Combustion and Flame 154 (2008) 740–760

www.elsevier.com/locate/combustflame

A dynamic model for the turbulent burning velocity for

large eddy simulation of premixed combustion
E. Knudsen ∗ , H. Pitsch
Department of Mechanical Engineering, Stanford University, Stanford, CA 94305, USA
Received 2 August 2007; received in revised form 17 April 2008; accepted 4 May 2008
Available online 30 July 2008

Abstract
Turbulent premixed combustion is particularly difficult to describe using large eddy simulation (LES). In LES,
premixed flame structures typically exist on subfilter length scales. Consequently, premixed LES models must be
capable of describing how completely unresolved flame structures propagate under the influence of completely
unresolved eddies. This description is usually accomplished through the implementation of a model for the tur-
bulent burning velocity. Here, a dynamic model for describing the turbulent burning velocity in the context of
LES is presented. This model uses a new surface filtering procedure that is consistent with standard LES filtering.
Additionally, it only uses information that comes directly from the flame front. This latter attribute is important
for two reasons. First, it guarantees that the model can be consistently applied when level set methods, where
arbitrary constraints can be imposed on field variables away from fronts, are used to track the flame. Second, it
forces the model to recognize that the physics governing flame front propagation are only valid locally at the front.
Results showing model validation in the context of direct numerical simulation (DNS), and model application in
the context of LES, are presented.
© 2008 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

Keywords: Premixed turbulent combustion; Burning velocity; Large eddy simulation; Level set methods; Dynamic model

1. Introduction are typically much broader than reaction zones, may

Turbulent premixed flames are particularly diffi-
cult to describe in the context of large eddy simulation
(LES). Most industrially relevant premixed flames ex-
ist in either the corrugated flamelets regime or the thin
reactions zones regime [1]. The width of the inner
reaction zone of a flame in these regimes is compa-
rable to, if not smaller than, the Kolmogorov length
scale that describes the size of the smallest turbu-
lent eddies in the flow. Flame preheat zones, which
* Corresponding author.
E-mail address: ewk@stanford.edu (E. Knudsen).
0010-2180/$ – see front matter © 2008 The Combustion Institute. Published by Elsevier Inc. All rights reserved.
doi:10.1016/j.combustflame.2008.05.024
also, in the corrugated flamelets regime, exist on sub-
Kolmogorov length scales. In LES, by definition, the
smallest length scales of a flow are filtered out. As a
result, in industrially relevant regimes the transitions
that occur between unburned and burned states occur
mostly on subfilter scales.
Premixed combustion models for implicitly fil-
tered LES that use standalone progress variable or
finite rate chemistry approaches will thus, it seems,
always fail. All models are limited by the accuracy of
the schemes they use for evaluating gradients, and no
scheme is capable of resolving the sharp subgrid tran-
sitions that occur in premixed implicit LES near flame
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
fronts. Premixed implicit LES models that attempt to
resolve flame structure are therefore especially prone
to numerical errors in the most critical regions of the
flowfield.
A variety of methods have been suggested in the
literature in response to the problem of subgrid transi-
tion. Two of the most widely discussed are level set, or
G-equation, methods, and artificial flame thickening
methods. The dynamically thickened flame model,
for example, uses finite rate chemistry, but addition-
ally broadens local reaction zones so that they can
be resolved on LES meshes [2]. This broadening is
achieved by increasing molecular diffusivities and,
in a proportionate manner so as to keep the laminar
flame speed constant, spreading out the influence of
reaction source terms. Thickened flame models there-
fore eliminate the problem of poor resolution. The
thickening procedure, however, has an important con-
sequence. The widened flame attenuates local turbu-
lence and prevents eddies smaller than the thickening
length scale from influencing the front. This effec-
tively decreases the velocity at which the front propa-
gates and creates the need for a compensating model.
The so-called “efficiency function” that is used acts to
ensure that the flame will propagate at appropriately
large speeds in the presence of turbulence [3]. This ef-
ficiency function may in one sense be viewed as the
empirical introduction of a model describing the tur-
bulent burning velocity.
In level set methods, flame fronts are character-
ized using isocontours of field variables and explicitly
tracked [1,4–7]. At the relevant isocontours, the field
variables are governed by equations describing how
the fronts propagate. Away from the relevant isocon-
tours, smooth gradients are prescribed for the field
variables to ensure numerical resolution of front dy-
namics. In level set methods, the inner reaction zones
of premixed flames are treated as coherent structures.
The effect of the chemical activity that occurs within
these reaction zones appears in the governing front
equation almost entirely as a front propagation speed.
This speed is approximately equivalent to the lami-
nar burning velocity in the fully resolved case. Due to
the coherent treatment of inner reaction zones, level
set approaches suffer from the drawback of not be-
ing able to inherently consider local flame quenching.
Conversely, they offer the advantage of not inducing
any artificial interactions between the heat released
by the flame and the flow field. Both the laminar un-
stretched burning speed and its dependence on the
local flame stretch rate appear as external parameters
in level set methods. These parameters are typically
well described by both experiments and computa-
tional chemical kinetics studies, and can be used with
confidence in simulations. When level sets are used in
the context of LES, however, filtered descriptions of
741
the burning velocity that account for subfilter turbu-
lence are additionally needed.
In non-premixed combustion, the great advantage
that LES offers is that the scalar mixing process is
reasonably resolved [8,9]. In premixed combustion,
the scalar mixing process is also important, but it re-
mains poorly resolved at the flame front in most LES
situations. The turbulent burning velocity, which is
needed in both of the premixed modeling approaches
discussed here, is a quantity that expresses how this
unresolved mixing process interacts with chemistry.
Traditional burning velocity models rely on a series of
coefficients that have been determined through anal-
yses of both experimental and direct numerical simu-
lation (DNS) data [6,10]. These coefficient-based ap-
proaches have been successfully applied in Reynolds
averaged Navier–Stokes (RANS) simulations, where
level set methods offer an alternative to the problem of
reaction rate closure [6,11]. In LES, however, where
instantaneous flame realizations are available, it is
possible to eliminate the use of constant coefficients
by employing dynamic procedures to determine coef-
ficients automatically.
The general dynamic procedure has been em-
ployed by a variety of researchers in their efforts
to deal with closure problems in reacting flows. Di-
rect reaction rate closure is a particularly challeng-
ing problem because reaction rates strongly depend
on the exact flow and chemical conditions that are
present on the smallest turbulent scales. Dynamic pro-
cedures have therefore been applied to combustion
models that use indirect techniques to account for re-
action rates. Two such approaches are the previously
discussed thickened flame model, and flamelet-type
models, where it is assumed that small scale flame
structures can be precomputed as a function of turbu-
lent parameters. Charlette et al. [12,13], for example,
use a dynamic approach to determine a parameter
in the “wrinkling factor” that appears in the thick-
ened flame model. Knikker et al. [14] dynamically
determine a parameter for this same wrinkling factor
but apply it in the context of the flame surface den-
sity model. Chakraborty and Cant [15] developed a
method of dynamically determining the surface aver-
aged curvature that appears in the flame surface den-
sity model.
Additionally, dynamic procedures have been ap-
plied in the context of level set modeling. Im et al.
[16] and Bourlioux et al. [17], for example, proposed
a dynamic propagation model that treats level set field
variables as scalars. Subfilter contributions to flame
propagation speed are determined by evaluating a
burning velocity model at two different filter levels
and comparing the results to differences in the mag-
nitude of the gradient of the level set field variable at
those same two levels. Im et al. [16] claim that this
742 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
approach can be physically interpreted as enforcing
flame consumption conservation among filter levels.
They base this claim on work by Kerstein et al. [4],
in which it is demonstrated that a volume integral of
the magnitude of the gradient of a level set field vari-
able is equivalent to a measure of the total front area
within the volume. It is important to note, however,
that in Kerstein et al.’s work a periodic and constant
density flow field is assumed, and each isocontour of
the level set field variable is treated as an equally valid
representation of the flame front. Under that assump-
tion, volume averaging is equivalent to averaging over
multiple front realizations. For variable density flows,
however, this is not the case.
More recent work has stressed that level set gov-
erning equations are only valid at the field variable
isocontour that they describe, and that traditional av-
eraging procedures therefore cannot be consistently
used [5,6,18]. Additionally, in real flames where den-
sity and vorticity vary through the flame front, it is
inappropriate to use flow field information from away
from the front to describe the propagation of the front
itself. In the present paper, then, a dynamic burning
velocity model is proposed that only considers infor-
mation directly from the 2-D front of interest. Al-
though this model will be developed by taking cues
from the level set approach, and it is therefore par-
ticularly straightforward to implement in that context,
it is not strictly dependent on the use of a level set
method. In general, it is applicable to other premixed
LES techniques. It is important to emphasize at the
outset that some of the quantities that appear in this
model also appear in slightly altered forms in other
premixed models. For example, the variable describ-
ing flame surface area density that is developed for
this model might also be developed using the wrin-
kling factor and the flame surface density that appear
in the flame surface density (FSD) formalism [13,14].
Unlike in the FSD formalism, however, the flame sur-
face area density that is used in this model is evaluated
using only information that comes from the 2-D flame
surface. This information can be generated from sim-
ple geometric knowledge of the surface, and no equa-
tions describing a transported flame surface density
need to be introduced. In general, the strength of this
new model is that it is consistent with the level set
method, and this consistency allows some of the well
known quantities related to the turbulent burning ve-
locity to be viewed and calculated in interesting ways.
The new model requires the use of a unique filter-
ing procedure that is developed and presented in Sec-
tion 2. In Section 3, the filtering procedure is applied
to develop the dynamic model. Section 4 presents an
evaluation of the model in the context of DNS. The
model is applied in an LES in Section 5. Brief con-
clusions are offered in Section 6.
2. An equation for the dynamics of premixed
flame fronts
Level set equations can be derived by setting the
substantial derivative of a generic field variable equal
to zero at a surface of interest. The resulting expres-
sion describes how the field variable isocontour asso-
ciated with that surface evolves. In premixed LES, the
derivation of an equation governing flame front be-
havior can be approached in a different way. A flame
front can generally be defined as an isocontour of a
generic progress variable c. This variable might rep-
resent, for example, a non-dimensionalized tempera-
ture. The equation governing the behavior of such a
variable is
 
∂c ∂c 1 ∂ ∂c 1
+ uj = ρD + ω̇R , (1)
∂t ∂xj ρ ∂xj ∂xj ρ
where uj is the local flow velocity in the j th direc-
tion, ρ is the fluid density, D is the diffusivity of the
variable c, and ω̇R is a source term that describes the
effects of chemical reactions. To derive an equation
describing the evolution of a particular c-isosurface,
information from Eq. (1) needs to be extracted di-
rectly from that isosurface, here arbitrarily defined as
c = c0 . This extraction operation can be performed by
multiplying Eq. (1) with a delta function, δ(c − c0 ),
 
∂c ∂c
δ(c − c0 ) + uj
∂t ∂xj
   
1 ∂ ∂c 1
= δ(c − c0 ) ρD + ω̇R . (2)
ρ ∂xj ∂xj ρ
This delta function does not necessarily need to be an
infinitesimally thin Dirac delta. Rather, here δ will be
defined as a normalized Gaussian of finite width. As
long as this width is small compared with the length
scale of the inner reaction zone of the flame being
considered, multiplication with δ(c − c0 ) will effec-
tively give a null result everywhere except at the flame
front. This finite width definition of δ is convenient
because it eliminates the problem of dealing with the
special mathematical properties of the Dirac delta, yet
it remains easy to write.
Just as Dirac delta functions may equivalently be
written as derivatives of Heaviside functions, Gaus-
sians may equivalently be written as derivatives of
error functions. Since the δ function that appears in
Eq. (2) only depends on c, the chain rule may be used
to rewrite the left-hand side of Eq. (2). Remember-
ing that what here will be referred to as the Heavi-
side function H represents an error function of finite
width, this procedure gives
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760 743
 
∂c ∂c
δ(c − c0 ) + uj set variable G. However, while both of these vari-
∂t ∂xj ables are used to describe a front, they differ in an
∂[H (c − c0 )] ∂[H (c − c0 )] important way. The level set variable G can be arbi-
= + uj . (3)
∂t ∂xj trarily defined away from the front it describes, while
the variable G cannot. Since there is nothing arbitrary
To move the delta function on the right-hand side of about the definition of a Heaviside function, G can be
Eq. (2) into the relevant derivatives, the gradient of the directly substituted into Eq. (8), which with Eq. (6)
progress variable must first be written in terms of the produces an equation that is well defined at all loca-
front normal direction at c0 , which will be denoted tions in space,
nj ,
∂G ∂G
∂c ∇c + uj = (Dκ)c0 |∇G| + sL,c0 |∇G|, (10)
= |∇c| = nj |∇c|, (4) ∂t ∂xj
∂xj |∇c|
where
where it is assumed that nj = ∇c/|∇c| points toward   
the burned gases. Use of the product rule on the dif- 1 ∂  
sL,c0 = nj ρD|∇c| + ω̇R .
fusive term then gives [19,20] |∇c|ρ ∂xj c=c0
  (11)
∂ ∂c
ρD Following Peters [6,20], the quantity sL,c0 describes
∂xj ∂xj
the propagation velocity of the isosurface c = c0 ,
∂nj ∂  
= ρD|∇c| + nj ρD|∇c| . (5) which is a function of the density, the normal diffu-
∂xj ∂xj sion term, and the reaction source term in the progress
Finally, the delta function acts on |∇c| as variable equation, as expected. In the absence of cur-
  vature and strain effects, sL,c0 reduces to the un-
δ(c − c0 )|∇c| = δ(c − c0 )∇c stretched 1-D laminar flame speed sL . In realistic


= ∇ H (c − c0 ) . (6) settings, however, these effects exist and might be im-
portant [19,21]. Although they cause sL,c0 to deviate
Applying the results of these manipulations to the from sL , sL is the best first-order approximation that
right-hand side of Eq. (2) produces can be made [6,20]. While it would be possible to
    introduce a more elaborate dependency such as a re-
1 ∂ ∂c 1
δ(c − c0 ) ρD + ω̇R action source term that depends on flame stretch, this
ρ ∂xj ∂xj ρ
would not significantly affect the model development

 1
= Dκ ∇ H (c − c0 )  + δ(c − c0 ) that is to follow.
ρ Because there is nothing arbitrary about its def-
 
∂   inition, G can be volumetrically filtered. In the fol-
× nj ρD|∇c| + ω̇R , (7)
∂xj lowing, the goal will be to develop an equation that
describes the evolution of a filtered G field. This equa-
where κ is the divergence of the normal vector, or
tion will not be numerically tractable for the same rea-
the curvature. Operating on Eq. (1) with a δ function,
sons that filtered progress variable equations are not
then, gives
tractable. It will, however, provide the basis for devel-
∂[H (c − c0 )] ∂[H (c − c0 )] oping a consistent turbulent burning velocity model
+ uj
∂t ∂xj for LES. This will be possible because the filtered

 1 G equation will indicate exactly how to separate re-
= Dκ ∇ H (c − c0 )  + δ(c − c0 ) solved and subfilter propagation components. This
ρ
  burning velocity model can then be used in a level set
∂  
× nj ρD|∇c| + ω̇R . (8) description of the filtered front, or in any other suit-
∂xj able front tracking scheme.
Equation (8) governs the behavior of a Heaviside Defining F (rj ; xj ) to be some appropriately nor-
function that tracks the flame front. malized filter kernel and then applying it to the G field
A new variable will now be introduced for the pur- gives
poses of notational convenience. The variable G will 
be defined as G(xj , t) = F (rj ; xj )G(xj − rj , t) drj (12)
  V
G(xj , t) = H c(xj , t) − c0 . (9) 
 
= F (rj ; xj )H c(xj − rj , t) − c0 drj .
The G symbol is intentionally used to denote this vari-
able because of its close relation to the standard level V (13)
744 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760

This filtering procedure is consistent with LES in the this conservation property produces, considering con-
sense that the same filter kernel F (rj ; xj ) that is used ditioning on the unburned gas for example,
to filter the Navier–Stokes equations can be used to
nj uj,u = nj 
uj
filter a representation of the flame front. It will be as-  
sumed here that the filter kernel F (rj ; xj ) commutes ρu − ρb  
− (Dκ)T ,u + sT ,u pb , (19)
with spatial derivatives. ρ
This same filter may be applied to Eq. (10). Doing where pb is the fraction of gas in the filter volume
so produces that is burned and where the tilde denotes Favre fil-
∂G ∂G tering. The insight gained by this manipulation is that
+ uj = (Dκ)T ,c |∇G| + sT ,c |∇G|, (14) the move from a conditional to an unconditional ve-
∂t ∂xj 0 0
locity produces a term that scales with the burning
where velocity. If the c = c0 surface is chosen to represent
the unburned side of the front, Eq. (19) can be used in
sT ,c |∇G| = sL,c0 |∇G| (15) Eq. (18),
0

and ∂G ∂G
uj,u =uj
∂xj ∂xj
(Dκ)T ,c |∇G| = (Dκ)c0 |∇G| (16)  
0
ρu − ρb  
have been used. Equations (15) and (16) represent the − (Dκ)T ,u + sT ,u pb |∇G|
ρ
introduction of definitions describing the burning ve-
locity associated with a certain filter level, sT ,c , and + Γ u. (20)
0
the filtered curvature, (Dκ)T ,c . Equation (20) can then be substituted into Eq. (14),
0
Since the convective term in Eq. (14) is not known, where again c = c0 will denote unburned condition-
it must also be modeled. This term is interesting be- ing,
cause the velocity component under the filter appears
∂G ∂G
next to the gradient of G. This gradient acts as a δ +uj +Γu
function and the filtered quantity therefore only con- ∂t ∂xj
  
tains information about velocities that are conditioned ρu − ρb
= pb + 1
on the density that coincides with the front location, ρ
ρc=c0 . It therefore must be true that

× (Dκ)T ,u + sT ,u |∇G|. (21)

∂G ∂G The term in the brackets may be further manipulated,
uj = uj,c=c0 . (17)  
∂xj ∂xj ρu − ρb
pb + 1
The conditioning that appears in this relationship ρ
makes it clear that Favre filtered velocities cannot be ρu pb − ρb pb + ρu (1 − pb ) + ρb pb
directly used to model this term. Conditionally fil- =
ρ
tered velocities must instead be employed. Specifi- ρu
cally, the resolved and subfilter contributions to the = , (22)
ρ
convective term will here be separated by introducing
the variable Γ c0 , and the final filtered equation may be written as
∂G ∂G
∂G ∂G +uj +Γu
uj,c=c0 = uj,c=c0 + Γ c0 . (18) ∂t ∂xj
∂xj ∂xj
ρu

= (Dκ)T ,u + sT ,u |∇G|. (23)

Physically, Γ c0 describes how subfilter velocity fluc-
tuations convect the filtered front. Equation (18) can
be treated as the definition of Γ c0 .
The problem with Eq. (18) is that conditionally
filtered velocities are usually not available in realis-
tic simulations. It would therefore be convenient to
transform the conditionally filtered quantity into an
unconditional quantity, and it has in fact been shown
that such a transformation can be derived using a
consistency condition describing mass conservation
through a flame front [5]. Modifying the notation used
in that transformation to match current convention,
j , t) = G0
ρ
Equation (23) will be used below to derive a dynamic
turbulent burning velocity model. It cannot be used
to track a flame front, however, because G is un-
derresolved on implicit LES meshes. Equation (23)
nonetheless informs the use of flame front tracking
methods. Specifically, the derivations of Eq. (10) and
Eq. (23) show explicitly how filtering affects level set
equations. For example, a level set description of a fil-
tered front can be developed using an isocontour of G,
G(x ∀ G(xj , t) = G0 , (24)
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760 745

 
∇ G(x

j , t) = 1 ∀ G(xj , t), (25) process will begin with a discussion of the subfilter
convective terms that appear in the equations.
where the hat on the G variable denotes that it is as-
sociated with a filtered front. A governing equation Unlike the subfilter Reynolds stress term that ap-

that is valid at the front can then be derived by
for G pears in the filtered Navier–Stokes equations, the sub-
taking the material derivative of Eq. (24), filter term Γ u that appears here is relatively insignifi-
cant. Specifically, Γ u describes how subfilter velocity

∂G
DG


∂G fluctuations affect the filtered front. While these ve-
= + uG,j

= 0, (26)
Dt ∂t ∂xj locities tend to wrinkle the front, they act only along
where uG,j is the velocity at which the level set a 2-D surface within the filter volume. Some of these


moves. Since the G0 and G0 surfaces must propagate velocity fluctuations tend to move the front location
in exactly the same fashion, the velocity uG,j

can be
forward, and some tend to move the front location
developed immediately from the filtered front equa- backward. When they are filtered, therefore, they will
tion, Eq. (23). The only intermediate step involves tend to have no effect on the mean front position. The
rewriting this equation using an expanded version unique behavior of this subfilter term is directly due
of |∇G|, to the gradient of G being non-zero only locally at
  the front. This insight further demonstrates the impor-
∂G ρu
∂G
+  uj − (Dκ)T ,u + sT ,u nj
∂t ρ ∂xj
+ Γ u = 0, (27)
where nj in this new context is nj = ∇G/|∇G|. If the
= G0 is to follow the actual filtered front,
levelset G
then uG,j
must explicitly contain all of the terms in velocity fluctuations would certainly exist, and the un-
the parentheses of Eq. (27). It must additionally ac-
count for the subfilter term Γ u .
3. Dynamic propagation model
The existence of an appropriate flame surface fil-
tering procedure, such as the one developed in the pre-
vious section, makes it possible to derive a dynamic
turbulent flame speed identity. This identity will relate
turbulent flame speeds associated with different filter
levels and provide an equation that can be solved for a
model constant. In some respects, this identity is anal-
ogous to Germano’s identity.
When a test-filter, which here will be denoted by
the hat operator, is applied to Eq. (23), that equation
becomes
∂G ∂G
tance of the present filtering approach, since the direct
application of a filter to a level set equation might not
produce an equivalent result.
The point can be illustrated by considering the
release of a non-propagating front in a flowfield of
isotropic turbulence. In this situation, subfilter scale
filtered front would become more and more wrinkled.
The mean front position, however, would remain sta-
tionary. Γ u should therefore be unable to contribute
to front propagation.
In general, subfilter velocity fluctuations can of
course contribute to the movement of a filtered front.
What is important to recognize is that these contribu-
tions are all made in conjunction with another phys-
ical mechanism. For example, velocity fluctuations
might promote an increase in flame surface area that
in turn leads to an increased consumption of unburned
mixture. Or, they might promote the growth of sub-
filter flame curvature, which would lead to increased
heat diffusion and an adjusted burning velocity. The
difference between these cases and the case of pas-
sive front release is the presence of self-sustaining
front propagation. But, since Γ u is uncorrelated with
a propagation speed, the approximations Γ u ≈ 0 and
+
uj + (Γ u )
∂t ∂xj
≈ 0 must hold. DNS data showing the validity
Γ u
of these approximations will be presented in Sec-
ρu ρu
= (Dκ)T ,u |∇G| + s |∇G|. (28) tion 4.
ρ ρ T ,u The broad idea behind deriving a dynamic model
When a filter and a test-filter are applied to Eq. (10) for sT ,u is to produce a front speed that ensures the
as a single operator, a different result is obtained, mean flame front position is independent of the filter
∂
∂

being used. This can be achieved by forcing equations
G G

+

uj +Γ u (28) and (29), which were created using different se-
∂t ∂xj quences of filter application, to produce equivalent
    ρu  
Dκ
∇
s
∇

ρu 
= G + G . (29) front positions. In other words, since the test-filter in


ρ T ,u

ρ T ,u
Eq. (28) commutes with derivatives, subtracting these
Equations (28) and (29) form the basis for the de- two equations produces a relation that will hold at all
velopment of the dynamic model. The development times if filtering is being properly performed,
746 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760

fronts in their level set based study of how premixed

  flames propagate. They leverage the fact that their
∂G ρu ρu

uj − (Dκ)T ,u + sT ,u |∇G| study permits treating multiple level set isocontours
∂xj ρ ρ
  as equally valid representations of flame fronts, and


∂G ρu   
s
∇

ρu go on to integrate this expression over all of these
=

uj − Dκ
+ G . (30)
∂xj
ρ T ,u
ρ T ,u
Like the identities used to describe subfilter stress
models, Eq. (30) does not hold locally and instan-
taneously. In part, this is because the propagation
speeds sT ,u and s
are only defined at their respec- one isosurface, a different approach will be taken.
T ,u
tive fronts, which exist at different spatial locations. It
is also in part due to the need for model constants to
be reasonably bounded, which will not occur if satis-
faction of the equation is forced for all time and space.
Rather, Eq. (30) holds statistically for a given number
of realizations. In dynamic subfilter stress models, an
appropriately large number of realizations are ensured
by averaging the terms in the dynamic equation over
some volume of the flow. The same technique will be
used here.
Given curvature and burning velocity models, the
terms in the parentheses of Eq. (30) are computable.
The terms that depend on G, however, are not. Again,
this is because jumps in G occur over single filter
widths, making the G variable computationally in-
tractable and unavailable. A similar problem occurs
with the gradients of G that appear in the convec-
tive terms. Another means of computing these terms
is therefore needed. An alternative approach can be
developed by starting from the definition of G,
  In the well resolved limit in which a front appears
  
∇G(xj , t) = ∇ F (rj ; xj )
 tor nj = ∇c/|∇c| in Eq. (33) can be brought outside
V
 the integral. Because |nj | = 1, |N j | would then re-
  

× H c(xj − rj , t) − c0 drj 

 
 
 
=  F nj δ(c − c0 )|∇c| drj , (31)
 
V iting case of very large filter widths, a filter volume
where nj = ∇c/|∇c|. In work by Maz’ja [22] and
subsequently Kollmann et al. [23], it has been shown
that the integral in Eq. (31) can be rewritten in terms
of the local flame surface area. Specifically, Kollmann
et al. present the relation

 
f (xj )δ Φ(xj ) − ψ |∇Φ| dxj
V

= f (xj ) dA(xj ), (32)
S,Φ(xj )=ψ
where Φ is a scalar and ψ is the value of that scalar
describing an isocontour of interest. Wenzel and Pe-
ters [24] use this relation to describe the area of flame
isocontours. They then recover Kerstein et al.’s [4]
expression showing that the magnitude of the volu-
metrically integrated level set gradient represents an
ensemble averaged flame area. In the present study
where the goal is to extract information from only
Equation (32) will be applied to only the surface of
interest, and no volumetric averaging will be per-
formed.
Applying Eq. (32) to Eq. (31) produces
 
 
 
 F nj δ(c − c0 )|∇c| dV 
 
V
  
 
 
= F nj dA = |N j |. (33)
 
S,c=c0
N j defines the front conditioned filtered normal vec-
tor. It is important to note that the presence of the filter
kernel F makes N j a volume weighted quantity, and
that there is no restriction on its magnitude as there is
in the case of a typical normal vector. While Eq. (33)
provides a useful way of looking at |∇G|, N j is no
more computationally tractable than |∇G|. N j does,
however, suggest a model for |∇G|.
in a filter volume as a planar surface, the normal vec-
duce to a measure of the front area per filter volume.
As the filter width in Eq. (33) grows, the possibility
of finding subfilter front wrinkling also grows. When
this wrinkling occurs, front normals within the filter
volume will point in misaligned directions. In the lim-
might encompass a series of opposed fronts. Front
normal vectors would then tend to cancel each other
out when integrated, and |N j | would approach zero.
In regimes between the completely resolved and the
completely unresolved limits, the value of |N j | re-
flects a balance between a tendency to decrease in
magnitude due to misaligned normals and a tendency
to increase in magnitude due to increased surface area
per volume.
These tendencies suggest that a link exists be-
tween |N j | and the flame surface density, Σ = |∇c|,
and that models from the flame surface density (FSD)
formalism might be applicable. The presence of δ(c −
c0 ) and nj in Eq. (31), however, differentiate Σ
and |N j |. The difference due to the δ function can
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760 747

be overcome by considering the “fine-grained” flame

surface density, Σ  [14,15], but Σ  does not account  
AΔ ρu ρu AΔ
for the normal vector nj . The “wrinkling factor” Ξ 
uj nj − (Dκ)T ,u + s
VΔ ρ ρ T ,u VΔ
must additionally be introduced to account for nj .

Apparently, |N j | is equivalent to the quantity Σ  /Ξ = uj

A 
nj Δ
[14] in the FSD formalism. This implies that FSD VΔ
models could be introduced to close Eq. (31). Knikker  
ρu  ρu AΔ
et al. [14], for example, model Σ  /Ξ using Σ and a −

Dκ +
T ,u
s
T ,u
, (36)
ρ ρ VΔ
fractal expression. While this approach is an option,
it requires the further development of equations and where Δ denotes the test-filter width. This form of
assumptions regarding Σ . Therefore, in an effort to the equation is very general and can account for
prevent the introduction of further equations, an ap- all premixed regimes except for the broken reaction
proach stemming from the level set framework will zones regime. Nonetheless, it is interesting to con-
be introduced. sider what happens to the equation in simplified lim-
The regime limits that have been described can be iting cases. For example, if convective effects are ne-
satisfied by modeling |N j | using the area of the fil- glected and if the limit of the corrugated flamelets
tered flame front that passes through the local filter regime, where curvature effects are small, is consid-
volume, AΔ , and the filter volume itself, VΔ , ered, the dynamic equation reduces to

AΔ
|∇G| = |N j | ≈ . (34) ρu AΔ ρu AΔ
VΔ sT ,u = s . (37)
ρ VΔ
ρ T ,u VΔ
In the fully resolved limit, the filtered flame area cor-
AΔ is zero at all locations in the test-filter domain that
responds to the actual planar flame area within the
are farther away from the flame front than the filter
filter volume, in exact agreement with Eq. (33). In
width Δ. If a box filter is used and it is assumed that
the poorly resolved limit, AΔ will still represent the
the filtered flame is homogeneous along the directions
area of a planar flame, and thus will be much smaller
parallel to its surface, the test-filtering operation that
than the actual flame area in the filter volume. This re-
acts on the left-hand side of Eq. (37) simply changes
duced area accounts for the effects of the misaligned
VΔ to VΔ , and AΔ to AΔ . AΔ describes the area of
normals. In areas where the flame front does not exist,
the filtered flame in the test-filter volume. This leaves
AΔ is zero, and Eq. (33) is still satisfied. As discussed
ρu ρu
s
above, most dynamic models need to consider multi-
s A = AΔ . (38)
ple realizations of a phenomenon, and a filter volume ρ T ,u Δ
ρ T ,u
might therefore need to encompass a multi-cell region
If the differences between the filtered and test-filtered
of space. In such a case, the area of the entire fil-
density are assumed to be small, the expression can
tered flame front that exists within that region would
be further simplified to
be used to represent AΔ .
With a model prescribed for |∇G|, the convec- sT ,u
AΔ
tive terms appearing in Eq. (30) are easily dealt with. = . (39)
The gradients in these terms must have the same s AΔ
T ,u
magnitude as |∇G|. Furthermore, it would be incon-
This expression is in agreement with Damköhler’s hy-
sistent if they pointed in any direction other than
pothesis that the turbulent flame speed scales with the
the normal vector associated with the filtered front,
turbulent flame surface area [25]. This dynamic model
n̄j = ∇G/|∇G|. Since the previously defined level
may therefore be viewed as an enforcement of a con-
set variable G describes the same front as G, it is
stant rate of flame mass consumption, independent of
acceptable to write nj as the computable quantity the filter being used. This condition satisfies the stated

n̄j = ∇ G/|∇ Therefore,
G|. goal of a having a propagation model that produces
filter-independent mean front positions.
∂G AΔ To apply this procedure to the determination of an

uj ≈
uj nj , (35) unknown constant, actual burning velocity and diffu-
∂xj VΔ
sive propagation models must be selected. A model
and a tractable form of the convective terms becomes for (Dκ)T ,u that considers interactions between re-
available. solved curvature and resolved mixing, and between
These models can now be introduced into a final resolved curvature and subfilter mixing, has been pro-
form of the dynamic equation, posed [5],
748 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760

(Dκ)T ,u = Du κ + Dt,Δ,u κ, (40) means that α must be locally solved for using an
iterative procedure such as Newton’s method. This
where Dt,Δ,u is the turbulent diffusivity conditioned
on the unburned gas density. In this formulation, inter- procedure requires information from both the local
actions between subfilter curvature and subfilter mix- computational cell and the surrounding cells that are
ing are left to be determined by the burning speed used in the test-filtering operation. While more ex-
model, sT ,u . For the remainder of this paper, this pro- pensive than a simple algebraic evaluation, the full
posed model will be adopted. A model for describ- operation is nonetheless still computationally cheap
ing the burning speed itself will be taken from Peters and does not significantly affect the cost of a simula-
[5,6], tion.
Additionally, as discussed earlier in this section,
sT ,u − sL,u
= −γ DaΔ Eq. (44) is not expected to hold instantaneously for
uΔ all flow realizations. The value of α that solves this
 1/2 equation in a practical LES, for example, will occa-
+ (γ DaΔ )2 + γ α DaΔ , (41)
sionally be negative. Negative values of α are incom-
where α and γ are RANS-type model constants, and patible with the burning velocity model that is used
where DaΔ is the Damköhler number associated with here because they can result in a negative value being
the filter width. The model constants have, relative to raised to the 1/2 power. To prevent the occurrence
Peters’ original model, been renamed and regrouped. of negative values of α, a certain amount of aver-
This regrouping permits the appearance of the model aging of the terms in Eq. (44) must be performed.
constant α in the limits of both large and small DaΔ . Just as in the computation of Smagorinsky subfilter
These limits correspond to the corrugated flamelets
viscosity coefficients, averaging can be performed lo-
and thin reaction zones regimes, respectively,
cally, in a homogeneous flow direction [26], or using
α
sT ,u = sL,u + uΔ , (42) more complex techniques [27]. In the remainder of
 2  this work, both the subfilter velocity fluctuation uΔ
 
Dt,Δ,u 1/2 and the filtered flame area per unit volume AΔ /VΔ
sT ,u = sL,u 1 + (γ α)1/2 . (43)
Du will be averaged. The filtered terms on the left-hand
side of Eq. (44) will be averaged before they are test-
In the model validation that follows, Eq. (41) will be
filtered for use in the model. The test-filtered terms on
used to model sT ,u , and α in Eq. (41) will be treated
the right-hand side of Eq. (44) will be averaged after
as a dynamic coefficient. The dynamic identity will
the test-filter is applied. The final form of the dynamic
then take the specific form,
equation, with these averaging procedures included, is
shown in Eq. (45),
AΔ ρu AΔ

uj nj − (Du κ + Dt,Δ,u κ)
VΔ ρ VΔ
   
AΔ ρu AΔ

uj nj − (Du κ + Dt,Δ,u κ)
ρu   VΔ ρ VΔ
− sL,u + uΔ −γ DaΔ
ρ
 1/2  AΔ ρu  
+ (γ DaΔ )2 + γ α DaΔ − sL,u + uΔ  −γ DaΔ 
VΔ ρ
 
A  ρu A   2 1/2  AΔ
uj
=
 nj Δ − (Du κ + Dt,Δ ,u
κ) Δ + γ DaΔ  + γ αDaΔ 
VΔ
ρ VΔ VΔ
ρu      
− sL,u + uΔ −γ DaΔ A  ρu A 

ρ uj
=
 nj Δ − (Du κ + Dt,Δ ,u κ) Δ
VΔ
ρ VΔ
 1/2 
AΔ ρu  

+ (γ DaΔ )2 + γ α DaΔ . (44) − sL,u + uΔ  −γ DaΔ 
VΔ
ρ
Equation (44) is somewhat more complex than a typ-  2 1/2 
+ γ DaΔ  + γ αDaΔ 
ical dynamic equation for a Smagorinsky subfilter
 
viscosity coefficient. This is because the dynamic co- A 
efficient α in Eq. (44) appears in an expression that × Δ , (45)
VΔ
is raised to the 1/2 power. This prevents α from be-
ing explicitly solved for in terms of the other vari- where the · operator denotes the averaging proce-
ables in the equation. In a practical simulation, this dure.
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760 749

Table 1
DNS parameters
Simulation constants Turbulence parameters
Mesh size = 512 × 256 × 256 Reλ = 48
Cell width, x = 1.0 × 10−3 m Ret = 149
ν = 1.87 × 10−5 m2 /s Integral length scale, lt = 32.0 × 10−3 m
ρ = 1.16 kg/m3 Eddy turnover time, τ = 0.32 s
Burning velocity, sL = 0.05, 0.15 m/s Kolmogorov scale, η = 5.0 × 10−4 m
Forcing coefficient, A = 1.038 Largest eddy size, l = 32.0 × 10−3 m

4. Validation in DNS Statistics involving the front are computed using

A DNS of a front propagating in forced isotropic
turbulence is performed to validate this model. The
parameters describing the DNS are shown in Table 1.
Turbulence is forced using the linear scheme of Ros-
ales and Meneveau [28]. The simulation is run using
a constant density flow field, so gas expansion is not
considered. A uniform Cartesian mesh is used, but in
the direction of front propagation the domain length is
doubled. This is done to allow the flame brush thick-
ness to increase to a width of multiple integral scales.
The forced Navier–Stokes equations,
∂uj
= 0,
∂xj
∂(ρui ) ∂(ρuj ui )
+ A parallel, structured code that is fourth-order ac-
∂t ∂xj
∂P ∂ 2 ui
=− + μ 2 + Aρui , (46)
∂xi ∂x j Because the linear forcing scheme used here adds en-
are solved to describe the flow physics. A is the forc-
ing coefficient used by Rosales and Meneveau [28].
These equations are solved in the incompressible
limit, and a Poisson equation for pressure is therefore
used to enforce the conservation of mass equation. Pe-
riodic boundary conditions are applied to the velocity
field in all three coordinate directions.
The level set equation
∂G ∂G
+ uj = sL |∇G|
∂t ∂xj
∀ G = G0 , |∇G| = 1, (47)
is solved to describe front evolution. Since no curva-
ture terms appear in this equation, the simulation ef-
fectively describes front propagation in the corrugated
flamelets regime. A reinitialization procedure is per-
formed after every three time steps to force the level
set field variable away from the front to conform to a
distance function. Reinitialization is accomplished by
using an iterative marker method to estimate the dis-
tance to the front, and then solving a PDE in pseudo-
time to improve the accuracy of the estimate. The
third-order WENO scheme of Peng et al. [29] is used
for the PDE step.
information from only one isocontour of the level set
field variable (G = G0 ). This isocontour is used to
compute the area of the flame, as needed in the model.
The first-order δ function scheme of Smereka [30]
is used for the area computation. Neumann bound-
ary conditions are prescribed for the level set at each
end of the domain in the direction of propagation. The
front, however, is never allowed to come so near these
boundaries that their treatment influences behavior.
When the front approaches the edge of the domain
along the long axis, a data shift is performed which
moves the front and all the field variables about a
domain length in the direction opposite of front prop-
agation. Periodic boundary conditions are prescribed
for the level set in the other two directions.
curate in space and second-order accurate in time is
used to compute the flow. The code is run using an
explicit solver, and the CFL number is limited to 0.8.
ergy to the flow at all wavenumbers, the turbulent
flowfield is initialized within a 2563 cube and then
copied to an adjacent cube. This prevents the gener-
ation of wavenumbers smaller than the inverse of the
box size. Fig. 1 shows two instantaneous realizations
of the flowfield and front.
Fig. 2 shows the average turbulent kinetic energy
in the domain as a function of time. Initially, the ki-
netic energy rapidly decays due to dissipation. Af-
ter several eddy turnovers, however, the effects of
the forcing scheme become apparent and the tur-
bulent energy begins to increase. Once the forcing
scheme strikes a balance with the viscous dissipa-
tion, the energy in the domain stabilizes in time. In
general, the trends seen in Fig. 2 agree well with
the results reported by the developers of the forc-
ing scheme [28]. The level set is released after sta-
bilization has occurred, at approximately 27.5 eddy
turnover times.
Fig. 3 shows the energy spectrum of the DNS,
computed at the point in time of level set release.
Even though this DNS only captures a very limited
inertial range, the spectrum does indicate that turbu-
lent physics are present. Additionally, the dissipative
750 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
Fig. 1. Snapshots from a DNS of front propagation with
u /sL,u ≈ 1.0. The level set is the wrinkled surface, and the
cut plane shows vorticity magnitude. The top image shows
an early time in the simulation, and the bottom image shows
the field several eddy turnover times later. From this perspec-
tive, the front propagates from the upper left to the lower
right.
falloff occurs smoothly and monotonically, indicating
that the smallest scales of the flow are resolved.
It was argued in Section 3 that Γ c0 does not con-
tribute to resolved front propagation. Kollmann’s rela-
tion, Eq. (32), can be used to write this term in a form
that can be evaluated directly from the DNS data. Af-
ter manipulation, the computable expression for Γ c0
is

Γ c0 = F uj,c=c0 nj dA
S,c=c0

− uj,c=c0 F nj dA. (48)
S,c=c0
Fig. 4 shows Γ c0 as a function of time for various
filter widths. The values of Γ c0 that are shown in
this figure are evaluated at the location in the com-
putational domain where the instantaneous unfiltered
front intersects a particular vector. This particular
vector lies along the domain centerline that points
Fig. 2. Average kinetic energy, k = 12 uj uj , in the DNS
as a function of time. The initial turbulent kinetic energy
has largely dissipated by 3 eddy turnover times. The forcing
scheme then increases and eventually stabilizes k. The level
set is initialized and released at 27.5 eddy turnover times.
Fig. 3. Energy spectrum from the DNS. This spectrum was
computed at the point in time of level set initialization and
release. DNS data (◦ ◦ ◦), κ −5/3 scaling (—).
in the direction of mean front propagation. The lo-
cation of the intersection changes with time as the
front moves through the domain, and the evaluation
therefore tracks the front. This tracking scheme is
not Lagrangian because it does not track a particu-
lar section of the front, but it does show how Γ c0
evolves at the front over a long period of time. Occa-
sionally, this tracking scheme suffers at smaller filter
widths when turbulence convects sections of the front
laterally across the domain centerline, and therefore
ahead of other sections of the front. This data sam-
pling scheme then switches to providing information
about the new front section rather than the section be-
hind it. This causes a small discontinuity in the time
signal of Γ c0 , but does not affect the conclusions that
can be drawn from the data.
Fig. 4 demonstrates that Γ c0 can grow to val-
ues that are comparable in magnitude to the resolved
front convection term, u · ∇G. The quantity Γ c0 de-
scribes seemingly small subfilter effects, but it can at
times reach relatively large magnitudes because the
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
Fig. 4. Time evolution of Γ c0 (—) and u · ∇G (- - -), evaluated at a position that follows the unfiltered front. Note the different
vertical axis scales.
volumetrically filtered velocity that the resolved con-
vection term sees differs from the velocity that the
front itself sees. In spite of the fact that Γ c0 may
instantaneously be large and positive or large and neg-
ative, Fig. 4 shows that it always fluctuates around
zero, regardless of the filter resolution. This tendency
to return to zero supports the prior assumption that
Γ c0 ≈ 0. When considered instantaneously, this ap-
pears to be a poor assumption, but when considered
in the context of describing multiple realizations of
a subfilter phenomenon, it is appropriate. Additional
support for this assumption is provided by the fact
that Γ c0 decreases in magnitude as the filter width in-
creases. Physically, this occurs because a larger flame
area is considered and the positive and negative con-
tributions to Γ c0 are more likely to cancel with one
another. This behavior suggests that Γ c0 is often in-
significant relative to the other terms that influence
front dynamics. The propagation speed of the front,
for example, does not decrease at larger filter widths,
and neither would resolved convective terms if bulk
flows existed. Finally, Γ c0 seemingly bears a resem-
blance to the subfilter stress terms in the momentum
transport equations that are typically modeled using a
turbulent viscosity. The tendency of Γ c0 to decrease
as the filter width increases, however, stands in con-
trast to the behavior of the subfilter stress. Modeled
turbulent viscosity increases as the filter width in-
creases because this term is forced to describe more
and more of the turbulent transport. The opposite
751
trend is observed in Γ c0 , which again supports the
assumption that this term does not require modeling.
The procedure that is used to evaluate Γ c0 can
also be used to evaluate the convective terms that ap-
pear in the dynamic model (Eq. (30), for example). In
Fig. 5, these convective terms are plotted as a func-
tion of time. Just as in Fig. 4, the terms are evaluated
at locations where the instantaneous position of the
unfiltered front intersects the vector lying along the
domain centerline that points in the direction of mean
front propagation. This choice is again made because
the terms are only interesting near the front. Fig. 5
shows the convective terms at the front when a fil-
ter and a test-filter are applied either sequentially or
concurrently. The terms are plotted using two differ-
ent filter resolutions, and the figure also shows the
differences in these terms as a function of filter resolu-
tion. Fig. 5 demonstrates that the differences between
the concurrently and sequentially filtered convective
terms decrease as the filter width is increased. This is
the same behavior that is observed with Γ c0 . Further-
more, the physical argument that is used to claim that
Γ c0 is negligible is equally applicable to the convec-
tive terms. The discrepancies that occur between the
filtered and test-filtered velocities at the front location
may be viewed as instantaneous deviations from the
test-filtered velocity field. When considered as an ef-
fect on the front over an entire filter volume, these
deviations should not be capable of moving the fil-
tered front because they should tend to, in the limit of
a large number of realizations, cancel one another out.
752 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760

Fig. 5. Time evolution of u· ∇G and u · ∇ G, evaluated at a position that follows the unfiltered front. The first two plots show

u · ∇G (—) and u · ∇ G (- - -) at different filter resolutions. The third plot shows the difference of these terms at Δ/η = 8 (- - -)
and Δ/η = 32 (—). In each plot, the test-filter width Δ is Δ = 1.5Δ.

These arguments and data suggest that it may be pos-

sible to neglect the resolved convective terms in the
dynamic model.
The DNS can also be used to evaluate the model
for |∇G| that was proposed in Eq. (34). Fig. 6 shows
the time evolution of both the model and |∇G| at a va-
riety of filter resolutions. These quantities are evalu-
ated at the same front locations that are used in Figs. 4
and 5. While not perfect, the model in Eq. (34) is val-
idated by the DNS data. Fig. 6 shows that |∇G| is
accurately represented by AΔ /VΔ . The model real-
istically describes the magnitude of the exact term at
all filter resolutions, and additionally captures the ma-
jority of the fluctuations of the exact term, including Fig. 6. Time evolution of |∇G| (—), and of the model for
how these fluctuations scale with filter width. Instan- |∇G| proposed in Eq. (34), AΔ /VΔ (- - -). From top to bot-
taneous model errors certainly appear, but these errors tom, the filter resolutions used in the plot are Δ/η = 4, 8, 16,
and 32.
are not cumulative, and are not biased toward either
overprediction or underprediction.
Since a reasonable level of support is established come fully wrinkled by the turbulence. The PDFs in
for the model describing |∇G|, the dynamic proce- this figure are parameterized by filter width. The tur-
dure itself can now be considered. Up to this point, it bulent burning velocity associated with a given filter
has been implicitly assumed that a dynamic procedure width and a given point in space is calculated using
is necessary, or at the very least helpful, in predict- the definition of sT ,u in Eq. (15). Fig. 7 illustrates
ing how propagating surfaces behave. The validity of that sT ,u /sL,u is unimodal, but spread out. This sug-
this assumption can be checked using the DNS. In gests that, even at a constant filter width, a model is
Fig. 7, some probability distribution functions (PDFs) needed to accurately describe sT ,u /sL,u .
describing the turbulent burning velocity are plotted. Fig. 8 demonstrates that it is in fact a dynamic
These PDFs are generated using the full flow and model, rather than a static model, that is needed. In
level set fields from a single instant in time. The time Fig. 8, the turbulent propagation velocity is shown in
instant that is chosen occurs after the front has be- a scatter plot as a function of uΔ . Since the front prop-
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760 753

Fig. 7. PDFs of sT ,u /sL,u , evaluated using an instantaneous

data field from the DNS. In this data set, u /sL,u ≈ 2.5 and
the test-filter resolutions are, from the curve with the high-
est peak to the curve with the lowest, Δ/η = 4, 8, 16, 32,
and 64.

agation in this DNS occurs in the corrugated flamelets

regime, the model for the turbulent propagation speed
in Eq. (41) is dependent only on uΔ . The scatter plot,
however, shows that a wide range of turbulent prop-
agation speeds exist for any given uΔ . In this figure,
uΔ is computed by determining the amount of kinetic
energy that is removed by the filtering procedure, Fig. 8. Distributions describing the turbulent burning veloc-
ity when u /sL,u ≈ 2.5. Top: sT ,u /sL,u vs uΔ when the
  2 2   filter width is Δ/η = 32. Bottom: PDFs of the dynamic
uΔ = k − kΔ  , (49)
3 coefficient α in Eq. (36), parameterized by the test-filter res-
where k is the kinetic energy of the unfiltered veloc- olution. The filter widths are, from the curve with the high-
ity field and kΔ is the kinetic energy of the filtered est peak to the curve with the lowest, Δ/η = 4, 8, 16, 32,
and 64.
velocity field. Note that the averaging procedure dis-
cussed at the end of Section 3 is applied to the ki-
netic energies, and that the averaging occurs over a
filter width in each direction. Because sT ,u /sL,u does
not unconditionally and monotonically increase with
uΔ , the value of the coefficient α cannot be constant.
This result is verified by the second plot in Fig. 8,
which shows PDFs of the dynamic coefficient α at
various filter widths. Interestingly, the mean value of
α approaches zero as the filter width is decreased.
This implies that the turbulent propagation speed ap-
proaches the laminar propagation speed faster than
uΔ approaches zero. This result is consistent with the
idea of the Gibson scale [6], which states that ed-
dies with velocities smaller than sL,u cannot affect Fig. 9. Time evolution of sT ,u /sL,u at a particular location
the flame front. While the static turbulent burning ve- on the propagating front. The filters used are: Δ/η = 4 (—),
locity model in Eq. (41) does not capture this effect, 16 (- - -), and 32 (- · · -).
the dynamic model does.
The turbulent burning velocity itself is shown as quantity. When a large filter is used, some sections of
a function of time and filter width in Fig. 9. sT ,u is the filtered front might be found propagating at the
evaluated directly from its definition, Eq. (15), and laminar burning speed, while others might propagate
at the same time-dependent front locations used in at four times this speed.
Figs. 4 and 5. Fig. 9 shows that, as expected, the use Dynamic models are designed to capture these
of a larger filter tends to result in a larger turbulent kinds of fluctuations, but it is well known that dy-
propagation speed. This figure also emphasizes that namic approaches do not produce desirable results
the turbulent propagation speed is a strongly varying when used instantaneously and without averaging.
754 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
Fig. 10. Time evolution of the unaveraged dynamic model at a particular location on the propagating front. The data in the second
and third columns are calculated both directly from the DNS (—), and from the dynamic model (- - -).
Fig. 10 shows this to be true for the proposed dy-
namic propagation model. In this figure, the dynamic
model is applied to the DNS data, but averaging is
only used to help describe the turbulent velocity fluc-
tuations, uΔ . All other quantities are left unaveraged.
The difference between the concurrently and sequen-
tially filtered areas appearing in the dynamic equa-
tion is shown in the left part of Fig. 10. This differ-
ence should, in an average sense, be positive. Locally,
however, it oscillates between positive and negative
values. Furthermore, one of the filtered areas that
comprises this difference occasionally decreases in
magnitude and becomes relatively small compared to
the other. When this happens the value of the com-
puted dynamic coefficient, α, strays outside reason-
able bounding limits. This behavior is shown in the
middle set of plots in the figure. While the values of
α that are computed directly from the DNS data all
fall within an expected range, those computed from
the dynamic model do not. In an effort to compensate
for these deviations, α is artificially bounded between
−1 and 5 in the figure. In spite of this bounding of the
coefficient, the resulting dynamic turbulent propaga-
tion speed predictions shown on the right in Fig. 10
still do not agree with the exact propagation speeds
calculated from the DNS. Although this problem is,
as discussed above, present in all dynamic models,
the failure that occurs in the model proposed here is
worth discussing. The particular mechanism of fail-
ure in this case is influenced by the fact, shown in
Fig. 8, that the turbulent burning velocity does not in-
stantaneously correlate with the fluctuating turbulent
velocity magnitude, uΔ . It only correlates with this
quantity in a very averaged sense, and burning speed
models that employ this quantity can therefore only
be used dynamically in a similarly averaged fashion.
Finally, in further support of this point, even the val-
ues of α that are calculated from the DNS data are no
longer well bounded when the velocity fluctuations
used in their computation are not averaged.
In Fig. 11, the implementation of the dynamic
model that employs the full averaging procedure sug-
gested in Eq. (45) is tested and compared to the DNS
data. The model is applied in a postprocessing step us-
ing three sets of filters and test-filters, and the terms
in the dynamic equation are averaged over one quar-
ter of the front. The plots in the figure show the mean
front position, the mean front speed, and the dynamic
model coefficient as a function of time. Note that the
“mean front” denotes the completely filtered, or pla-
nar, realization of the front. When a filter smaller than
the domain width is used, the mean front speed is cal-
culated by multiplying the filtered front speed by the
ratio of the filtered and planar front areas, thus en-
suring a front consumption rate that is independent of
filter width. In each plot, the dynamic model is com-
pared with the static turbulent burning velocity model
from Eq. (41), the laminar burning velocity, and the
DNS data.
In the first combination of filters shown in Fig. 11,
the filter width corresponds to the mesh cell width,
Δ/η = 2, and the filtered data is therefore fully
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
Fig. 11. Modeled time evolution of the DNS front. The propagation is computed using the: DNS data (◦ ◦ ◦), laminar burning
speed sL,u (
), static sT ,u model (  ), dynamic sT ,u model with: Δ/η = 2 and Δ /η = 512 (—), Δ/η = 3 and Δ /η = 6
(- - -), Δ/η = 6 and Δ /η = 12 (- · · -).
resolved. The test-filter width, Δ /η = 512, corre-
sponds to the domain size, and the test-filtered data
is fully unresolved. This is admittedly a purely aca-
demic filtering scheme, but when used with this filter
combination, the dynamic model produces excellent
results, as expected. The model captures the average
magnitude of the propagation velocity, as well as the
instantaneous propagation velocity fluctuations. The
front position is calculated by integrating the mean
propagation speed, and is accurately, although not
perfectly, predicted. The accuracy of these results is
directly attributable to the dynamic model’s access to
the area of the fully resolved front. With these filters,
the left-hand side of Eq. (36) reproduces the exact
area per filter volume of the front, and the filtered
propagation speed, sT ,u , is simply the laminar burn-
ing velocity. Solving the equation therefore forces the
dynamic coefficient α to guarantee that the mean front
consumes exactly the same volume of space that the
fully resolved front does. While this result does not
prove anything about the model’s capabilities in re-
alistic LES settings, it does suggest that the use of
surface area as a dynamic scaling parameter is sound.
The accuracy of the model suffers mildly when the
second and third combinations of filters are used. In
the second case, the filter width is Δ/η = 3, and the
test-filter width is twice the filter width, Δ /η = 6. In
the third case, the filter width is Δ/η = 6, and the test-
filter width is again twice the filter width, Δ /η = 12.
These filters correspond to realistic LES filters and
755
produce results that, while imperfect, are more accu-
rate than the corresponding static model results. The
final front displacement is somewhat overpredicted,
but the dynamic model for the most part reasonably
reproduces the DNS data. It especially improves pre-
dictions, as compared to the static model, during the
first few eddy turnover times. In this early devel-
opment period, the flame front exists in a transient
regime and has not yet fully equilibrated with the tur-
bulence. The dynamic model efficiently captures this
behavior. The overprediction errors that do occur ap-
pear to depend strongly on how uΔ and α interact.
Specifically, when α is allowed to change in order to
conserve front volume consumption among different
filter levels, it can drift away from values that pro-
duce ideal propagation speed estimates. For example,
for the second and third filter combinations applied in
Fig. 11, the dynamic model does ensure that the test-
filtered front consumes just as much volume as the
filtered front. The value of α that guarantees this prop-
erty, however, produces a somewhat underestimated
front speed when it is returned to the burning velocity
model. Unfortunately, the burning velocity model that
is employed here is susceptible to this issue. This be-
havior occurs because, as discussed above, sT ,u does
not appear to locally correlate well with uΔ . Addi-
tionally, sT ,u strongly depends on the product of α
and uΔ , especially in the corrugated flamelets regime.
Therefore, while a dynamically induced change in the
value of α might be correctly describing a trend based
756 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
on a changing value of uΔ , it is possible that the
changing α value is incorrectly describing the magni-
tude of that trend. Put another way, α might become
either too large or too small when modeled veloc-
ity fluctuation estimates either over- or underpredict
the local influence of turbulence on the propagation
speed.
In general, these DNS results validate the basic
premise of the dynamic model. In regimes where
subfilter velocity fluctuations can be accurately esti-
mated, the application of a dynamic procedure can
improve front speed predictions. In certain cases, such
as when a flame’s brush thickness is undergoing a
transient phase, it can dramatically improve them.
While the procedure is thus justified, the particular
burning velocity model that has been applied presents
some challenges to the dynamic procedure. Specifi-
cally, when Eq. (41) is used in regions where velocity
fluctuations do not correlate well with the propagation
speed, the accuracy of the dynamic model suffers.
Similar issues regarding turbulent velocity fluctu-
ations have been noted by other researchers. In their
formulation, for example, Im et al. [16] observed and
commented on a sensitivity to the methods they used
to model these velocity fluctuations. Charlette et al.
[12] showed that in the limit where sT ,u scales en-
tirely with uΔ , the dynamic equation, Eq. (36), is
an ill posed problem for determining α. Charlette et
al. subsequently suggested a burning velocity model
based on a power law in an effort to deal with this
problem. While the issue of ill-posedness offers in-
sight into the results presented in this section, it
should not, by itself, rule out a continued considera-
tion of Eq. (41). Primarily, this is because the limiting
case where the problem becomes ill-posed seems to
occur in only a small subset of industrially relevant
flows. If, for example, the burning velocity model in
Eq. (41) is to become a function of only uΔ , two con-
ditions must hold. The first condition is that the func-
tional dependency on the Damköhler number must
disappear. If it did not, the burning velocity would still
scale with the filter width. This dependency only dis-
appears for Damköhler numbers above approximately
20 [6]. The second condition is that uΔ must be sig-
nificantly larger than sL,u . If it is not, then the laminar
burning velocity itself still plays an important role in
turbulent propagation. These two conditions lead to a
constraint on the ratio of the filter width and the flame
thickness. If they are to be satisfied only weakly, so
that Da = 20 and uΔ /sL,u = 5, then the equation for
the Damköhler number shows that Δ/f = 100. In
most industrially relevant LES computations, the ra-
tio of Δ and f is significantly smaller than 100,
and the ill-posed limit is not approached. The DNS
that is considered here, however, does seem to reach
these conditions when the filter width is large. This is
because the Damköhler number in the DNS is effec-
tively infinity.
Apparently, at least two issues regarding the dy-
namic model proposed here require further attention.
The first is the need for an improved method of de-
scribing propagation in the limit where sT ,u ∼ uΔ .
In this limit, all propagation occurs on the subfilter
scale, and the dynamic approach of examining how
information transfers between the resolved and subfil-
ter scales becomes irrelevant. The second is the need
for improved methods of describing subfilter veloc-
ity fluctuations, and of describing precisely how front
propagation depends on them. Addressing these is-
sues would improve the precision with which a dy-
namic burning velocity model could operate.
5. Application in LES
To demonstrate the dynamic model’s applicability
in LES, it is tested in a simulation of the F3 turbu-
lent premixed Bunsen flame studied experimentally
by Chen et al. [31]. The F3 flame is fed by a stoi-
chiometric (φ = 1.0) methane–air mixture that exits a
jet nozzle at a Reynolds number of ∼23,000. A pilot
flame of equivalent composition is used for stabiliza-
tion. The nozzle diameter of the Bunsen jet is 12 mm,
and the gas exit velocity is 30 m/s. In the context of
the premixed regime diagram, the F3 flame exists in
the thin reaction zone regime.
This flame is experimentally well-characterized,
and therefore, due to the scarcity of experimental tur-
bulent premixed data, has been the subject of a num-
ber of numerical studies [1,32,33]. The simulation
performed here closely follows prior work in which a
level set based premixed combustion model was ap-
plied in the context of LES [1]. More specifically,
the code used in the present work is a variant of the
code described in that earlier work. The differences
that do exist, such as the present use of Lagrangian
rather than Germano averaging in the subgrid viscos-
ity model [27], and the use of a parallel fast marching
method for level set reinitialization, do not qualita-
tively affect the computational approach.
A cylindrical coordinate mesh consisting of 323 ×
117 × 64 nodes in the axial, radial, and azimuthal di-
rections, respectively, is used for this LES. This ∼2.4
million node mesh is run in parallel on 32 processors.
The computation of a flow through time takes roughly
35 wall clock hours. The computational domain of the
simulation reaches from a half jet diameter upstream
of the nozzle to 30 jet diameters downstream of the
nozzle, and from the nozzle centerline to a distance
of 6 jet diameters in the radial direction.
The dynamic burning velocity model developed
above is used to describe the propagation speed of the
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
flame front. At each time step and at each mesh point,
a slightly modified version of the dynamic equation,
Eq. (44), in which the convective terms have been
neglected is solved. There are two reasons for neglect-
ing the dynamic convective terms in this LES. First,
it was argued in Section 4 that these terms should
not contribute to front propagation when considered
in a statistically averaged sense. Second, in this par-
ticular flame the jet velocity is significantly larger
than the turbulent burning velocity. Therefore, in sit-
uations where these terms are not considered on a
sufficiently averaged basis, they will not appropriately
cancel out and will, due to their magnitude, inappro-
priately dominate the dynamic equation. All curvature
terms, which are known to be important in the thin re-
action zone regime, are dynamically considered.
In practice, the process of solving the dynamic
equation is initiated by evaluating a Heaviside func-
tion everywhere in the domain using the value of
the local level set field variable. The introduction of
this Heaviside function allows the flame front to be
test-filtered in a mathematically consistent way, as de-
scribed earlier in this paper. The test-filtering of the
Heaviside function is performed using the standard
LES test-filter prescribed in the code. Similarly, the
local value of the kinetic energy is computed every-
where both before and after application of the test-
filter to the velocity field. These kinetic energy esti-
mates will be needed in evaluating the velocity fluc-
tuations required by the model. Next, again using the
scheme of Smereka [30], the areas of the filtered and
test-filtered fronts are computed, each on a per-filter-
volume basis. Finally, to ensure that a statistically
appropriate number of realizations are used in the
dynamic calculations, all of these quantities are av-
eraged in space. In this LES, averaging is performed
in the homogeneous (azimuthal) flow direction. Since
the burning velocity is only needed in the vicinity of
the flame front, this averaging procedure neglects all
mesh points that are not within a distance of several
cells from the actual flame front.
Estimates of the subfilter velocity fluctuations
needed in the burning velocity model are then made.
Velocity fluctuations at the filter level are estimated
by dividing the subfilter turbulent viscosity by the
mesh cell size, uΔ = νt,Δ /x. While it might seem
reasonable to then estimate the test-filtered velocity
fluctuations, uΔ , by computing a test-filtered tur-
bulent viscosity, this approach does not work well
in conjunction with Peters’ turbulent burning veloc-
ity model. Specifically, nothing guarantees that test-
filtered turbulent viscosities will produce test-filtered
velocity fluctuations that are larger than their filtered
counterparts. If this implementation of the dynamic
model were to encounter a situation where uΔ < uΔ ,
it would fail. To avoid this situation, the test-filtered
757
velocity fluctuations are computed by taking the dif-
ference of the computed filtered and test-filtered ki-
netic energies, backing out the associated fluctuation
magnitude, and adding it to the uΔ estimate,
  2   2 2  
uΔ = uΔ + kΔ  − kΔ  . (50)
3
This choice guarantees that the test-filter scale will
contain more energy than the filter scale. The dy-
namic coefficient α is finally solved for using a simple
Newton–Raphson technique.
Fig. 12 shows a single snapshot from the LES. The
local influence of the dynamic model can be seen in
the coloring of the flame front. This coloring shows
the ratio of the filtered and test-filtered flame areas,
on a per-test-filter-volume basis. In the flat, planar
regions of the flame, this ratio is nearly one and is in-
dicated by the blue portion of the color spectrum. In
regions of the flame where turbulence has produced a
large amount of wrinking, the ratio increases to 1.15.
The red end of the color spectrum represents these
larger values.
Two subsequent LES runs were performed for the
purposes of comparison. In the first of these, the tur-
bulent burning velocity is taken to simply be equiv-
alent to the laminar burning velocity. In the second,
the turbulent burning velocity is computed using the
full model from Eq. (41), but with the value of the
coefficient α left constant. In Fig. 13, instantaneous
data from these simulations is used to plot the turbu-
lent burning velocity for both the dynamic and static
model cases, as a function of the turbulent veloc-
ity fluctuation magnitude. This figure shows that the
dynamic model significantly impacts the computed
flame propagation speed. Specifically, the dynami-
cally computed flame speeds are approximately 50%
smaller than the statically computed flame speeds
over a wide range of uΔ . This suggests either that the
flame front itself is fairly well resolved in this LES,
or that the flame brush thickness does not often equi-
librate with the turbulence in this flame. The former
of these explanations seems to be most plausible.
Fig. 14 shows time averaged statistics from the
LES simulations. The axial velocity, temperature, and
water mass fraction plots each compare favorably,
although not perfectly, with experiments. The most
interesting feature of these statistics is their very
weak dependence on the turbulent propagation speed.
Specifically, the dynamic propagation model results
and the results computed using simply the laminar
burning velocity virtually overlap. While this behav-
ior is supported by Fig. 13, it would be expected
that the static model results, which predict 50% faster
burning rates, would be significantly different. Fig. 14
shows that this is not the case, and that the static
model results differ from the dynamic results by a few
758 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760

Fig. 12. Instantaneous snapshot from an LES of the F3 flame. The cut plane shows a contour plot of temperature, which ranges
= G0 isosurface, is colored according
from 300 K (black) to 2200 K (bright yellow). The premixed flame, represented by the G

to the ratio of AΔ /VΔ and AΔ /VΔ .

Fig. 13. Instantaneous scatter plot of the turbulent burning velocity as a function of the local turbulent velocity fluctuation
magnitude. Left: sT ,u is computed locally in the LES using a static value of α. Right: sT ,u is computed locally in the LES using
a dynamically determined value of α.

percent at most. These small differences can be seen
most clearly at the x/Do = 12.5 station. No experi-
mental data is available at this location, but the faster
flame speed predicted by the static model is evident
in the slightly higher temperatures and water mass
fractions, as compared to the dynamic model, that are
produced.
A comparison of the convection and flame propa-
gation time scales in this burner helps to explain these
time averaged results. A convection time scale can be
formed by considering how long it takes a particle
injected at the nozzle to propagate one burner diame-
ter downstream, tc = Do /Uo ∼ 4 × 10−4 s. A flame
propagation time scale can be formed by considering
how long it takes a laminar CH4 /air premixed flame to
propagate from the burner nozzle to the centerline of
the burner, tf = (Do /2)/sL,u ∼ 1.6 × 10−2 s. Com-
paring these time scales reveals tf = 40tc , which in
turn implies that the premixed front will be convected
far downstream before flame propagation can close it
off at the centerline. This time scale disparity necessi-
tates the use of a pilot flame in this burner. Addition-
ally, since this flame in reality closes off 12 to 15 di-
ameters downstream of the nozzle [31], this time scale
analysis supports the idea that turbulence, and not
propagation, plays the dominant role in determining
the flame height. As a result, the flame displays only a
weak dependency on the modeled propagation speed.
These results show that the proposed dynamic
burning velocity model can be effectively used in
large eddy simulations, and furthermore that the dy-
namic model predicts significantly different propaga-
tion speeds than the corresponding static model. Due
to the F3 burner’s lack of sensitivity to flame speed,
however, these results do not yet suggest that the dy-
namic model is superior to the static model. Further
testing is needed to demonstrate this in a comparative
sense.
6. Conclusions
In consistent LES procedures, subfilter models
should act to ensure that all mean predicted quantities
 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760 759

Fig. 14. Time averaged statistics from LES of the F3 flame. Left to right: axial velocity, temperature, and H2 O mass fraction
at sequential downstream locations. Experimentally measured values (◦ ◦ ◦) are plotted along with LES results computed using
propagation speeds of sL,u (—), a dynamically determined sT ,u (- - -), and a statically determined sT ,u (- · · -). The sL,u and
dynamic sT ,u results virtually overlap.

are, within certain ranges, independent of the filters Acknowledgments

being used. In this paper, a dynamic model for cal-
culating the propagation speed of flame fronts was
presented. This model was derived in a way that en-
sures this filter independence criteria is met. In the
derivation of the model, a new approach to writing
a filtered flame front equation was developed. This
approach was useful because it both worked in con-
junction with standard LES filtering techniques and
because it used information from only a 2-D sur-
face. Using this approach, the terms that describe
subfilter influences on the turbulent burning velocity
were explicitly determined. Furthermore, these terms
were used to derive a dynamic identity for the burn-
ing velocity. When enforced, this identity ensures
that evolving a flame front and then filtering the re-
sult yields the same answer that evolving a filtered
front does. A DNS was performed to validate the
proposed dynamic model. Results showed that the
model predicted the speed of a propagating turbu-
lent front with more accuracy than a static turbulent
burning velocity model. Applying the model in an
LES of a turbulent Bunsen flame produced results
that were in agreement with experimental measure-
ments.
Support from the United States Department of De-
fense NDSEG Program, the United States Air Force
Office of Scientific Research (AFOSR), and from
NASA is gratefully acknowledged.
References
[1] H. Pitsch, L. Duchamp De Lageneste, Proc. Combust.
Inst. 29 (2002) 2001–2008.
[2] J.P. Legier, T. Poinsot, D. Veynante, in: Proc. of the
CTR Summer Program, Stanford University, 2000, pp.
157–168.
[3] O. Colin, F. Ducros, D. Veynante, T. Poinsot, Phys. Flu-
ids 12 (7) (2000) 1843–1862.
[4] A.R. Kerstein, W.T. Ashurst, F.A. Williams, Phys. Rev.
A 37 (7) (1988) 2728–2731.
[5] H. Pitsch, Combust. Flame 143 (2005) 587–598.
[6] N. Peters, Turbulent Combustion, Cambridge Univ.
Press, Cambridge, UK, 2000.
[7] F.A. Williams, in: J.D. Buckmaster (Ed.), The Mathe-
matics of Combustion, SIAM, 1985, p. 97.
[8] H. Pitsch, Proc. Combust. Inst. 29 (2002) 1971–1978.
[9] H. Pitsch, H. Steiner, Phys. Fluids 12 (10) (2000) 2541–
2554.
760 E. Knudsen, H. Pitsch / Combustion and Flame 154 (2008) 740–760
[10] R.G. Abdel-Gayed, D. Bradley, Philos. Trans. R. Soc.
London A 301 (1457) (1981) 1–25.
[11] M. Herrmann, Ph.D. thesis, RWTH Aachen, 2000.
[12] F. Charlette, C. Meneveau, D. Veynante, Combust.
Flame 131 (2002) 159–180.
[13] F. Charlette, C. Meneveau, D. Veynante, Combust.
Flame 131 (2002) 181–197.
[14] R. Knikker, D. Veynante, C. Meneveau, Phys. Flu-
ids 16 (11) (2004) L91–L94.
[15] N. Chakraborty, R.S. Cant, Phys. Fluids 19 (10) (2007)
5101–5122.
[16] H.G. Im, T.S. Lund, J.H. Ferziger, Phys. Fluids 9 (12)
(1997) 3826–3833.
[17] A. Bourlioux, H.G. Im, J.H. Ferziger, in: Proc. of the
CTR Summer Program, Stanford University, 1996, pp.
137–148.
[18] M. Oberlack, H. Wenzel, N. Peters, Combust. Theory
Modelling 5 (2001) 363–383.
[19] T. Echekki, J.H. Chen, Combust. Flame 118 (1999)
308–311.
[20] N. Peters, J. Fluid Mech. 384 (1999) 107–132.
[21] M. Matalon, J. Matkowsky, J. Fluid Mech. 124 (1982)
239–259.
[22] V.G. Maz’ja, Sobolev Spaces, Springer-Verlag, Berlin,
1985.
[23] W. Kollmann, J.H. Chen, Proc. Combust. Inst. 25
(1994) 1091–1098.
[24] H. Wenzel, N. Peters, Combust. Sci. Technol. 158
(2000) 273–297.
[25] G. Damköhler, Z. Elektrochem. 46 (1941) 601–652.
[26] M. Germano, U. Piomelli, P. Moin, W.H. Cabot, Phys.
Fluids A 3 (7) (1991) 1760–1765.
[27] C. Meneveau, T.S. Lund, W.H. Cabot, J. Fluid
Mech. 319 (1996) 353–385.
[28] C. Rosales, C. Meneveau, Phys. Fluids 17 (9) (2005)
1–8.
[29] D. Peng, B. Merriman, S. Osher, H. Zhao, M. Kang, J.
Comput. Phys. 155 (2) (1999) 410–438.
[30] P. Smereka, J. Comput. Phys. 211 (1) (2006) 77–90.
[31] Y.-C. Chen, N. Peters, G.A. Schneemann, N. Wruck,
U. Renz, M.S. Mansour, Combust. Flame 107 (1996)
223–244.
[32] M. Herrmann, Combust. Flame 145 (2006) 357–375.
[33] R.P. Lindstedt, E.M. Vaos, Combust. Flame 145 (2006)
495–511.