Dissipation rate, higher-Ka 0.96 0.96
Dissipation rate, lower-Ka 0.96 0.96
Table E.1: Pearson’s correlation coefficient, r, and distance correlation, dCorn , between dissipation
rate and progress variable source term at c = cpeak for the higher-Ka unity Lewis number flame.
E.2 Non-unity Lewis numbers
Figure E.2 shows the joint probablity density function of ω̇c vs. χ at cpeak for three intervals of ξ, for
the higher-Ka non-unity Lewis number flame. The correlation coefficients for this figure are listed
in Table. E.2.
r dCorn
Dissipation rate |χ ∈ Ξlow 0.90 0.89
Dissipation rate |χ ∈ Ξmid 0.86 0.86
Dissipation rate|χ ∈ Ξhigh 0.78 0.80
Table E.2: Pearson’s correlation coefficient, r, and distance correlation, dCorn , between dissipation
rate and progress variable source terms for various intervals of ξ at c = cpeak for the higher-Ka
non-unity Lewis number flame.
192
1.4 4 1.4 4
1.2 3.5 1.2 3.5
ωc/ωc,lam,peak
ωc/ωc,lam,peak
1 3 1 3
2.5 2.5
0.8 0.8
2 2
0.6 0.6
1.5 1.5
0.4 1 0.4 1
0.2 0.5 0.2 0.5
0 0 0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
χ/χlam,c-peak χ/χlam,c-peak
(a) χ ∈ Ξlow (b) χ ∈ Ξmid
1.4 4
1.2 3.5
ωc/ωc,lam,peak
1 3
2.5
0.8
2
0.6
1.5
0.4 1
0.2 0.5
0 0
0 0.5 1 1.5 2 2.5
χ/χlam,c-peak
(c) χ ∈ Ξhigh
Figure E.2: Joint PDF of the progress variable source term vs. the dissipation rate of the progress
variable, normalized by their peak laminar values (at c = clam,peak ), on the isosurface of c = cpeak in
the
5 higher-Ka3 non-unity Lewis number 7 flame, conditional on ξ ∈ Ξk , with
1 k = low, mid, high. Ξlow =
9 3
4 ξ lam,c-peak , 2 ξlam,c-peak , Ξ mid = 8 ξ lam,c-peak , 8 ξlam,c-peak , Ξ high = 2 ξlam,c-peak , 4 ξ lam,c-peak .
193
Bibliography
[1] Abdel-Gayed, R., and Bradley, D. Combustion regimes and the straining of turbulent
premixed flames. Combust. Flame 76 (1989), 213–218.
[2] Alexander, R. Diagonally implicit runge-kutta methods for stiff O.D.E.’s. SIAM J. Numer.
Anal. 14, 6 (1977), 1006–1021.
[3] Almgren, A., Bell, J., Colella, P., Howell, L., and Welcome, M. A conservative
adaptive projection method for the variable density incompressible Navier–Stokes equations.
J. Comput. Phys. 142 (1998), 1–46.
[4] Altantzis, C., Frouzakis, C., Tomboulides, A., Matalon, M., and Boulouchos, K.
Hydrodynamic and thermodiffusive instability effects on the evolution of laminar planar lean
premixed hydrogen flames. J. Fluid Mech. 700 (2012), 329–361.
[5] Amato, A., Day, M., Cheng, R., Bell, J., and Lieuwen, T. Lead-
ing edge statistics of turbulent, lean, H2 /air flames. Proc. Comb. Inst. (2014),
http://dx.doi.org/10.1016/j.proci.2014.05.143.
[6] Aro, C. J., and Rodrigue, G. H. Preconditioned time differencing for stiff ODEs in diurnal
atmospheric kinetics. Comput. Phys. Commun. 92, 1 (1995), 27–53.
[7] Ashurst, W., Kerstein, A., Kerr, R., and Gibson, C. Alignment of vorticity and scalar
gradient with strain rate in simulated Navier-Stokes turbulence. Phys. Fluids 30 (1987), 2343–
2353.
194
[8] Aspden, A., Bell, J., and Woosley, S. Distributed flames in type Ia supernovae. Astro-
phys. J. 710 (2010), 1654–1663.
[9] Aspden, A., Day, M., and Bell, J. Lewis number effects in distributed flames. Proc.
Comb. Inst. 33 (2011), 1473–1480.
[10] Aspden, A., Day, M., and Bell, J. Turbulence-flame interactions in lean premixed hydro-
gen: transition to the distributed burning regime. J. Fluid Mech. 680 (2011), 287–320.
[11] Aspden, A., Day, M., and Bell, J. Turbulence-chemistry interaction in lean premixed
hydrogen combustion. Proc. Comb. Inst. 35, 2 (2015), 1321–1329.
[12] Bagrinovskii, K., and Godunov, S. Difference schemes for multidimensional problems (in
Russian). Dokl. Akad. Nauk. USSR 115 (1957), 431–433.
[13] Balarac, G., Pitsch, H., and Raman, V. Modeling of the subfilter scalar dissipation rate
using the concept of optimal estimators. Phys. Fluids 20 (2008), 091701.
[14] Barlow, R., Frank, J., Karpetis, A., and Chen, J.-Y. Piloted methane/air jet flames:
Transport effects and aspects of scalar structure. Combust. Flame 143 (2005), 433–449.
[15] Bhagatwala, A., Chen, J. H., and Lu, T. Direct numerical simulations of HCCI/SACI
with ethanol. Combust. Flame 161 (2014), 1826–1841.
[16] Bisetti, F. Integration of large chemical kinetic mechanisms via exponential methods with
Krylov approximations to Jacobian matrix functions. Combust. Theory Model. 16, 3 (2012),
387–418.
[17] Bisetti, F., Blanquart, G., Mueller, M., and Pitsch, H. On the formation and early
evolution of soot in turbulent nonpremixed flames. Combust. Flame 159 (2012), 317–335.
[18] Blanquart, G. CaltechMech v2.1. Available at http://theforce.caltech.edu/resources/.

195
[19] Blanquart, G., Pepiot-Desjardins, P., and Pitsch, H. Chemical mechanism for high
temperature combustion of engine relevant fuels with emphasis on soot precursors. Combust.
Flame 156, 3 (Mar. 2009), 588–607.
[20] Borghi, R. Recent Advances in the Aerospace Science. Plenum, New York, 1985, ch. On the
structure and morphology of turbulent premixed flames, pp. 117–138.
[21] Bradley, D., Lau, A., and Lawes, M. Flame stretch rate as a determinant of turbulent
burning velocity. Phil. Trans. R. Soc. Lond. 338 (1992), 359–387.
[22] Bradley, D., Lawes, M., Liu, K., and Mansour, M. Measurements and correlations
of turbulent burning velocities over wide ranges of fuels and elevated pressures. Proc. Comb.
Inst. 34 (2013), 1516–1526.
[23] Brenan, K. E., and Campbell, S. L. A Description of DASSL: a differential/algebraic
equation solver. In Scientific Computing. North-Holland, Amsterdam, The Netherlands, 1983,
pp. 65–68.
[24] Brenblatt, G., Zel’dovich, Y., and Istratov, A. On diffusional thermal stability of
laminar flame. Prikl. Mekh. Tekh. Fiz. 2 (1962), 21–26.
[25] Brown, P. N., Byrne, G. D., and Hindmarsh, A. C. VODE: A variable-coefficient ODE
solver. SIAM J. Sci. Stat. Comput. 10, 5 (1989), 1038–1051.
[26] Brown, P. N., Shumaker, D. E., and Woodward, C. S. Fully implicit solution of large-
scale non–equilibrium radiation diffusion with high order time integration. J. Comput. Phys.
204, 2 (Apr. 2005), 760–783.
[27] Candler, G., and Olynick, D. Hypersonic flow simulations using a diagonal implicit
method. In Computing Methods in Applied Sciences and Engineering, R. Glowinski, Ed. Nova
Science Publishers, 1991, pp. 29–47.
[28] Candler, G., Subbareddy, P., and Nompelis, I. Decoupled implicit method for aerother-
modynamics and reacting flows. AIAA J. 51, 5 (2013), 1245–1254.

196
[29] Cao, R., and Pope, S. The influence of chemical mechanisms on PDF calculations of
nonpremixed piloted jet flames. Combust. Flame 143 (2005), 450–470.
[30] Carbone, F., Smolke, J., Fincham, A., and Egolfopoulos, F. Characteristics of
piloted premixed turbulent-jet flames of methane and C6 -C8 hydrocarbons. In Spring Technical
Meeting of the WSSCI (Pasadena, USA, March 2014).
[31] Carroll, P., and Blanquart, G. A proposed modification to Lundgren’s physical space
velocity forcing method for isotropic turbulence. Phys. Fluids 25 (2013), 105114.
[32] Carroll, P., and Blanquart, G. The effect of velocity field forcing techniques on the
Karman-Howarth equation. J. Turbul. 15 (2014), 429–448.
[33] Chakraborty, N., and Cant, R. Unsteady effects of strain rate and curvature on turbulent
premixed flames in an inflow-outflow configuration. Combust. Flame 137 (2004), 129–147.
[34] Chakraborty, N., Hawkes, E., Chen, J., and Cant, R. The effects of strain rate and
curvature on surface density function in turbulent premixed methane-air and hydrogen-air
flames: a comparative study. Combust. Flame 154 (2008), 259–280.
[35] Chen, J., and Im, H. Correlation of flame speed with stretch in turbulent premixed
methane/air flames. Twenty-Seventh Symposium (International) on Combustion 1-2 (1998),
819–826.
[36] Chen, J., and Im, H. Stretch effects on the burning velocity of turbulent premixed hydro-
gen/air flames. Proc. Comb. Inst. 28 (2000), 211–218.
[37] Chen, J. H. Petascale direct numerical simulation of turbulent combustion - fundamental
insights towards predictive models. Proc. Comb. Inst. 33, 1 (2011), 99–123.
[38] Chen, Y., and Bilger, R. Experimental investigation of three-dimensional flame-front
structure in premixed turbulent combustion - II. Lean hydrogen/air Bunsen flames. Combust.
Flame 138 (2004), 155–174.

197
[39] Chen, Z., Burke, M. P., and Ju, Y. Effects of Lewis number and ignition energy on the
determination of laminar flame speed using propagating spherical flames. Proc. Comb. Inst.
32, 1 (2009), 1253–1260.
[40] Cheng, R., Littlejohn, D., Nazeer, W., and Smith, K. Laboratory studies of the flow
field characteristics of low-swirl injectors for adaptation to fuel-flexible turbines. J. Eng. Gas
Turb. Power 130 (2008), 021501.
[41] Cheng, R., Littlejohn, D., Strakey, P., and Sidwell, T. Laboratory investigations of
low-swirl injectors with H2 and CH4 at gas turbine conditions. Proc. Comb. Inst. 32 (2009),
3001–3009.
[42] Clavin, P., and Joulin, G. High-frequency response of premixed ames to weak stretch and
curvature: a variable-density analysis. Combust. Theory Modelling 1 (1997), 429–446.
[43] Clavin, P., and Williams, F. Effects of molecular diffusion and of thermal expansion on the
structure and dynamics of premixed flames in turbulent flows of large scale and low intensity.
J. Fluid Mech. 116 (1982), 251–282.
[44] Colket, M., Edwards, T., Williams, S., Cernansky, N., Miller, D., Egolfopoulos,
F., Lindstedt, P., Seshadri, K., Dryer, F., Law, C., Friend, D., Lenhert, D.,
Pitsch, H., Sarofim, A., Smooke, M., and Tsang, W. “Development of an experimental
database and kinetic models for surrogate jet fuels” in 45th AIAA Aerospace Sciences Meeting
and Exhibit (Reno, Nevada) (Jan. 8-11 2007).
[45] Cook, A., and Riley, J. Subgrid-scale modeling for turbulent, reacting flows. Combust.
Flame 112 (1998), 593–606.
[46] Curran, H.J. Gaffuri, P., Pitz, W., and Westbrook, C. A comprehensive modeling
study of n-heptane oxidation. Combust. Flame 114 (1998), 149–177.
[47] Dagaut, P., and Cathonnet, M. The ignition, oxidation, and combustion of kerosene: A
review of experimental and kinetic modeling. Prog. Energy Combust. Sci. 32 (2006), 48–92.
198
[48] Damkhöler, G. Der einfluss der turbulenz auf die flammengeschwindigkeit. Z. Elektrochem.
46 (1940), 601–652.
[49] D’Angelo, Y., and Larrouturou, B. Comparison and analysis of some numerical schemes
for stiff complex chemistry problems. Modélisation mathématique et analyse numérique 29, 3
(1995), 259–301.
[50] Daniele, S., Mantzaras, J., Jansohn, P., Denisov, A., and Boulouchos, K. Flame
front/turbulence interaction for syngas fuels in the thin reaction zones regime: turbulent and
stretched laminar flame speeds at elevated pressures and temperatures. J. Fluid Mech. 724
(2013), 36–68.
[51] Davis, S., and Law, C. Determination of and fuel structure effects on laminar flame speeds
of C1 to C8 hydrocarbons. Comb. Sci. Tech. 140 (1998), 427–449.
[52] Day, M., and Bell, J. Numerical simulation of laminar reacting flows with complex chem-
istry. Combust. Theory Model. 4, 4 (1999), 535–556.
[53] Day, M., Bell, J., Bremer, P.-T., Pascucci, V., Beckner, V., and Lijewski, M.
Turbulence effects on cellular burning structures in lean premixed hydrogen flames. Combust.
Flame 156 (2009), 1035–1045.
[54] Dennis Jr, J. E., and Schnabel, R. B. Numerical methods for unconstrained optimization
and nonlinear equations, vol. 16. SIAM, 1996.
[55] Desjardins, O., Blanquart, G., Balarac, G., and Pitsch, H. High order conservative
finite difference scheme for variable density low Mach number turbulent flows. J. Comput.
Phys. 227, 15 (2008), 7125–7159.
[56] Dworkin, S., Smooke, M., and Giovangigli, V. The impact of detailed multicomponent
transport and thermal diffusion effects on soot formation in ethylene/air flames. Proc. Comb.
Inst. 32, 1 (2009), 1165–1172.

199
[57] Eberhardt, S., and Imlay, S. Diagonal implicit scheme for computing flows with finite
rate chemistry. J. Thermophys. Heat Tr. 6, 2 (1992), 208–216.
[58] Echekki, T., and Chen, J. Unsteady strain rate and curvature effects in turbulent premixed
methane-air flames. Combust. Flame 106 (1996), 184–202.
[59] Ern, A., and Giovangigli, V. Impact of multicomponent transport on planar and coun-
terflow hydrogen/air and methane/air flames. Comb. Sci. Tech. 149 (1999), 157–181.
[60] Eucken, A. The heat-carrying capabilities, the specific heat and the internal friction of gas.
Phys. Z. 14 (1913), 324–333.
[61] Falgout, R. D., and Yang, U. M. HYPRE: A library of high performance preconditioners.
In Computational Science—ICCS 2002. Springer, 2002, pp. 632–641.
[62] Ferreira, J. Flamelet modelling of stabilization in turbulent non-premixed combustion. Dis-
sertation, ETH Zürich, 1996.
[63] Fiorina, B., Vicquelin, R., Auzillon, P., Darabiha, N., Gicquel, O., and Vey-
nante, D. A filtered tabulated chemistry model for LES of premixed combustion. Combust.
Flame 157, 3 (2010), 465–475.
[64] Gicquel, L., Staffelbach, G., and Poinsot, T. Large Eddy Simulations of gaseous
flames in gas turbine combustion chambers. Prog. Energy Combust. Sci. 38, 6 (2012), 782–
817.
[65] Gicquel, O., Darabiha, N., and Thévenin, D. Laminar premixed hydrogen/air coun-
terflow flame simulations using flame prolongation of ILDM with differential diffusion. Proc.
Comb. Inst. 28 (2000), 1901–1908.
[66] Gou, X., Sun, W., Chen, Z., and Ju, Y. A dynamic multi-timescale method for combustion
modeling with detailed and reduced chemical kinetic mechanisms. Combust. Flame 157, 6
(2010), 1111–1121.
200
[67] Goyal, G., Paul, P., Mukunda, H., and Deshpande, S. Time dependent operator-split
and unsplit schemes for one dimensional premixed flames. Comb. Sci. Tech. 60, 1-3 (1988),
167–189.
[68] Gruber, A., Sankaran, E. R., Hawkes, E. R., and Chen, J. Turbulent flame-wall
interaction: a direct numerical simulation study. J. Fluid Mech. 658 (2010), 5–32.
[69] Halter, F., Chauveau, C., and Gökalp, I. Characterization of the effects of hydrogen
addition in premixed methane/air flames. Int. J. Hydrogen Energy 32 (2007), 2585–2592.
[70] Hawkes, E., Chatakonda, O., Kolla, H., Kerstein, A., and Chen, J. A petascale
direct numerical simulation study of the modelling of flame wrinkling for large-eddy simulations
in intense turbulence. Combust. Flame 159 (2012), 2690–2703.
[71] Hawkes, E., and Chen, J. Comparison of direct numerical simulation of lean premixed
methane-air flames with strained laminar flame calculations. Combust. Flame 144 (2006),
112–125.
[72] Hawkes, E., Sankaran, R., Sutherland, C., and Chen, J. Direct numerical simulation
of turbulent combustion: fundamental insights towards predictive models. J. Phys. Conf. Ser.
16 (2005), 65–79.
[73] Haworth, D., and Poinsot, T. Numerical simulations of Lewis number effects in turbulent
premixed flames. J. Fluid Mech. 244 (1992), 405–436.
[74] Herrmann, M., Blanquart, G., and Raman, V. Flux corrected finite volume scheme
for preserving scalar boundedness in reacting large–eddy simulations. AIAA Journal 44, 12
(2006), 2879–2886.
[75] Hiremath, V., Lantz, S., Wang, H., and Pope, S. Large-scale parallel simulations of
turbulent combustion using combined dimension reduction and tabulation of chemistry. Proc.
Comb. Inst. 34 (2013), 205–215.

201
[76] Hirschfelder, O., Curtiss, C., and Bird, R. Molecular Theory of Gases and Liquids.
John Wiley and Sons, New York, 1954.
[77] Hong, Z., Davidson, D., and Hanson, R. An improved H-2/O-2 mechanism based on
recent shock tube/laser absorption measurements. Combust. Flame 158, 4 (2011), 633–644.
[78] Hult, J., Gashi, S., Chakraborty, N., Klein, M., Jenkins, K., Cant, S., and Kamin-
ski, C. Measurement of flame surface density for turbulent premixed flames using PLIF and
DNS. Proc. Comb. Inst. 31 (2007), 1319–1326.
[79] Ihme, M., and Pitsch, H. Prediction of extinction and reignition in nonpremixed turbulent
flames using a flamelet/progress variable model. 1. A priori study and presumed PDF closure.
Combust. Flame 155, 1-2 (Oct. 2008), 70–89.
[80] Ihme, M., and Pitsch, H. Prediction of extinction and reignition in nonpremixed turbulent
flames using a flamelet/progress variable model. 2. A posteriori study with application to
Sandia flames D and E. Combust. Flame 155, 1-2 (Oct. 2008), 90–107.
[81] Im, H., and Chen, J. Effects of flow transients on the burning velocity of laminar hydro-
gen/air premixed flames. Proc. Comb. Inst. 28 (2000), 1833–1840.
[82] Im, H., and Chen, J. Preferential diffusion effects on the burning rate of interacting turbulent
premixed hydrogen-air flames. Combust. Flame 131 (2002), 246–258.
[83] Jay, L. Inexact simplified newton iterations for implicit runge-kutta methods. SIAM J.
Numer. Anal. 38, 4 (2000), 1369–1388.
[84] Joulin, G. On the response of premixed flames to time-dependent stretch and curvature.
Combust. Sci. Technol. 97 (1994), 219–229.
[85] Ju, Y. Lower-upper scheme for chemically reacting flow with finite rate chemistry. AIAA J.
33, 8 (1995), 1418–1425.

202
[86] Ju, Y., Sun, W., Burke, M., Gou, X., and Chen, Z. Multi-timescale modeling of
ignition and ame regimes of n-heptane-air mixtures near spark assisted homogeneous charge
compression ignition conditions. Proc. Comb. Inst. 33, 1 (2011), 1245–1251.
[87] Kailasanathan, R. K. A., Book, E., Fang, T., and Roberts, W. Hydrocarbon species
concentrations in nitrogen diluted ethylene-air laminar jet diffusion flames at elevated pres-
sures. Proc. Comb. Inst. 34, 1 (2013), 1035–1043.
[88] Kailasanathan, R. K. A., Yelverton, T., Fang, T., and Roberts, W. Effect of
diluents on soot precursor formation and temperature in ethylene laminar diffusion flames.
Combust. Flame 160, 3 (2013), 656–670.
[89] Kanevsky, A., Carpenter, M., Gottlieb, D., and Hesthaven, J. Application of
implicit-explicit high order Runge-Kutta methods to discontinuous-Galerkin schemes. J. Com-
put. Phys. 225 (2007), 1753–1781.
[90] Kassam, A.-K., and Trefethen, L. N. Fourth-order time-stepping for stiff PDEs. SIAM
J. Sci. Comput. 26, 4 (2005), 1214–1233.
[91] Kee, R., Warnatz, J., and Miller, J. A Fortran computer code package for the evaluation
of gas-phase viscosities, conductivities, and diffusion coefficients. Tech. rep., Sandia National
Laboratories, 1983.
[92] Kelly, A., Smallbone, A., Zhu, D., and Law, C. Laminar flame speeds of C5 to C8
n-alkanes at elevated pressures and temperatures. In 48th AIAA Aerospace Sciences Meeting
(Orlando, USA, January 2010).
[93] Kennedy, C. A., and Carpenter, M. H. Additive Runge-Kutta schemes for convection-
diffusion-reaction equations. Appl. Numer. Math. 44, 1–2 (2003), 139–181.
[94] Kitagawa, T., Nakahara, T., Maruyama, K., Kado, K., Hayakawa, A., and
Kobayashi, S. Turbulent burning velocity of hydrogen-air premixed propagating flames at
elevated pressures. Int. J. Hydrogen Energy 33 (2008), 5842–5849.

203
[95] Knio, O. M., Najm, H. N., and Wyckoff, P. S. A semi-implicit numerical scheme for
reacting flow: II. Stiff, operator-split formulation. J. Comput. Phys. 154, 2 (1999), 428–467.
[96] Knudsen, E., Kolla, H., Hawkes, E., and Pitsch, H. LES of a premixed jet flame DNS
using a strained flamelet model. Combust. Flame 160 (2013), 29112927.
[97] Knudsen, E., and Pitsch, H. A general flamelet transformation useful for distinguishing
between premixed and non-premixed modes of combustion. Combust. Flame 156 (2009), 678–
696.
[98] Knudsen, E., and Pitsch, H. Capabilities and limitations of multi-regime flamelet com-
bustion models. Combust. Flame 159 (2012), 242–264.
[99] Kobayashi, H., Kawabata, Y., and Maruta, K. Experimental study on general correla-
tion of turbulent burning velocity at high pressure. Proc. Comb. Inst. 27 (1998), 941–948.
[100] Kobayashi, H., Otawara, Y., Wang, J., Matsuno, F., Ogami, Y., Okuyama, M.,
Kudo, T., and Kadowaki, S. Turbulent premixed flame characteristics of a CO/H2 /O2
mixture highly diluted with CO2 in a high-pressure environment. Proc. Comb. Inst. 34 (2013),
1437–1445.
[101] Lam, S. Singular perturbation for stiff equations using numerical methods. In Recent Advances
in the Aerospace Sciences. Springer, 1985, pp. 3–19.
[102] Lam, S., and Coussis, D. Understanding complex chemical kinetics with computational
singular perturbation. In Symposium (International) on Combustion (1989), vol. 22, Elsevier,
pp. 931–941.
[103] Li, J., Zhao, Z., Kazakov, A., and Dryer, F. L. An updated comprehensive kinetic
model of hydrogen combustion. Int. J. Chem. Kinet. 36, 10 (2004), 566–575.
[104] Lignell, D. O., Chen, J. H., and Smith, P. J. Three-dimensional direct numerical
simulation of soot formation and transport in a temporally evolving nonpremixed ethylene jet
flame. Combust. Flame 155, 1–2 (2008), 316–333.

204
[105] Lignell, D. O., Chen, J. H., Smith, P. J., Lu, T., and Law, C. K. The effect of
flame structure on soot formation and transport in turbulent nonpremixed flames using direct
numerical simulation. Combust. Flame 151, 1–2 (2007), 2–28.
[106] Lu, T., and Law, C. K. A directed relation graph method for mechanism reduction. Proc.
Comb. Inst. 30, 1 (2005), 1333–1341.
[107] Lu, T., and Law, C. K. A criterion based on computational singular perturbation for the
identification of quasi steady state species: A reduced mechanism for methane oxidation with
NO chemistry. Combust. Flame 154, 4 (2008), 761–774.
[108] Lu, T., and Law, C. K. Strategies for mechanism reduction for large hydrocarbons: n-
heptane. Combust. Flame 154, 1–2 (2008), 153–163.
[109] Lu, T., and Law, C. K. Toward accommodating realistic fuel chemistry in large-scale
computations. Prog. Energy Combust. Sci. 35, 2 (2009), 192–215.
[110] Lu, T., Law, C. K., Yoo, C. S., and Chen, J. H. Dynamic stiffness removal for direct
numerical simulations. Combust. Flame 156, 8 (2009), 1542–1551.
[111] Lundgren, T. “Linear forced isotropic turbulence” in Annual Research Briefs (Center for
Turbulence Research, Stanford) (2003) 461-473.
[112] Maas, U., and Pope, S. Simplifying chemical kinetics: Intrinsic low-dimensional manifolds
in composition space. Combust. Flame 88, 3–4 (1992), 239–264.
[113] Markstein, G. Cell structure of propane flames burning in tubes. J. Chem. Phys. 17 (1949),
428.
[114] Matalon, M., and Matkowsky, B. Flames as gasdynamic discontinuities. J. Fluid Mech.
124 (1982), 239–259.
[115] Mathur, S., Tondon, P., and Saxena, S. Thermal conductivity of binary, ternary and
quaternary mixtures of rare gases. Mol. Phys. 12, 6 (1967), 569–579.

205
[116] McNenly, M. J., Whitesides, R. A., and Flowers, D. L. Faster solvers for large kinetic
mechanisms using adaptive preconditioners. Proc. Comb. Inst. 35, 1 (2015), 581–587.
[117] Mehl, M., Pitz, W. J., Westbrook, C. K., and Curran, H. J. Kinetic modeling of
gasoline surrogate components and mixtures under engine conditions . Proc. Comb. Inst. 33
(2011), 193 – 200.
[118] Menon, S., Boettcher, P., and Blanquart, G. Enthalpy based approach to capture
heat transfer effects in premixed combustion. Combust. Flame 160 (2002), 1242–1253.
[119] Moreau, A., Teytaud, O., and Bertoglio, J. Optimal estimation for large-eddy simula-
tion of turbulence and application to the analysis of subgrid models. Phys. Fluids 18 (2006),
105101.
[120] Mueller, C., Driscoll, J., Sutkus, D., Roberts, W., Drake, M., and Smooke, M.
Effect of unsteady stretch rate on OH chemistry during a flame-vortex interaction: to assess
flamelet models. Combust. Flame 100 (1995), 323–331.
[121] Mueller, M. A., Kim, T. J., Yetter, R. A., and Dryer, F. L. Flow reactor studies
and kinetic modeling of the H2 /O2 reaction. Int. J. Chem. Kinet. 31, 2 (1999), 113–125.
[122] Mueller, M. E., and Pitsch, H. LES model for sooting turbulent nonpremixed flames.
Combust. Flame 159, 6 (June 2012), 2166–2180.
[123] Mukhopadhyay, S. Modeling turbulent combustion using filtered flamelets. Dissertation,
Technische Universiteit Eindhoven, 2014.
[124] Muller, J.-M. Elementary functions: algorithms and implementation, 2 ed. Birkhäuser,
2005.
[125] Muppala, S., Papalexandris, M., Manickam, B., Aluri, N, K., and Dinkelacker,
F. Numerical simulation of lean premixed turbulent hydrogen/hydrocarbon flames at elevated
pressures. In 10th International Workshop on Premixed Turbulent Flames (Mainz, Germany,
August 2006).
206
[126] Mydlarski, L., and Warhaft, Z. On the onset of high-Reynolds-number grid-generated
wind tunnel turbulence. J. Fluid Mech. 320 (1996), 331–368.
[127] Najm, H., and Wyckoff, P. Premixed flame response to unsteady strain rate and curvature.
Combust. Flame 110 (1997), 92–112.
[128] Najm, H. N., Wyckoff, P. S., and Knio, O. M. A semi-implicit numerical scheme for
reacting flow: I. Stiff chemistry. J. Comput. Phys. 143, 2 (1998), 381–402.
[129] Niemeyer, K., and Sung, C. Mechanism reduction for multicomponent surrogates: A case
study using toluene reference fuels. Combust. Flame 161, 11 (2014), 2752–2764.
[130] Ober, C. C., and Shadid, J. N. Studies on the accuracy of time-integration methods for
the radiation-diffusion equations. J. Comput. Phys. 195, 2 (2004), 743–772.
[131] Park, C., and Yoon, S. Fully coupled implicit method for thermochemical nonequilibrium
air at suborbital flight speeds. AIAA J. 28, 1 (1991), 31–39.
[132] Pearson, K. Notes on regression and inheritance in the case of two parents. P. R. Soc.
London 58 (1895), 240–242.
[133] Pelce, P., and Clavin, P. Influence of hydrodynamics and diffusion upon the stability
limits of laminar premixed flames. J. Fluid Mech. 124 (1982), 219–237.
[134] Pepiot-Desjardins, P., and Pitsch, H. An efficient error-propagation-based reduction
method for large chemical kinetic mechanisms. Combust. Flame 154, 1 (2008), 67–81.
[135] Perini, F., Galligani, E., and Reitz, R. D. A study of direct and Krylov iterative sparse
solver techniques to approach linear scaling of the integration of chemical kinetics with detailed
combustion mechanisms. Combust. Flame 161, 5, 1180–1195.
[136] Peters, N. Local quenching due to flame stretch and non-premixed turbulent combustion.
Combust. Sci. Technol 30, 1 (1983), 1–17.

207
[137] Peters, N. Laminar flamelet concepts in turbulent combustion. Twenty-First Symposium
(International) on Combustion (1986), 1231–1250.
[138] Peters, N. Numerical Approaches to Combustion Modelling, Progress in Astronautics and
Aeronautics, Vol. 135. 1991, ch. Length scales in laminar and turbulent flames, pp. 155–182.
[139] Peters, N. Reducing mechanisms. In Reduced kinetic mechanisms and asymptotic approx-
imations for methane-air flames, M. D. Smooke, Ed., vol. 384 of Lecture Notes in Physics.
Springer Berlin Heidelberg, 1991, pp. 48–67.
[140] Peters, N. Fifteen lectures on laminar and turbulent combustion. Ercoftac Summer School,
September 1992.
[141] Peters, N. The turbulent burning velocity for large-scale and small-scale turbulence. J.
Fluid Mech. 384 (1999), 107–132.
[142] Peters, N. Turbulent Combustion. Cambridge University Press, Cambridge, 2000.
[143] Petzold, L. R. A Description of DASSL: a differential/algebraic system solver. Sandia
National Laboratories Report, SAND282-8637 (1982).
[144] Pierce, C., and Moin, P. Progress-variable approach for large-eddy simulation of non-
premixed turbulent combustion. J. Fluid Mech. 504 (2004), 73–97.
[145] Pierce, C. D. Progress-variable approach for large-eddy simulation of turbulent combustion.
PhD thesis, Stanford University, 2001.
[146] Pitsch, H. Flamemaster: A C++ computer program for 0D combustion and 1D laminar flame
calculations. available at http://www.itv.rwth-aachen.de/downloads/flamemaster/. 1998.
[147] Pitsch, H. A consistent level set formulation for large-eddy simulation of premixed turbulent
combustion. Combust. Flame. 143 (2005), 587–598.
[148] Pitsch, H. Large-eddy simulation of turbulent combustion. Ann. Rev. Fluid Mech. 38 (2006),
453–482.
208
[149] Pitsch, H., and Steiner, H. Large-eddy simulation of a turbulent piloted methane/air
diffusion flame (sandia flame D). Phys. Fluids 12 (2000), 2541–2554.
[150] Poinsot, T., Veynante, D., and Candel, S. Diagrams of premixed turbulent combustion
based on direct simulation. Twenty-Third Symposium (International) on Combustion (1990),
613–619.
[151] Poludnenko, A., and Oran, E. The interaction of high-speed turbulence with flames:
Global properties and internal flame structure. Combust. Flame 157 (2010), 995–1011.
[152] Pope, S. Turbulent Flows. Cambridge University Press, Cambridge, 2000.
[153] Quinlan, J., McDaniel, J., Drozda, T., Lacaze, G., and Oefelein, J. A
priori analysis of flamelet-based modeling for a dual-mode scramjet combustor. 50th
AIAA/ASME/SAE/ASEE Joint Propulsion Conference 2014 (2014), 2879–2886.
[154] Richardson, L. F. The approximate arithmetical solution by finite differences of physical
problems involving differential equations, with an application to the stresses in a masonry
dam. Philos. Trans. R. Soc. Lond. A 210 (1911), 307–357.
[155] Robertson, H. The solution of a set of reaction rate equations. Numerical analysis: an
introduction (1966), 178–182.
[156] Ropp, D. L., Shadid, J. N., and Ober, C. C. Studies of the accuracy of time integration
methods for reaction-diffusion equations. J. Comput. Phys. 194, 2 (2004), 544–574.
[157] Rosales, C., and Meneveau, C. Linear forcing in numerical simulations of isotropic tur-
bulence: Physical space implementations and convergence properties. Phys. Fluids 17 (2005),
095106.
[158] Sandu, A., Verwer, J., Blom, J., Spee, E., Carmichael, G., and Potra, F. Bench-
marking stiff ode solvers for atmospheric chemistry problems II: Rosenbrock solvers. Atmos.
Environ. 31, 20 (1997), 3459–3472.

209
[159] Sandu, A., Verwer, J., Loon, M. V., Carmichael, G., Potra, F., Dabdub, D., and
Seinfeld, J. Benchmarking stiff ode solvers for atmospheric chemistry problems-I. implicit
vs explicit. Atmos. Environ. 31, 19 (1997), 3151–3166. EUMAC: European Modelling of
Atmospheric Constituents.
[160] Sankaran, R., Hawkes, E., Chen, J., Lu, T., and Law, C. Structure of a spatially
developing turbulent lean methane-air bunsen flame. Proc. Comb. Inst. 31 (2007), 1291–1298.
[161] Saylor, R. D., and Ford, G. D. On the comparison of numerical methods for the inte-
gration of kinetic equations in atmospheric chemistry and transport models. Atmos. Environ.
29, 19 (1995), 2585–2593.
[162] Shunn, L., Ham, F., and Moin, P. Verification of variable-density flow solvers using
manufactured solutions. J. Comput. Phys. 231, 9 (2012), 3801–3827.
[163] Smith, G. P., Golden, D. M., Frenklach, M., Moriarty, N. W., Eite-
neer, B., Goldenberg, M., Bowman, C. T., Hanson, R. K., Song, S., Gar-
diner, W. C., Lissianski, V. V., and Qin, Z. GRI-Mech 3.0. Available at
http://www.me.berkeley.edu/gri mech/.
[164] Smooke, M., Mitchell, R., and Keyes, D. Numerical-solution of 2-dimensional axisym-
metric laminar diffusion flames. Combust. Sci. Technol 67, 4–6 (1989), 85–122.
[165] Staffelbach, G., Gicquel, L., Boudier, G., and Poinsot, T. Large Eddy Simulations
of self excited azimuthal modes in annular combustors. Proc. Comb. Inst. 32 (2009), 2909–
2916.
[166] Strang, G. On the construction and comparison of difference schemes. SIAM J. Numer.
Anal. 5, 3 (1968), 506–517.
[167] Swanson, R., Turkel, E., and Rossow, C.-C. Convergence acceleration of Runge-Kutta
schemes for solving the Navier-Stokes equations. J. Comput. Phys. 224 (2007), 365–388.
210
[168] Székely, G., Rizzo, M., and Bakiroz, N. Measuring and testing dependence by correlation
of distances. Ann. Stat. 35 (2007), 2769–2794.
[169] Tranquilli, P., and Sandu, A. Rosenbrock-Krylov methods for large systems of differential
equations. SIAM J. Sci. Comput. 36, 3 (2014), A1313–A1338.
[170] van Lipzig, J., Nilsson, E., de Goey, L., and Konnov, A. Laminar burning velocities
of n-heptane, iso-octane, ethanol and their binary and tertiary mixtures. Fuel 90 (2011),
2773–2781.
[171] van Oijen, J., Lammers, F., and de Goey, L. Modeling of complex premixed burner
systems by using flamelet-generated manifolds. Combust. Flame 127 (2001), 2124–2134.
[172] Venkateswaran, P., Marshall, A., Shin, D., and Noble, D. Measurements and anal-
ysis of turbulent consumption speeds of H2 /CO mixtures. Combust. Flame 158 (2011), 1602–
1614.
[173] Verma, S., Xuan, Y., and Blanquart, G. An improved bounded semi-Lagrangian scheme
for the turbulent transport of passive scalars. J. Comput. Phys. 272, 0 (2014), 1 – 22.
[174] Viswanathan, S., Wang, H., and Pope, S. Numerical implementation of mixing and
molecular transport in LES/PDF studies of turbulent reacting flows. J. Comp. Phys. 230
(2011), 6916–6957.
[175] Wang, H., and Frenklach, M. A detailed kinetic modeling study of aromatics formation
in laminar premixed acetylene and ethylene flames. Combust. Flame 110, 1–2 (1997), 173–221.
[176] Wang, W., Luo, K., and Fan, J. Direct numerical simulation and conditional statistics of
hydrogen/air turbulent premixed flames. Energ. Fuel 27 (2013), 549–560.
[177] Warnatz, J. Numerical methods in laminar flame propagation. Friedr. Vieweg and Sohn
Verlagsgesellschaft mbH, Braunschweig, ch. Influence of transport models and boundary con-
ditions on flame structure.

211
[178] Wilcox, D. Turbulence Modeling for CFD. DCW Industries, Anaheim, 2000.
[179] Wilke, C. A viscosity equation for gas mixtures. J. Chem. Phys. 18, 4 (1950), 517–519.
[180] Williams, F. A. Combustion Theory 2nd Edition. Addison-Wesley, 1985.
[181] Xin, Y., and Law, C. K. A mechanistic evaluation of Soret diffusion in heptane/air ames.
Combust. Flame 159 (2012), 2345–2351.
[182] Xuan, Y., and Blanquart, G. Numerical modeling of sooting tendencies in a laminar
co-flow diffusion flame. Combust. Flame 160, 9 (2013), 1657–1666.
[183] Xuan, Y., and Blanquart, G. Effects of aromatic chemistry-turbulence interactions on
soot formation in a turbulent non-premixed flame. Proc. Comb. Inst. 35, 2 (2015), 1911–1919.
[184] Xuan, Y., Blanquart, G., and Mueller, M. E. Modeling curvature effects in diffusion
flames using a laminar flamelet model. Combust. Flame 161, 5 (2014), 1294–1309.
[185] Yeung, P., Girimaji, S., and Pope, S. Straining and scalar dissipation on material-surfaces
in turbulence - implications for flamelets. Combust. Flame 79 (1990), 340–365.
[186] Yoo, C. S., Lu, T., Chen, J. H., and Law, C. K. Direct numerical simulations of
ignition of a lean n-heptane/air mixture with temperature inhomogeneities at constant volume:
Parametric study. Combust. Flame 158, 9 (2011), 1727–1741.
[187] Yoo, C. S., Luo, Z., Kim, H., and Chen, J. H. A DNS study of ignition characteristics of
a lean iso-octane/air mixture under HCCI and SACI conditions. Proc. Comb. Inst. 34 (2013),
2985–2993.
[188] Yoo, C. S., Richardson, E., Sankaran, R., and Chen, J. H. A DNS study on the
stabilization mechanism of a turbulent lifted ethylene jet flame in highly-heated coflow. Proc.
Comb. Inst. 33, 1 (2011), 1619–1627.
[189] Yoshida, H. Construction of higher order symplectic integrators. Phys. Lett. A 150, 5 (1990),
262–268.
212
[190] Yu, R., Yu, J., and Bai, X.-S. An improved high-order scheme for DNS of low Mach
number turbulent reacting flows based on stiff chemistry solver. J. Comput. Phys. 231, 16
(2012), 5504–5521.
[191] Zhong, X. Additive semi-implicit Runge–Kutta methods for computing high-speed nonequi-
librium reactive flows. J. Comput. Phys. 128, 1 (1996), 19–31.

Characterization and Modeling of Premixed Turbulent N-Heptane Flames in The Thin Reaction Zone Regime

Uploaded by

Copyright:

Available Formats

Characterization and Modeling of Premixed Turbulent N-Heptane Flames in The Thin Reaction Zone Regime

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Characterization and Modeling of Premixed Turbulent N-Heptane Flames in The Thin Reaction Zone Regime

Uploaded by

Copyright:

Available Formats

Characterization and modeling of premixed turbulent

n-heptane flames in the thin reaction zone regime

In Partial Fulfillment of the Requirements

for the Degree of

California Institute of Technology

(Defended May 12, 2015)

All Rights Reserved

Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231.

my (ex-)office mates who have contributed to parts of this work.

these dimensionally-reduced manifolds.

1.2 Turbulent premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2.1 Laminar premixed flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2.3 Regimes of turbulent premixed flames . . . . . . . . . . . . . . . . . . . . . . 7

1.3 Direct numerical simulation of high Karlovitz premixed turbulent flames . . . . . . . 9

1.4 Time-integration for stiff chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.5 Flame structure and differential diffusion . . . . . . . . . . . . . . . . . . . . . . . . 15

1.6 Reaction zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.8 Objectives and outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 Governing equations and numerical solver 21

2.1 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.1.1 Fluid mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.1.2 Chemical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.2 Numerical algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.2.1 Overview of the numerical solver . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.2.2 Extension to multi-dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.3 Application to turbulent reacting flows . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Semi-implicit time-integration for stiff chemistry 32

3.1 Species vs. temperature equation stiffness . . . . . . . . . . . . . . . . . . . . . . . . 33

3.2 Preconditioned iterative method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.3 Semi-implicit preconditioning for stiff chemistry . . . . . . . . . . . . . . . . . . . . . 34

3.3.1 Proposed preconditioner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.3.2 Extension to multi-dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3.4 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.4.1 One-dimensional premixed flame . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.4.2 Three-dimensional turbulent premixed flame . . . . . . . . . . . . . . . . . . 39

3.5.1 Theoretical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.5.2 Eigenvalue analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.5.3 Convergence of the sub-iterations . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.5.4.1 Theoretical stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.5.4.2 Numerical stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.5.5.1 Order of accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.5.5.2 Magnitude of errors . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.5.6 Mass conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.5.7 Computational efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.6.1.1 Premixed flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3.6.1.2 Non-premixed flames . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.6.2 Advantages over other methods . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.6.2.1 Fully-implicit method . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.6.2.2 Operator-splitting methods . . . . . . . . . . . . . . . . . . . . . . . 67

3.6.2.3 Stiffness removal through QSSA . . . . . . . . . . . . . . . . . . . . 69

4 Direct numerical simulations 76

4.1 Numerical approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.1.1 Flow configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.1.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.1.3 Turbulence forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

4.2 Chemistry and transport models validation . . . . . . . . . . . . . . . . . . . . . . . 82

4.2.1 Reduced mechanism performance . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.2.2 Transport model validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

4.3 Grid refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86