Combustion Theory and Modelling

ISSN: 1364-7830 (Print) 1741-3559 (Online) Journal homepage: http://www.tandfonline.com/loi/tctm20

LES/PDF for premixed combustion in the DNS limit

Ranjith R. Tirunagari & Stephen B. Pope

To cite this article: Ranjith R. Tirunagari & Stephen B. Pope (2016) LES/PDF for premixed
combustion in the DNS limit, Combustion Theory and Modelling, 20:5, 834-865, DOI:
10.1080/13647830.2016.1188991

To link to this article: http://dx.doi.org/10.1080/13647830.2016.1188991

Published online: 16 Jun 2016.

Submit your article to this journal

Article views: 183

View related articles

View Crossmark data

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tctm20

Download by: [Ranjith Tirunagari] Date: 17 December 2016, At: 21:05

Combustion Theory and Modelling, 2016
Vol. 20, No. 5, 834–865, http://dx.doi.org/10.1080/13647830.2016.1188991

LES/PDF for premixed combustion in the DNS limit

Ranjith R. Tirunagari and Stephen B. Pope∗

Sibley School of Mechanical and Aerospace Engineering, Cornell University, Ithaca, USA
(Received 26 September 2015; accepted 5 May 2016)

We investigated the behaviour of the composition probability density function (PDF)

model equations used in a large-eddy simulation (LES) of turbulent combustion in the
direct numerical simulation (DNS) limit; that is, in the limit of the LES resolution length
scale  (and the numerical mesh spacing h) being small compared to the smallest flow
length scale, so that the resolution is sufficient to perform a DNS. The correct behaviour
of a PDF model in the DNS limit is that the resolved composition fields satisfy the DNS
equations, and there are no residual fluctuations (i.e. the PDF is everywhere a delta
function). In the DNS limit, the treatment of molecular diffusion in the PDF equations
is crucial, and both the ‘random-walk’ and ‘mean-drift’ models for molecular diffusion
are investigated. Two test cases are considered, both of premixed laminar flames (of
thickness δ L ). We examine the solutions of the model PDF equations for these test
cases as functions of /δ L and h/δ L . Each of the two PDF models has advantages and
disadvantages. The mean-drift model behaves correctly in the DNS limit, but it is more
difficult to implement and computationally more expensive. The random-walk model
does not have the correct behaviour in the DNS limit in that it produces non-zero
residual fluctuations. However, if the specified mixing rate  normalised by the reaction
timescale τ c is sufficiently large (τ c  1), then the residual fluctuations are less than
10% and the observed flame speed and thickness are close to their laminar values. Away
from the DNS limit (i.e. h/δ L  1), the observed flame thickness scales with the mesh
spacing h, and the flame speed scales with h. For this case it is possible to construct
a non-general specification of the mixing rate  such that the flame speed matches the
laminar flame speed.
Keywords: PDF methods; large-eddy simulation; turbulent combustion; DNS limit;
premixed flames

Nomenclature
Roman
CD mixing rate constant, Equation (36)
CL mixing rate constant, Equation (44)
Cm mixing rate constant, Equation (10)
D nozzle diameter
D molecular diffusivity
De effective diffusivity, Equation (7)
Dn numerical diffusivity, Equation (38)
Dr residual diffusivity

∗
Corresponding author. Email: [email protected]


C 2016 Informa UK Limited, trading as Taylor & Francis Group
 Combustion Theory and Modelling 835

f(ψ; x, t) density-weighted PDF of composition

h mesh spacing
N number of particles
S reaction source term
sL laminar flame speed
T temperature
t time
U fluid velocity
uF flame speed
v specific volume
W(t) isotropic Wiener process
X∗ (t) particle position
x position
YOH mass fraction of OH
z coordinate along the axis of the opposed nozzles

Greek
 LES resolution length scale
δ smallest flow length scale
δF flame thickness, Equation (34)
δL laminar flame thickness
ρ fluid density
σ φ, max maximum residual standard deviation
τc reaction timescale
φ(x, t) composition
ψ sample-space variables corresponding to φ
 mixing rate, Equation (10)

Superscripts
φ ∗ particle composition

φ resolved composition
∗

φ resolved composition at particle position

Subscripts
Tb burnt temperature
Tu unburnt temperature

Abbreviations
LES large-eddy simulation
PDF probability density function
DNS direct numerical simulation
836 R.R. Tirunagari and S.B. Pope

1. Introduction
Probability density function (PDF) methods [1] are seeing increased use in conjunction
with large-eddy simulation (LES) as an effective means of accounting for the turbulence–
chemistry interactions occurring in turbulent combustion [2,3]. While most previous ap-
plications of LES/PDF have been to non-premixed combustion [4–6], recently there have
been several applications to premixed turbulent combustion [7–10].
In LESs [11–13], modelled equations are solved for the larger-scale resolved motions,
while the effects of the smaller-scale (unresolved) residual motions are modelled. A speci-
fied LES resolution length scale, , demarcates the resolved and residual scales. The most
common approach to LES is to define the resolved fields by a filtering operation, with 
being the characteristic width of the filter [11,14]. However, as discussed below, for the
present considerations, it is preferable to define the resolved fields in terms of conditional
means [13,15].
In practice, the resolution length scale  is usually linked to the mesh spacing h used
in the numerical solution of the LES equations, e.g.  = h or  = 2h, which leads to a
mingling of modelling and numerical errors [11,16]. However, it is important to retain the
distinction between these two quantities, in particular so that numerically-accurate solutions
can be considered for h/  1.
An obviously-desirable property of an LES formulation is that it correctly converges
to the direct numerical simulation (DNS) limit. That is, as the resolution scale  becomes
small compared to the smallest flow length scale δ (so that the residual motions tend to
zero), then the LES equations should yield the same solutions as those obtained from the
fundamental conservation laws (e.g. the Navier–Stokes equations) that apply in a DNS. The
purpose of this paper is to examine the behaviour of the composition PDF models in this
DNS limit.
In a typical application of LES to a high-Reynolds-number flow, /δ is large, and
molecular transport has a negligible direct effect on the resolved fields: its primary effects
are the dissipation of kinetic energy and the mixing of species below the resolved scale. For
this reason, in LES, molecular transport is often neglected or treated in an unrealistically
simplified way [17]. But in the DNS limit, molecular transport is a dominant process, and
hence must be implemented correctly if the DNS limit is to be attained.
We now provide three reasons why it is important for LES/PDF models to converge
correctly to the DNS limit.
First, in laboratory flames, the molecular diffusivity can be much larger than one might
expect, given the typical flow Reynolds numbers of order 10,000. This is primarily because
the molecular diffusivity D (of dimensions length squared divided by time) increases
strongly with temperature T (e.g. as T1.7 ), and hence at flame temperatures can be 30 times
its value at room temperature. Kemenov and Pope [17] show that, in a typical LES of
the Barlow and Frank flames [18], the molecular diffusivity is generally several times the
residual diffusivity for temperatures above 1000 K.
Second, there is recent interest in premixed combustion in turbulent counter-flow burn-
ers [9,10,19]. When LES is applied to these flames, it is typically found that the resolution
scale  is of the same order of magnitude as the laminar flame thickness, δ L . Hence, the
direct effects of molecular transport are important, as is the convergence to the DNS limit.
Third, as computer power continues to increase, it is becoming practicable to perform
high-fidelity LES for a larger class of flows. Whereas, in conventional LES, the resolution
scale  is chosen to resolve only the larger, energy-containing motions, in high-fidelity
LES,  is chosen to be much smaller to provide partial resolution of the smallest-scale
 Combustion Theory and Modelling 837

processes [20]. For many occurrences of turbulent combustion, in which the rate-controlling
processes are at the smallest scales, high-fidelity LES may be required in order to provide
reliable simulations [3]. By definition, high-fidelity LES approaches the DNS limit, and
hence for this purpose it is essential that the LES/PDF models used converge correctly in
this limit.
The effects of molecular diffusion on the joint PDF of compositions can be decomposed
into: spatial transport, which affects the mean compositions; and mixing, which decreases
the composition variances, while not affecting the means. These two effects are modelled
separately. Here it is sufficient to consider the simplest mixing model, namely the ‘inter-
action by exchange with the mean’ (IEM) model [21], which (like other models) causes
the mixing to occur at a specified rate . Molecular transport can be implemented in two
ways, and we consider both, with details given in Section 2. These are referred to as the
Random-Walk (RW) model and the Mean-Drift (MD) model.
In the DNS limit, the residual diffusivity Dr vanishes in comparison to the molecular
diffusivity D. However, the mixing rate  does not vanish, but remains as an important
parameter.
We investigate the DNS limit by applying the LES/PDF model equations to two
test cases, both of premixed laminar flames. The first is an idealised, unstrained, one-
dimensional, freely-propagating, premixed laminar flame, with one-step chemistry. For this
case, the LES/PDF equations are solved using a simple 1D MATLAB code. The second
R

case is a strained methane/air premixed laminar flame in an opposed jet configuration, in

which a cold methane/air jet is opposed to a jet of hot combustion products. In this case
the LES/PDF equations are solved using the same 3D code used in many previous applica-
tions of LES/PDF to turbulent flames, (including turbulent premixed flames in the present
counter-flow configuration [10]) and the chemistry is described by a 16-species reduced
mechanism.
The LES/PDF equations are solved by particle-mesh methods, using N particles, mesh
spacing h, and time steps t. Except where mentioned explicitly otherwise, the number of
particles can be considered to be sufficiently large, and the time step sufficiently small, so
that the only significant numerical errors arise from the mesh spacing, h.
The focus of the investigation is on the solution of the LES/PDF equations as a function
of /δ L , i.e. the ratio of the resolution length scale to the laminar flame thickness: the DNS
limit corresponds to this ratio tending to zero. We consider both the numerically-accurate
solutions (corresponding to small h/), and also numerical solutions with h/ of order
unity.
The remainder of the paper is organised as follows. In Section 2 we describe the
two LES/PDF models considered, namely the random-walk model and the mean-drift
model. In Section 3 we consider the correct behaviour in the DNS limit, and the behaviour
of the LES/PDF equations corresponding to the MD and RW models. In Section 4 we
describe the first of the two test problems considered, namely an idealised, unstrained, one-
dimensional, freely-propagating laminar, premixed flame. Numerically-accurate solutions
to the LES/PDF equations in the DNS limit are presented and discussed. These solutions
are obtained using a mesh-free particle method that is described in Appendix A. Then, we
present numerical solutions obtained using a conventional cloud-in-cell (CIC) particle/mesh
method, which is described in Appendix B. In this case, the solutions depend on the specified
normalised mesh size h/. In Section 5 we describe the second test case of a strained
methane/air opposed-jet, laminar, premixed flame. The numerically-accurate solution of
the laminar-flow equations for this case is obtained using CHEMKIN-PRO [22]. The
numerical solutions to the LES/PDF equations in the DNS limit are obtained using the
838 R.R. Tirunagari and S.B. Pope

same 3D particle/mesh method used in many previous LES/PDF studies (e.g. [6,9]). Based
on the results obtained and other considerations, in Section 6 we discuss the relative merits
of the random-walk and mean-drift models. Conclusions are drawn in Section 7.

2. Composition PDF methods

2.1. Governing equations
We consider the standard case of a low-Mach-number, single-phase, turbulent reacting
flow. As functions of position, x, and time, t, the flow is described by the fields of velocity,
U(x, t), and composition, φ(x, t). In general, the nφ components of the composition vector
φ can be taken to be the mass fractions of the ns chemical species and the enthalpy. For
the low-Mach-number flow considered, as far as the thermochemistry is concerned, the
pressure can be approximated as being constant and uniform, and is not shown explicitly
in the notation. Thus, the equations of state determining the density ρ and temperature T
are of the form

ρ(x, t) = ρ̂(φ(x, t)), (1)

and

T (x, t) = T̂ (φ(x, t)). (2)

The velocity field is governed by the usual equations for the conservation of mass and
momentum. The focus here is on the composition field, which we take to be governed by
the equation

Dφ ∂φ
ρ =ρ + ρU · ∇φ = ∇ · (ρD∇φ) + ρS, (3)
Dt ∂t

where S is the reaction source term (i.e. the rate of change of φ due to chemical reactions),
and D is the molecular diffusivity. Both S and D are known functions of φ. The form of
the diffusion term embodies the assumption of unity Lewis numbers, which is discussed
further below.

2.2. LES/PDF formulation

In an LES/PDF simulation of the flow considered, the primary quantities involved are the
resolved density, ρ(x, t), the density-weighted resolved velocity, U(x, t), and the density-
weighted PDF, f(ψ; x, t), where ψ are the sample-space variables corresponding to φ.
While a decomposition of the composition field into resolved and residual components is
not needed, it is nevertheless useful to consider the (density-weighted) resolved composition
field defined as


φ(x, t) ≡ f(ψ; x, t)ψ dψ, (4)

where, here and below, the integration is over the whole of the composition space. The PDF
then represents the residual fluctuations about this mean.
 Combustion Theory and Modelling 839

The most common way to define resolved fields is by filtering (see for example [11,14])
and then the PDF is, more correctly, the filtered density function (FDF) [23,24]. The
less common alternative, and that considered here, is to define f(ψ; x, t) as the PDF of
composition, conditional on the resolved velocity field [13,15]. For most purposes the
difference between the two approaches is more conceptual than practical, with the same
models being used regardless of the approach taken. However, when applying the LES-
PDF equations in the DNS limit or to laminar flows, there is a major difference: with the
conditioning approach there are no residual fluctuations φ about the resolved mean  φ;
whereas, with the filtering approach, there are non-zero residual fluctuations, which scale
with . Clearly, in considering the DNS limit, it is far preferable to adopt the approach that
has the simple limiting behaviour φ = 0 rather than that in which the limiting behaviour
depends non-trivially on the non-physical parameter .
For some purposes, it is more convenient to use the specific volume v ≡ 1/ρ rather than
the density. For example, the resolved density is obtained from the PDF by

ρ(x, t)−1 = 
v (x, t) = f(ψ; x, t)v̂(ψ) dψ, (5)

where v̂ ≡ 1/ρ̂ is the equation of state for specific volume.

2.3. Exact PDF equation

The exact evolution equation for the PDF f(ψ; x, t) is

∂ f
ρ U · ∇ f = −∇ · (ρfU − 
+ ρ U | ψ)
∂t
∂  
− ρf[v∇ · (ρD∇φα ) | ψ + Sα (ψ)] + R, (6)
∂ψα

where  · | ψ denotes the mean conditional on the composition and on the resolved velocity
field. (This is the variable-density version of Equation (C10) of [13].) The term R is discussed
in [13]: suffice it to say that it is generally taken to be zero, and it certainly is zero for the
laminar test cases considered here.

2.4. Lagrangian particle method

In PDF methods, the modelling and numerical solution are usually performed using a
Lagrangian particle method [1,2,11]. The position of the general particle is denoted by
X∗ (t), and its composition by φ ∗ (t). The modelling is performed by specifying the evolution
of these particle properties; and the PDF considered, f(ψ; x, t), is the density-weighted
PDF of φ ∗ (t) conditional on X∗ (t) = x.

2.5. Random-walk model

In the random-walk (RW) model, the particle is specified to move with the resolved velocity,
plus a random walk, the magnitude of which is determined by the mean effective diffusivity
(which has dimensions of length squared divided by time). The molecular diffusivity is
D, and part of the LES modelling is to define a residual diffusivity, Dr , which accounts
for transport due to the residual motions. The effective diffusivity is then defined as their
840 R.R. Tirunagari and S.B. Pope

sum:

De ≡ D + Dr . (7)

According to the RW model, the particle position evolves by the stochastic differential
equation (SDE)

dX∗ = [
U +
v ∇(ρD ∗e )1/2 dW,
e )]∗ dt + (2D (8)

where W(t) is an isotropic Wiener process. Note that De∗ denotes the effective diffusiv-
ity based on the particle composition (and Dr ); whereas D e∗ denotes the mean effective
e (X∗ (t), t).
diffusivity evaluated at the particle location, i.e. it is a shorthand notation for D
The evolution equation for the particle composition is

dφ ∗ ∗
= −∗ (φ ∗ − 
φ ) + S(φ ∗ ). (9)
dt

On the right-hand side, the first term is the IEM mixing model, and the second term is the
reaction source term. The standard model for the mixing rate, , is

e
D
 = Cm , (10)
2

where Cm is a model constant.

(Note that, in Equations 8 and 10, the diffusion coefficient and the mixing rate are based
on the mean effective diffusivity, D ∗e . One can consider three additional model variants in
which one or both of these quantities is instead based on the particle effective diffusivity,
De∗ . Tests – for the case described in Section 4 – show that there are no qualitative differences
between the four variants, and only modest quantitative differences.)
The PDF equation deduced from these particle equations is

∂ f  
e ∇ f) − ∂ ρf −(ψα − φ

ρ U · ∇ f = ∇ · (ρD
+ ρ α ) + Sα (ψ) , (11)
∂t ∂ψα

where the summation convention applies to the repeated composition suffix α. Henceforth
we refer to this as the RW-PDF equation.
Multiplying this PDF equation by ψ and integrating over the composition space, we
obtain the implied conservation equation for the resolved composition:

∂
φ
ρ U · ∇
+ ρ e ∇
φ = ∇ · (ρD φ) + ρ
S. (12)
∂t

Note that the random walk (and in particular the drift term in Equation (8)) is constructed
to yield the diffusion term in Equation (12); and, by construction, the IEM model does not
(directly) affect the resolved composition φ, and hence the mixing rate  does not appear
in Equation (12).
 Combustion Theory and Modelling 841

2.6. Mean-drift model

In the random-walk model, molecular transport is implemented via the random walk in the
SDE for the particle position, X∗ (t). In contrast, in the mean-drift (MD) model [25,26],
molecular transport is implemented in the ODE for composition. The particle evolution
equations for the MD model are

dX∗ = [
U +
v ∇(ρD ∗r )1/2 dW,
r )]∗ dt + (2D (13)

dφ ∗   ∗
= [
v ∇ · (ρD∇ φ)]∗ − ∗ (φ ∗ − 
φ ) + S(φ ∗ ). (14)
dt

Thus, in the SDE for position (Equation 13) the random walk is based on D r , and accounts
solely for the transport due to the residual motions; and in the particle composition equation
(Equation 14), the first term of the right-hand side accounts for transport by molecular
diffusion.
For the mean-drift model, the corresponding PDF equation (deduced from Equations 13
and 14) is

∂ f  
r ∇ f) − ∂ ρf   φ
ρ U · ∇ f = ∇ · (ρD
+ ρ v ∇ · (ρD∇ α )
∂t ∂ψα

−(ψα − φ α ) + Sα (ψ) . (15)

Henceforth we refer to this as the MD-PDF equation. The corresponding equation for the
resolved composition is again Equation (12), i.e. it is identical to that for the random-walk
model.

3. The PDF methods in the DNS limit

3.1. The DNS limit
The general idea of the DNS limit of an LES is simple and obvious: as the resolution length
scale becomes small (/δ  1), it is highly desirable that the numerically-accurate solution
to the LES model equations tends to the solution of the DNS equations. For the general case
it is non-trivial to make this idea precise by quantifying the departure of the LES solution
from the DNS solution as a function of /δ.
The matter is more tractable for cases in which the flow is laminar (as in the test cases
considered here), or in which /δ is sufficiently small to resolve the velocity field fully. In
these cases there is no randomness in the evolution of the composition field, and from the
definitions of the resolved fields we therefore have simply


U = U, ρ = ρ, 
φ = φ,  = D,
D (16)

i.e. in the DNS limit the resolved quantities are equal to the instantaneous quantities. There
are no composition fluctuations, so the PDF is the delta function:

f(ψ; x, t) = δ(φ(x, t) − ψ). (17)

842 R.R. Tirunagari and S.B. Pope

We observe that, in the DNS limit, the exact PDF equation (Equation 6) reduces to

∂ f ∂    φ

ρ U · ∇ f = −
+ ρ f ∇ · (ρD∇ α ) + ρSα (ψ) . (18)
∂t ∂ψα

Henceforth we refer to this as the DNS-PDF equation.

The question now to address is whether the LES-PDF equations (Equations 11 and 15)
yield solutions consistent with the DNS limit.
Considering the resolved composition  φ(x, t) obtained from an LES, we define the
difference between the LES and the DNS limit by

φ () = ||
φ(x, t) − φ(x, t)||, (19)

where || · || denotes a suitable norm (e.g. the maximum absolute difference over all composi-
tions, x and t). We say that the LES is consistent with the DNS provided that this difference
tends to zero, i.e.

lim φ () = 0. (20)

/δ→0

3.2. The approach to the limit

For an LES model, more than consistency is desirable: it is desirable also that the model
have the correct approach to the DNS limit. Below, we encounter models for which the
difference φ varies quadratically as
 2

φ () ∼ . (21)
δ

This is an incorrect approach to the DNS limit, since (for the cases being considered) 
φ(x, t)
is independent of /δ (provided that the velocity field is fully resolved).
The correct approach to the DNS limit is an important consideration in LES, since
resolution is at a premium. Typically, the computational work W required to solve the
LES equations varies as W ∼ (/δ)−4 . Consequently, if φ varies quadratically with /δ
(Equation 21), then this error decreases slowly with increasing computational work as
W−1/2 .
We say that a model behaves correctly in the DNS limit if it is consistent and it
approaches the limit correctly.

3.3. Numerical solutions and the DNS limit

The preceding discussion pertains to the accurate solutions of the LES equations, whereas
in practice the solutions are obtained numerically and hence contain a numerical error
dependent upon the mesh spacing h. Consistency then requires that the numerical solutions
of the LES equations converge to the DNS solutions as /δ → 0 and h/δ → 0. A remaining
questions is: For consistency, is there a requirement on the ratio h/ as both h and  tend
to zero? For the LES-PDF equations applied to laminar flames, this question is addressed
in Section 4.6, with a possibly surprising answer.
 Combustion Theory and Modelling 843

3.4. Residual covariance and the DNS limit

The residual covariance is defined as

φ
α φβ ≡ f(ψ) (ψα − φ
α )(ψα − φ
β ) dψ. (22)

With the viewpoint taken here, that f is the PDF of the composition conditional on the
resolved velocity field, in the DNS limit, this PDF is a delta function, Equation (17).
Consequently, the residual covariance is zero.
Note that if the alternative, filtering, viewpoint were taken, then, in the DNS limit, the
residual covariance would instead be non-zero. With the usual definition of the filter width,
to leading order in , the residual covariance is

2
φ
α φβ =
α · ∇ φ
∇φ β (23)
12

(see Equation 13.157 of [11]). For the present purposes, it is clearly preferable to adopt the
conditional PDF viewpoint so that the residual covariance is zero in the DNS limit.

3.5. The mean-drift model in the DNS limit

We now examine the behaviour of the mean-drift model in the DNS limit by comparing the
MD-PDF equation (Equation 15) with the DNS-PDF equation (Equation 18).
The first term on the right-hand side of the MD-PDF equation is absent from the
DNS-PDF equation. This term represents turbulent transport by the residual motions, and
is zero approaching the DNS limit (since U =  U, Equation 17). For the laminar cases
r to
considered below, it is certainly correct, therefore, to take the residual diffusivity D
be zero. However, we note that, if instead the residual diffusivity were obtained from the
standard Smagorinsky model and a specified turbulent Prandtl number, then this would lead
r varying as (/δ)2 , and hence would result in incorrect behaviour in the DNS limit.
to D
r taken to be zero, the MD-PDF equation can be written
With D

∂ f ∂    φ

ρ U · ∇ f = −
+ ρ f ∇ · (ρD∇ α ) + ρSα (ψ) − ρ(φα − φ
α ) .
∂t ∂ψα
(24)

This is identical to the DNS-PDF equation except for the addition of the term in , which
arises from the IEM mixing model. This term causes compositions φ to relax to the mean φ
at the rate . However, approaching the DNS limit, there are no departures from the mean
(i.e. φ = φ, Equation 16) and hence the term is zero. Thus, in the approach to the DNS
limit, the LES-PDF-MD equation admits the same solution as the DNS-PDF equation, and
hence we conclude that the mean drift model is consistent and has the correct approach to
the DNS limit, i.e. it behaves correctly.
This important conclusion notwithstanding, we note that, approaching the DNS limit,
the relaxation rate is


D  
Cm D −2
 = Cm 2 = 2 , (25)
 δ δ
844 R.R. Tirunagari and S.B. Pope

and hence tends to infinity. Thus, while the delta-function PDFs given as solutions to the
DNS-PDF and MD-PDF equations are identical, the equations themselves are not. For this
reason, our definitions of consistency and the correct approach are based on solutions, not
on equations.

3.6. The random-walk model in the DNS limit

Compared to the DNS-PDF equation (Equation 18), the differences in the RW-PDF equation
(Equation 11) in the DNS limit are the addition of the term in  stemming from the IEM
model, and that the effects of molecular diffusion appear as a diffusion of the PDF (the first
term on the right-hand side of Equation 11, with D e = D  in the DNS limit) instead of as a
drift in composition space. The effects of these differences are best revealed by examining
the evolution equation for the residual covariance.
Looking first at the mean-drift model, the evolution equation for the residual covariance
(deduced from Equation 15) is

∂ 
ρ φ φ + ρU · ∇ φ  
α φβ = ∇ · (ρD r ∇ φα φβ )
∂t α β
+ 2ρDr ∇ φ β − 2ρ φ
α · ∇ φ α φβ

+ ρφ 
α Sβ + ρφβ Sα . (26)

In the DNS limit, the residual covariance and Dr are zero, and hence the right-hand side of
this equation is zero. Thus there is no mechanism causing the residual covariance to depart
from zero. This again shows that the mean-drift model approaches the DNS limit correctly,
with the residual covariance being zero (independent of /δ).
For the random-walk model, the residual covariance equation is almost the same, but
with the effective diffusivity D r in the first two
e in place of the residual diffusivity D
terms on the right-hand side. But this difference is crucial, because, in the DNS limit, the
second term on the right-hand side is then non-zero, and yields the spurious production
term 2ρD∇ φ β . This spurious production causes the residual variances to depart
α · ∇ φ
from zero (at locations where there are non-zero composition gradients).
It may be expected, and is confirmed by the results presented below, the dominant bal-
ance in the residual-covariance equation is between the spurious production and relaxation
term in . A balance between these two terms yields


D 2
φ
α φβ =
α · ∇ φ
∇φ β = α · ∇ φ
∇φ β . (27)
 Cm

This result shows that the random-walk model has an incorrect approach to the DNS limit,
with residual fluctuations scaling with . It is however consistent (as the residual covariance
tends to zero, and the mean evolves by the correct equation).
It is interesting to observe that, with the random-walk model and the specification Cm =
12, the residual covariance given by Equation (27) is the same as that given by Equation (23).
(We do not ascribe any significance to this observation.)
 Combustion Theory and Modelling 845

3.7. Summary of the DNS limit

To summarise the conclusion from this section: as the DNS limit is approached there are
no residual fluctuations, the residual covariance is zero, and the PDF is a delta function,
independent of /δ (provided that it is sufficiently small to resolve the velocity field).
Both the mean-drift and random-walk models are consistent with this limit. The mean-drift
model approaches this limit correctly (e.g. with zero residual covariance); whereas, because
of the spurious production of residual fluctuations, the random-walk model has an incorrect
approach, with the residual covariance varying as (/δ)2 .

4. Idealised, unstrained, plane, premixed laminar flame

4.1. Definition of the test case
The first test case we consider is that of an idealised, unstrained, one-dimensional, plane,
premixed laminar flame. The single direction of variation is denoted by x. The chemistry
is described by a single composition variable, φ(x, t), which may be considered to be a
normalised product mass fraction, or a reaction progress variable. At x = −∞ there are
pure reactants (φ = 0), and at x = ∞ there are completely burnt products (φ = 1). In a
frame fixed with the unburnt reactants, the flame propagates to the left at the flame speed
uF . Alternatively, in the frame fixed to the flame, the reactants flow to the right at speed uF .
The very simple thermochemistry is defined in terms of φ as follows. The temperature
T is specified as

T = Tu + φ(Tb − Tu ), (28)

where Tu and Tb are the temperatures of the unburnt and burnt mixture. For definiteness, we
take Tu = 300 K and Tb = 2100 K, although it is only the ratio Tb /Tu = 7 that is significant.
Consistent with the ideal gas law, the density and specific volume are then specified as

Tu T
ρ(T ) = ρu , v(T ) = vu , (29)
T Tu

where ρ u = 1/v u are the unburnt values.

The molecular diffusivity is specified as

 1.72
T
D(T ) = Du , (30)
Tu

which is based on an accurate approximation for methane combustion [17], with Du being
the unburnt value.
For each quantity introduced, the value in the burnt stream is denoted by ρ b , v b and Db .
The density and specific volume ratios are 7 (like the temperature), while the diffusivity
ratio is Db /Du ≈ 28.
The chemical source term is specified to be
⎧
⎨0 for φ < φR
S(φ) = 1 4 (31)
⎩ z(1 − z)2 for φ ≥ φR ,
τc 27
846 R.R. Tirunagari and S.B. Pope

0.8

0.6

S(φ)τc
0.4

0.2

0 0.2 0.4 0.6 0.8 1

φ
Figure 1. Normalised reaction source term, Equation (31). (Colour online).

where τ c is the chemical timescale, φ R is taken to be 0.4, and z is defined by

1−φ
z≡ . (32)
1 − φR

The resulting normalised function S(φ)τ c is shown in Figure 1. The factor of 4/27 in
Equation (31) is chosen to make the maximum value of S(φ)τ c unity.
For this simple 1D case, it is not necessary to consider the momentum equation, as the
velocity is determined by mass conservation. Specifically, in the frame fixed with the flame,
the mass flux is uniform, so that the velocity U is obtained from the relation

ρ U = ρu uF . (33)

We arbitrarily set ρ u , Du and τ c to unity, and so all quantities obtained can be considered
to be normalised by them.

4.2. Solutions considered

For this test case, we consider the following five solutions.

(1) The numerically-accurate solution of the governing equation (Equation 3), i.e. the
laminar-flame solution; or, equivalently, the DNS solution.
(2) The numerically-accurate solution of the LES/PDF random-walk equations in the
DNS limit.
(3) The numerically-accurate solution of the LES/PDF mean-drift equations in the
DNS limit.
(4) The numerical solution of the LES/PDF random-walk equations in the DNS limit.
(5) The numerical solution of the LES/PDF mean-drift equations in the DNS limit.
 Combustion Theory and Modelling 847

0.9

0.8

0.7

0.6

φ 0.5

0.4

0.3

0.2

0.1

0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x/δL

Figure 2. Laminar flame profile. The symbols are the points on the profile where φ = [1/4 3/4], and
the dashed line through them is the secant used to define the flame thickness. (Colour online).

4.3. DNS solution

The laminar-flame or DNS solution is obtained by solving Equation (3) for the case con-
sidered by an accurate finite-difference method. The resulting composition profile is shown
in Figure 2. In general, the flame speed is denoted by uF and the flame thickness (defined
below) is denoted by δ F . However, for this reference, DNS solution, these quantities are
denoted by sL and δ L , i.e. the laminar flame speed and thickness.
The measure of flame thickness we use is the secant thickness, defined as follows. Let
x1/4 and x3/4 denote the locations at which the composition φ has the values 1/4 and 3/4,
respectively. Then we define the thickness as

δF = 2(x3/4 − x1/4 ). (34)

As illustrated in Figure 2, the geometric interpretation of this thickness is that it is the

distance between the intersections of the secant through x1/4 and x3/4 and the pure stream
values, φ = 0 and φ = 1. We use the secant thickness because it has the simple geometric
interpretation given above, and because it can be extracted from numerical simulations
more robustly than the gradient thickness. We also define the centre of the flame to be at
the mid-point between x1/4 and x3/4 . Thus, when the flame is centred at the origin (x = 0),
it follows that x1/4 /δ F = −1/4 and x3/4 /δ F = 1/4 .
The (normalised) laminar flame speed and thickness are found to be sL /(Du /τc )1/2 =
0.81 and δL /(Du τc )1/2 = 7.6.

4.4. Numerically-accurate solution for the mean-drift model

As mentioned in Section 3.5, the numerically-accurate solution of the LES/PDF mean-drift
model equations have the correct behaviour in the DNS limit. Thus the solution for φ  is
just the laminar flame solution, and the variance of φ is zero. The value specified for the
mixing rate  is immaterial.
848 R.R. Tirunagari and S.B. Pope

1.4

1.2

0.8
uF /sL
0.6

0.4

0.2

0
10-2 10-1 100 101
Ωu τ c

Figure 3. For the random-walk model, the normalised flame speed as a function of the normalised
mixing rate. Dashed line: 3(u τ c )1/2 . (Colour online).

4.5

3.5

3
δF /δL

2.5

1.5

0.5

0
10-2 10-1 100 101
Ωu τ c

Figure 4. For the random-walk model, the normalised flame thickness as a function of the normalised
mixing rate. Dashed line: 3.7(u τ c )−1/2 . (Colour online).

4.5. Numerically-accurate solution for the random-walk model

In contrast, the random-walk model does not have the correct behaviour in the DNS limit.
In this case the solution depends on the specified (normalised) mixing rate τ c .
Numerically-accurate solutions for the random-walk model are obtained using the
mesh-free particle method described in Appendix A. Consistent with Equation (10), the
mixing rate is taken to be proportional to the molecular diffusivity, i.e.

D
 = u , (35)
Du

with the unburnt mixing rate, u , being a specified parameter.

 Combustion Theory and Modelling 849

0
10

σφ,max 10-1

10-2 -2 -1 0 1
10 10 10 10
Ωu τ c

Figure 5. For the random-walk model, the maximum standard deviation of composition as a function
of the normalised mixing rate. Dashed line: 0.14(u τ c )−1/2 . (Colour online).

The computed normalised flame speed and thickness are shown in Figures 3 and 4 as
functions of the normalised unburnt mixing rate, u τ c . As may be seen from these figures,
as u τ c increases beyond unity, the flame speed and thickness become close to their laminar
values.
Since the composition is non-reactive for φ < φ R = 0.4, only by mixing does the
composition rise from zero to φ R . As a consequence, for small u τ c , mixing is the rate-
limiting process. If one assumes that (for small u τ c ) the flame speed and thickness are
determined by D and , independently of τ c , then dimensional arguments dictate that uF
scales with (Du u )1/2 , and that δ F scales with (Du / u )1/2 . As may be seen from Figures 3
and 4, the results for the four smallest values of u τ c investigated are consistent with these
scalings.
It is interesting to observe that, as u τ c decreases, uF /sL first increases and attains a
maximum of about 1.3 around u τ c ≈ 0.3, before decreasing, eventually with the mixing-
limited scaling.
Figure 5 shows σ φ, max , the maximum value (over all x) of the standard deviation of the
composition. For small u τ c , the standard deviation has the quite large value around 0.4,
independent of τ c . For large u τ c , it decreases as (u τ c )−1/2 , consistent with the balance
of production and dissipation.
Figure 6 shows a scatter plot of particle composition through the flame. For the case
shown (u τ c = 0.01), mixing is rate limiting, and reaction in comparison is very fast. As
a consequence, there are very few partially-burnt particles (i.e. 0.4 < φ ∗ < 1), and many
fully-burnt particles (i.e. φ ∗ = 1).
In summary, the numerically-accurate solution to the random-walk model in the DNS
limit depends on the normalised mixing rate u τ c . Because the residual variance is non-
zero, the model behaves incorrectly in the DNS limit. However, for u τ c greater than
unity, the flame speed and flame thickness are close to the laminar values, and the residual
standard deviation is less than 10%. Consequently, in spite of the spurious residual variance,
the random-walk model may be considered to be a useful and accurate model provided that
the mixing rate is sufficiently large.
850 R.R. Tirunagari and S.B. Pope

0.8

0.6

φ*, φ
0.4

0.2

-6 -4 -2 0 2 4 6 8
x/δL

Figure 6. For the random-walk model with u τ c = 0.01, a scatter plot of particle composition
. Dashed line: laminar flame profile. (Colour
φ ∗ against position. Solid line: resolved composition φ
online).

4.6. Cloud-in-cell solutions to the model equations

In practice, the LES/PDF equations are not solved accurately (as considered in the previous
subsection), but instead are solved by a numerical method that may incur significant error.
This is because the mesh spacing h is typically chosen not to be very small compared
to the resolution length scale , and indeed may be larger than the smallest scale in the
accurate solution to the LES/PDF equations (e.g. the flame thickness δ F ). We therefore
consider here the numerical solution of the LES/PDF equations using the random-walk and
mean-drift models obtained using a cloud-in-cell particle/mesh method. This is the type of
method typically used in practice. The details of the current implementation are given in
Appendix B.
For the given test case being considered, the numerical solution to the LES/PDF equa-
tions depends on three quantities: the specified resolution scale , the specified mixing-
model constant Cm , and the mesh spacing h. (We take the number of particles N to be large
enough, and the time step t to be small enough, such that only the spatial discretisation
error is significant.) It is convenient to reduce the number of parameters that need to be
considered to two, which is possible since Cm and  enter solely through the combination
Cm /2 (in Equation 10). To this end, we re-express the mixing rate as


D
 = CD , (36)
h2

where comparison with Equation (10) shows that CD and Cm are related by

 2
h
CD = Cm . (37)


We then take the two parameters to be considered to be CD and h/δ L . Note that, for example,
the value CD = 1 corresponds both to Cm = 1, h/ = 1, and to Cm = 4, h/ = 1/2, etc.
 Combustion Theory and Modelling 851

1.8
RW, CD =1
1.6 RW, CD =4
MD, CD =1
1.4 MD, CD =4

1.2

uF /sL
1

0.8

0.6

0.4

0.2

0
10-1 100 101
h/δL

Figure 7. Using the cloud-in-cell method, the normalised flame speed as a function of the normalised
mesh spacing for (from bottom to top at the right): the RW model with CD = 1; the MD model with
CD = 1; the RW model with CD = 4; the MD model with CD = 4. (Colour online).

Figure 7 shows the computed flame speed as a function of the normalised mesh spacing
for both models and for two values of the model coefficient CD . For small h/δ L both models
attain the same asymptote, but with uF /sL being greater than unity, specifically 1.2 for CD =
1 and 1.6 for CD = 4. As with the accurate solution, with increasing mesh spacing, the
flame speed of the random-walk model achieves a maximum, before decreasing. On the
other hand the flame speed of the mean-drift model decreases monotonically, and is similar
to that of the random walk model for large h/δ L .
The key to understanding many of these observations is to appreciate that the CIC im-
plementation of the IEM model incurs a numerical smearing error. A simple analysis shows
that, to first order, this error is equivalent to there being an additional numerical diffusivity
Dn , which is proportional to h2 . With α denoting the coefficient of proportionality, we
have


Dn = αh2 = αCD D. (38)

Note that, for fixed CD , this numerical diffusivity does not vanish as h/δ L tends to zero,
and it increases with CD . This explains the observed values of uF /sL (for small h/δ L ) being
greater than unity, and larger for CD = 4 than for CD = 1. (Below, we empirically determine
the value α = 0.4.)
This reasoning can be made quantitative as follows. We define the augmented diffusivity
to be

 + Dn = D(1
Da ≡ D  + αCD ). (39)

Just as the laminar flame speed and thickness scale as (Du /τc )1/2 and (Du τc )1/2 , respec-
tively, the numerical flame speed and thickness can be expected to scale as (Da /τc )1/2 and
852 R.R. Tirunagari and S.B. Pope

1.2
RW, CD =1
RW, CD =4
1 MD, CD =1
MD, CD =4

0.8

uF /sa 0.6

0.4

0.2

0
10-2 10-1 100 101
h/δa

Figure 8. Using the cloud-in-cell method, the flame speed normalised by sa as a function of the
mesh spacing normalised by δ a for (from bottom to top at the right): the RW model with CD = 1; the
RW model with CD = 4; the MD model with CD = 1; the MD model with CD = 4. (Colour online).

(Da τc )1/2 , respectively, and hence to be larger than the laminar values by a factor of
 1/2
Da
= (1 + αCD )1/2 . (40)
Du

We thus define the augmented laminar flame speed and thickness to be

sa ≡ (1 + αCD )1/2 sL , (41)

and

δa ≡ (1 + αCD )1/2 δL , (42)

and, in Figure 8, we re-plot the computed flame speeds, but now with normalisation by sa
and δ a . As may be seen, with the empirically-determined value of α = 0.4, this scaling is
successful in accurately yielding uF = sa for small h/δ a for all four cases.
Figure 9 shows the computed flame thicknesses normalised by δ a . As may be seen, in
all four cases, δ F converges to δ a as the mesh is refined. For large h, the flame thickness
scales essentially with h, with δ F ≈ 4h for the random-walk model with CD = 1, and δ F ≈
h for the mean-drift model with CD = 4.
In the above we consider CD to be a fixed parameter. If instead we revert to considering
Cm to be fixed, then from Equation (41) we obtain (for small h/δ L and h/)

 2 1/2
sa h
= 1 + αCm
sL 
 2
1 h
≈ 1 + αCm , (43)
2 
 Combustion Theory and Modelling 853

RW, CD =1
RW, CD =4
MD, CD =1
MD, CD =4

101

δF /δa

0
10
-2 -1 0 1
10 10 10 10
h/δa

Figure 9. Using the cloud-in-cell method, the flame thickness as a function of the mesh spacing,
both normalised by δ a for (from bottom to top at the right): MD model, CD = 4; MD model, CD = 1;
RW model, CD = 4; RW model, CD = 1. The lower and upper dashed lines are δ F = h and δ F = 4h,
respectively. (Colour online).

and similarly for δ a /δ L . The important conclusions are that, for fixed Cm , the correct laminar
flame speed and thickness are obtained as the grid spacing h/δ L tends to zero when /δ L is
fixed, but not when h/ is fixed.
The last result – that the correct solution is not obtained in the limit /δ L → 0, h/δ L →
0, for fixed h/ – is perhaps surprising and deserves further comment.
Consider first the numerical solution of the implied equation (Equation 12) for the re-
solved mean composition  r is given by the Smagorinsky
φ in which the residual diffusivity D
model instead of being taken to be zero. In this case, compared to the DNS equation for com-
position (Equation 3), the implied equation for the mean composition contains a spurious
diffusion term of order (/δ L )2 . Even if a numerical error of order one relative magnitude
is incurred (e.g. by using the coarse mesh h/ = 1), since the term itself vanishes in the
limit /δ L → 0, the numerical error does not prevent the solution from converging to the
DNS solution.
Returning to the CIC solutions of the PDF equations, the fundamental reason that they
do not converge to the DNS solutions (for fixed h/) is that (as previously observed) the
LES-PDF equations do not converge to the DNS-PDF equation: the IEM model introduces
the term in  which, far from vanishing, in fact tends to infinity in the limit /δ L → 0.

4.7. Matching the laminar flame speed

It is natural to ask: With the CIC implementation of the LES/PDF models, is it possible
to specify the mixing rate  so that the flame speed uF matches the laminar flame speed
sL , even if the mesh size h is not small compared to the laminar flame thickness δ L ? The
answer is: Yes.
First consider the case when h/δ L is large. In this case: the numerical diffusion Dn =
αh2 dominates the molecular diffusion; the flame thickness scales with h (see Figure 9);
and the reaction is not rate limiting. Hence the flame speed is determined dominantly by
Dn and h, with D and τ c not being significant. It follows (on dimensional grounds) that
the flame speed scales as h, and hence, to match the laminar flame speed (which scales
854 R.R. Tirunagari and S.B. Pope

100

CL

-1
10

-2 -1 0 1
10 10 10 10
h/δL

Figure 10. Empirical coefficients CL (h/δ L ) (Equations 45 and 46) yielding flame speeds equal to the
laminar flame speed for the CIC implementation of the random-walk model (lower curve at left) and
of the mean-drift model (upper curve at left). (Colour online).

as D/δL ),  has to be specified to scale as D/(hδL ). Thus, for large h/δ L , the appropriate
specification is


D
 = CL , (44)
hδL

where CL is a constant. More generally, we can consider CL to be a coefficient dependent

on h/δ L , which tends to a constant for large h/δ L .
Before proceeding, we make the following observation based on the flame speed scaling
with h for large h/δ L . It follows from Equation (44) that, by construction, for fixed CL ,
uF /sL is independent of h/δ L . It further follows from the relations between CL , CD and Cm
that uF /sL varies as (h/δ L )−1 for fixed CD , and as hδ L /2 = (h/)2 (h/δ L )−1 for fixed Cm .
We determine empirically that the function CL (h/δ L ) which yields uF = sL is well
approximated for the random-walk model by
  
h
CL = 3.3 1 − exp −2 , (45)
δL

and for the mean-drift model by

 1/2 
h
CL = min 2.5, 3.5 . (46)
δL

These functions are shown in Figure 10.

Note that, with everything except h fixed: for fixed CD ,  varies as h−2 ; for large h/δ L
the above expressions for CL imply that  varies as h−1 ; while for small h/δ L , CL for
the RW model (Equation 45) implies that  varies as h0 , whereas CL for the MD model
(Equation 46) implies that  varies as h−1/2 .
 Combustion Theory and Modelling 855

1.5
RW
1.4 MD

1.3

1.2

1.1

uF /sL 1

0.9

0.8

0.7

0.6

0.5
10-1 100 101
h/δL

Figure 11. Normalised flame speed against normalised mesh spacing for the CIC implementations
of the random-walk and mean-drift models with the mixing rate specified by Equations (44)–(46).
(Colour online).

RW
MD

1
10
δF /δL

0
10 -1 0 1
10 10 10
h/δL

Figure 12. Normalised flame thickness against normalised mesh spacing for the CIC implementa-
tions of the random-walk model using Equation (45) (upper curve) and the mean-drift model using
Equation (46) (lower curve). The dashed line is δ F = h. (Colour online).

The computed flame speeds using  given by Equations (44)–(46) are shown in
Figure 11. As may be seen, with these specifications of CL , both models do indeed yield
flame speeds close to the laminar flame speed.
While both models yield the same flame speed, their solutions are quite different. While
the mean-drift model is consistent with the DNS limit in producing no residual fluctuations,
the random-walk model has large residual fluctuations (around 30%), even for small h/δ L .
This is because, for the RW model with CL given by Equation (45), as h/δ L tends to zero,
 tends to the value 6.6D/δ  2 , independent of h; whereas, in order to suppress residual
L
fluctuations and hence to reproduce the DNS limit,  must vary as a negative power of h/δ L
(as it does for a fixed value of CD ).
856 R.R. Tirunagari and S.B. Pope

As a further manifestation of the differences between the model solutions, Figure 12

shows the normalised flame thicknesses. For small h/δ L , the mean-drift model correctly
yields δ F = δ L ; and for large h/δ L it yields δ F ≈ h. In contrast, the random-walk model
yields consistently thicker flames, with δ F ≈ 2δ L for small h/δ L .
We conclude this subsection with the following caveats and reservations. A great
virtue of PDF methods is that they are generally applicable to all modes and regimes
of combustion. However, the specification of the mixing rate  by Equations (44)–(46)
is anything but general. It is particular to homogeneously-premixed combustion, so that a
unique laminar flame thickness δ L can be defined. Furthermore, the appropriate specification
of the coefficient CL depends on the details of the numerical implementation.

5. Strained methane/air premixed laminar flame

5.1. Definition of the test case
The flow considered consists of two opposed round jets emanating from nozzles of diameter
D = 12.7 mm placed at a distance of d = 16 mm apart. One of the nozzles emits a
fresh (unburnt) mixture of CH4 /O2 /N2 at equivalence ratio φ u = 0.85 (O2 /N2 = 30/70
molar ratio), temperature Tu = 294 K and pressure 1 atm. The other nozzle emits fully-
burnt combustion products at temperature Tad = 2435 K, corresponding to the chemical-
equilibrium mixture with the same enthalpy, elemental composition and pressure as the
reactant stream. The axial velocity of the reactant stream is Uu = 1 m/s, and that of
the product stream is taken to be twice this value so that the resulting laminar flame is
approximately mid-way between the nozzles. The nominal mean strain rate is thus 2Uu /d =
125s−1 . The radial velocities at the nozzle exits are taken to be zero.
The axial coordinate is denoted by z, with z = − 8 mm being the product-stream nozzle
exit, and z = 8 mm being the reactant-stream nozzle exit.
The methane combustion is described by a 16-species augmented reduced mechanism
for methane oxidation [27]. The thermal diffusivity is obtained from the CHEMKIN trans-
port library, and then the (equal) species diffusivities are obtained by taking the Lewis
numbers to be unity.

5.2. OPPDIF calculations

As is well known [28], for the flow considered, the profiles along the centreline can be
obtained from the solution of 1D equations. These equations are solved using the OPPDIF
code within CHEMKIN-PRO [22]. Figure 13(a) shows the computed axial profiles of the
axial velocity and temperature. Figure 13(b) shows the profile of ρU/ρ u . The value of the
laminar flame speed sL is taken to be the value of ρU/ρ u at the location of maximum
temperature gradient. This location is around z = 4 mm, and the resulting value of sL is
55.8 cm/s. The secant thickness (obtained from the normalised temperature profile (T −
Tu )/(Tad − Tu ) and Equation (34)) is δ L = 0.334 mm.

5.3. LES/PDF calculations

LES/PDF calculations are performed in the DNS limit using the coupled NGA/HPDF
code [6,29,30]. The number of grid points in the radial and azimuthal directions are kept
constant at 48 and 4 respectively, while the number of grid points in the axial direction,
nx , is varied from 12 to 384 in order to span the range of values of /δ L typically used in
 Combustion Theory and Modelling 857

5 2500 1

4
2000
0.5
3

ρU ρu (m/s)
1500
U (m/s)

T (K)
0
1
1000

0
−0.5
500
−1

−2 0 −1
−8 −6 −4 −2 0 2 4 6 8 −8 −6 −4 −2 0 2 4 6 8
z (mm) z (mm)
(a) (b)

Figure 13. Laminar profiles from the OPPDIF calculations of (a) velocity and temperature and (b)
ρU/ρ u as a function of axial coordinate between the two nozzles. (Colour online).

Figure 14. Contour plot of normalised filter-width from the previous 3D LES/PDF simulations of
turbulent premixed counterflow flames described in [10]. (Colour online).

LES/PDF calculations. To illustrate this range, Figure 14 shows a contour plot of the values
of /δ L observed in recent 3D LES/PDF simulations of turbulent premixed counterflow
flame [10]. Note that the filter-width in the present simulations is taken to be equal to the
(uniform) grid spacing in the axial direction, i.e.  = h.
The results reported below depend on: the model used (random walk or mean drift); the
value of the mixing-rate constant Cm ; and the specified axial grid spacing and resolution
length scale h = . The values of Cm used are 1.0 and 4.0. By using different grids, h/δ L
is varied between 0.13 and 4.0.

5.4. Results
Figures 15 and 17 show the normalised flame speed and flame thickness as functions of
the normalised grid spacing for both models and for both values of Cm investigated. The
858 R.R. Tirunagari and S.B. Pope

Cm = 4.0, RW
Cm = 4.0, MD
Cm = 1.0, RW
0 Cm = 1.0, MD
10

uF /sL −1
10

−2
10 −1 0 1
10 10 10
h δL

Figure 15. Normalised flame speed as a function of normalised grid spacing for: Cm = 4.0, RW
(blue dot); Cm = 4.0, MD (blue star); Cm = 1.0, RW (red dot); Cm = 1.0, MD (red star). (The curves
for Cm = 1.0 are below the corresponding curves for Cm = 4.0.) (Colour online).

flame speed uF and the flame thickness δ F are obtained from the centreline profiles of the
resolved velocity, density and temperature in the same way as for the OPPDIF calculations
(as described in Section 5.2).
For the flame speed, the observed behaviour (Figure 15) is broadly the same as for the
CIC solution for the idealised, unstrained 1D flame considered in Section 4.6 (see Figure 7).
For the random-walk model, uF /sL has an asymptote, larger than unity, for small h/δ L , and
then decreases with increasing h/δ L . The values of uF /sL for Cm = 4 are consistently higher
than those for Cm = 1. The only qualitative difference compared to the results in Figure 7
is that a local maximum in uF /sL around h/δ L ≈ 0.4 is not observed.
For the mean-drift model, the results are qualitatively similar. The only point of note
is that, for Cm = 1, the asymptotic value of uF /sL is about 0.9, i.e. less than unity. Further
investigation reveals that this is due to time-stepping error. As shown in Figure 16, as the
time step decreases, uF /sL increases, plausibly exceeding unity for smaller time steps than
investigated.
Figure 17 shows the normalised flame thickness as a function of the normalised grid
spacing. The observations are the same as for the CIC solutions for the idealised, unstrained
flame (see Figure 9). The flame thickness is generally larger for the random-walk model
than for the mean-drift model, and decreases with increasing Cm . On the finest grids, δ F /δ L
approaches an asymptote greater than unity; while on coarse grids the flame thickness is
approximately δ F ≈ 2.5h (in three out of the four cases).
Figure 18 shows scatter plots on the centreline of the particle temperature versus
position, colour coded by the mass fraction of OH. The left-hand column is for the random-
walk model, and the right-hand column for the mean-drift model. The top two rows are
for Cm = 4, and the bottom two for Cm = 1. The first and third rows are for the relatively
coarse grid (h/δ = 1.4), while the second and fourth rows are for the finer grid (h/δ = 0.5).
The symbols on the top axis of each plot show the locations of the grid nodes.
The first observation to be made from these scatter plots is that the implementation
of the mean-drift model is successful in producing negligible residual fluctuations in all
cases. It may also be observed that the particles are, to some extent, able to resolve the
 Combustion Theory and Modelling 859

1.4
Cm = 4.0, MD, h/δL = 0.25
Cm = 1.0, MD, h/δL = 0.25

1.2

uF /sL

0.8 −3 −2
10 10
Δt/τc

Figure 16. Normalised flame speed as a function of normalised time step for Cm = 4.0, MD, h/δ L =
0.25 (upper curve) and Cm = 1.0, MD, h/δ L = 0.25 (lower curve). Here, τ c is defined by τ c ≡ δ L /sL .
(Colour online).

Cm = 4.0, RW
Cm = 4.0, MD
1 Cm = 1.0, RW
10 Cm = 1.0, MD
δF δL

0
10 −1 0 1
10 10 10
h δL

Figure 17. Normalised flame thickness as a function of normalised grid spacing for: Cm = 4.0, RW
(blue dot); Cm = 4.0, MD (blue star); Cm = 1.0, RW (red dot); Cm = 1.0, MD (red star). The dashed
line is δ F = 2.5h. (The curves for Cm = 1.0 are above the corresponding curves for Cm = 4.0.) (Colour
online).

flame profile below the grid scale. For the random-walk model there is appreciable scatter
(inconsistent with the DNS limit), which decreases with increasing Cm and with decreasing
h/δ L .
The results presented in this section serve to confirm the validity of the model im-
plementations in the 3D LES/PDF code in the DNS limit. They also confirm that the
results obtained for the idealised, unstrained flame using the 1D MATLAB script are R

representative of those obtained with the much more complex 3D LES/PDF code.
860 R.R. Tirunagari and S.B. Pope

Figure 18. Scatter plots of particle temperature on the centreline colour coded by particle OH
species mass fraction for (a) Cm = 4.0, RW, h/δ L = 1.4, (b) Cm = 4.0, MD, h/δ L = 1.4, (c) Cm = 4.0,
RW, h/δ L = 0.5, (d) Cm = 4.0, MD, h/δ L = 0.5, (e) Cm = 1.0, RW, h/δ L = 1.4, (f) Cm = 1.0, MD,
h/δ L = 1.4, (g) Cm = 1.0, RW, h/δ L = 0.5, (h) Cm = 1.0, MD, h/δ L = 0.5. The grid points are shown
as black circles at the top of the figures. (Colour online).
 Combustion Theory and Modelling 861

6. Discussion
Based on the results presented above and other considerations, it is clear that the mean-
drift model has significant theoretical advantages over the random-walk model. These
advantages are as follows.

(1) Differential diffusion can be implemented in the mean-drift model, whereas with
the random-walk model the unity-Lewis-number assumption is unavoidable.
(2) The mean-drift model behaves correctly in the DNS limit, whereas the random-
walk model has spurious production of variance leading to significant residual
fluctuations.
(3) In the CIC implementation of the models, on a given mesh, the flame thickness δ F
given by the mean-drift model is significantly smaller (i.e. closer to δ L ) than that
given by the random-walk model.

However, the mean-drift model also has distinct disadvantages in ease of implementation
and computational cost. Specifically:

(i) with the mean-drift model, special measures must be taken to guarantee realizability
(e.g. imposing a lower limit on the mixing rate to ensure that species mass fractions
remain non-negative [25,26]);
(ii) the mean-drift model requires the evaluation of the nφ resolved composition fields

φ(x, t) (which is computationally expensive) and the solution of nφ PDEs to deter-
mine the mean drift.

In contrast, the random-walk model can be implemented in an efficient, mesh-free algo-

rithm that automatically guarantees realizability and does not require the evaluation of the
resolved compositions. Because of these practical advantages, and in spite of its theoretical
disadvantages, we do not advocate abandoning the random-walk model: it may be ade-
quate (and simpler and cheaper) for a range of problems in which differential diffusion and
consistency with the DNS limit are not crucial.

7. Conclusions
The principal conclusions to be drawn from this work are as follows.

(1) In considering LES in the DNS limit, it is advantageous to define the resolved
quantities as conditional means rather than as filtered fields. With the conditional-
mean approach, in the DNS limit, the resolved fields in LES are just the DNS fields,
and there are no residual fluctuations. In contrast, with the filtering approach, the
resolved fields differ from the DNS fields (by of order (/δ)2 ), and the residual
variances are non-zero (and are, again, of order (/δ)2 ).
(2) An LES method is said to be consistent in the DNS limit if the numerical solution
of the LES equations converges to the solution of the DNS equations in the limit
/δ → 0 and h/δ → 0 (possibly with additional requirements on the ratio h/).
(3) An LES method is said to behave correctly in the DNS limit if it is consistent and
it approaches the limit correctly. For the LES-PDF equations applied to laminar
flames, the correct behaviour is that the residual composition fluctuations are zero,
independently of /δ L .
862 R.R. Tirunagari and S.B. Pope

(4) The behaviour of LES/PDF models in the DNS limit has been studied by applying
them to two laminar premixed flames: an idealised unstrained flame, and a strained
methane/air flame.
(5) The numerically-accurate solutions to the LES/PDF equations incorporating the
mean-drift model exhibit the correct behaviour in the DNS limit.
(6) A mesh-free particle method has been implemented to obtain numerically-accurate
solutions to the LES/PDF equation incorporating the random-walk model (i.e. the
RW-PDF equation).
(7) These numerically-accurate solutions to the RW-PDF equation exhibit an incorrect
behaviour in the DNS limit, in that the residual fluctuations are non-zero. However,
if the mixing rate  is sufficiently large (e.g. τ c  1), then the flame speed and
thickness are close to the laminar values, and the residual fluctuations are not large
(e.g. less than 10%).
(8) The cloud-in-cell implementation of the IEM mixing model incurs a smearing error
which is equivalent to a numerical diffusion Dn that scales as h2 .
(9) With the standard specification of the mixing rate, with a constant value of Cm , as
the mesh is refined (i.e. h/δ L → 0), the LES/PDF solutions are consistent with the
DNS limit if the resolution scale is fixed (i.e. /δ L is fixed), but not if h/ is fixed.
(10) On coarse meshes (h/δ L  1): the flame thickness δ F scales with the mesh spacing
h; the mixing rate  and the mesh spacing h are the controlling parameters (with
the molecular diffusivity D and the reaction timescale τ c not being controlling);
and hence the flame speed uF scales as h.
(11) On coarse meshes, and with Cm and h/ fixed, the normalised flame speed uF /sL
scales with (h/δ L )−1 .
(12) With both the RW and the MD model, it is possible to specify the mixing rate (by
Equations 44–46) so that the flame speed uF matches the laminar flame speed sL ,
even on coarse meshes. It is emphasised that this is a non-general model.
(13) Computations have been performed of a strained, premixed, methane/air, opposed-
jet flame, using the same 3D code used in previous LES/PDF studies. The results
confirm the validity of the model implementations in the 3D code in the DNS limit;
and they confirm that the results obtained for the idealised, unstrained flame are
representative of those obtained with the much more complex 3D LES/PDF code.
(14) Whereas the mean-drift model has theoretical advantages (in allowing differential
diffusion to be implemented and behaving correctly in the DNS limit), the random-
walk model has the practical advantages of being simpler to implement and being
computationally less expensive.

Acknowledgements
We gratefully acknowledge the California Institute of Technology, the University of Colorado at
Boulder and Stanford University for licensing the NGA code used in this work.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This material is based upon work supported by the US Department of Energy Office of Science,
Office of Basic Energy Sciences [Award Number DE-FG02-90ER14128].
 Combustion Theory and Modelling 863

References
[1] S.B. Pope, PDF methods for turbulent reactive flows, Prog. Energy Combust. Sci. 11 (1985),
pp. 119–192.
[2] D.C. Haworth, Progress in probability density function methods for turbulent reacting flows,
Prog. Energy Combust. Sci. 36 (2010), pp. 168–259.
[3] S.B. Pope, Small scales, many species and the manifold challenges of turbulent combustion,
Proc. Combust. Inst. 34 (2013), pp. 1–31.
[4] F.A. Jaberi, P.J. Colucci, S. James, P. Givi and S.B. Pope, Filtered mass density function for
large-eddy simulation of turbulent reacting flows, J. Fluid Mech. 401 (1999), pp. 85–121.
[5] V. Raman and H. Pitsch, A consistent LES/filtered-density function formulation for the simula-
tion of turbulent flames with detailed chemistry, Proc. Combust. Inst. 31 (2007), pp. 1711–1719.
[6] Y. Yang, H. Wang, S.B. Pope and J.H. Chen, Large-eddy simulation/probability density function
modeling of a non-premixed CO/H2 temporally evolving jet flame, Proc. Combust. Inst. 34
(2013), pp. 1241–1249.
[7] I.A. Dodoulas and S. Navarro-Martinez, Large eddy simulation of premixed turbulent flames
using the probability density function approach, Flow Turbul. Combust. 90 (2013), pp. 645–
678. Available at http://dx.doi.org/10.1007/s10494-013-9446-z.
[8] J. Kim and S.B. Pope, Effects of combined dimension reduction and tabulation on the simula-
tions of a turbulent premixed flame using large-eddy simulation/probability density function,
Combust. Theory Model. 18 (2014), pp. 388–413.
[9] S.B. Pope and R. Tirunagari, Advances in probability density function methods for turbulent
reactive flows, in Proceedings of the Nineteenth Australasian Fluid Mechanics Conference,
RMIT University, Melbourne, 2014.
[10] R.R. Tirunagari and S.B. Pope, An investigation of turbulent premixed counterflow flames using
large-eddy simulations and probability density function methods, Combust. Flame 166 (2016),
pp. 229–242.
[11] S.B. Pope, Turbulent Flows, Cambridge University Press, Cambridge, UK, 2000.
[12] H. Pitsch, Large-eddy simulation of turbulent combustion, Annu. Rev. Fluid. Mech. 38 (2006),
pp. 453–482.
[13] S.B. Pope, Self-conditioned fields for large-eddy simulations of turbulent flows, J. Fluid Mech.
652 (2010), pp. 139–169.
[14] R.S. Rogallo and P. Moin, Numerical simulation of turbulent flows, Annu. Rev. Fluid. Mech.
17 (1985), pp. 99–137.
[15] R.O. Fox, Computational Models for Turbulent Reactive Flows, Cambridge University Press,
New York, 2003.
[16] S.B. Pope, Ten questions concerning the large-eddy simulation of turbulent flows, New J. Phys.
6 (2004), Article No. 35. Available at http://dx.doi.org/10.1088/1367-2630/6/1/035.
[17] K.A. Kemenov and S.B. Pope, Molecular diffusion effects in LES of a piloted methane–air
flame, Combust. Flame 157 (2011), pp. 240–254.
[18] R.S. Barlow and J.H. Frank, Effects of turbulence on species mass fraction in methane/air jet
flames, Proc. Combust. Inst. 27 (1998), pp. 1087–1095.
[19] O.T. Stein, B. Boehm, A. Dreizler and A.M. Kempf, Highly-resolved LES and PIV analysis
of isothermal turbulent opposed jets for combustion applications, Flow Turbul. Combust. 87
(2011), pp. 425–447.
[20] A.M. Ruiz, G. Lacaze and J.C. Oefelein, Flow topologies and turbulence scales
in a jet-in-cross-flow, Phys. Fluids 27 (2015), Article ID 045101. Available at
http://dx.doi.org/10.1063/1.4915065.
[21] J. Villermaux and J.C. Devillon, Représentation de la coalescence et de la redispersion des
domaines de ségrégation dans un fluide par un modèle d’interaction phénoménologique, in
Proceedings of the 2nd International Symposium on Chemical Reaction Engineering, 2–4 May
1972, Amsterdam. Elsevier, New York, 1972, pp. B1–B13.
[22] R.J. Kee, F.M. Rupley, J.A. Miller, M.E. Coltrin, J.F. Grcar, E. Meeks, H.K. Moffat, A.E. Lutz,
G. Dixon-Lewis, M.D. Smooke, J. Warnatz, G.H. Evans, R.S. Larson, R.E. Mitchell, L.R.
Petzold, W.C. Reynolds, M. Caracotsios, W.E. Stewart, P. Glarborg, C. Wang and O. Adigun,
CHEMKIN Collection, Release 3.6, Reaction Design, San Diego, CA, 2000.
[23] S.B. Pope, Computations of turbulent combustion: progress and challenges, Proc. Combust.
Inst. 23 (1990), pp. 591–612.
864 R.R. Tirunagari and S.B. Pope

[24] P.J. Colucci, F.A. Jaberi, P. Givi and S.B. Pope, Filtered density function for large eddy simulation
of turbulent reacting flows, Phys. Fluids 10 (1998), pp. 499–515.
[25] R. McDermott and S.B. Pope, A particle formulation for treating differential diffusion in filtered
density function methods, J. Comput. Phys. 226 (2007), pp. 947–993.
[26] S. Viswanathan, H. Wang and S.B. Pope, Numerical implementation of mixing and molecular
transport in LES/PDF studies of turbulent reacting flows, J. Comput. Phys. 230 (2011), pp.
6916–6957.
[27] C.J. Sung, C.K. Law and J.Y. Chen, An augmented reduced mechanism for methane oxidation
with comprehensive global parametric validation, Proc. Combust. Inst. 27 (1998), pp. 295–304.
[28] R.J. Kee, M.E. Coltrin and P. Glarborg, Chemically Reacting Flow: Theory and Practice, Wiley,
Hoboken, NJ, 2003.
[29] O. Desjardins, G. Blanquart, G. Balarac and H. Pitsch, High order conservative finite difference
scheme for variable density low Mach number turbulent flows, J. Comput. Phys. 227 (2008),
pp. 7125–7159.
[30] H. Wang and S.B. Pope, Large eddy simulation/probability density function modeling of a
turbulent CH4 /H2 /N2 jet flame, Proc. Combust. Inst. 33 (2011), pp. 1319–1330.
[31] S.B. Pope, A model for turbulent mixing based on shadow-position conditioning, Phys. Fluids
25 (2013), Article ID 110803. Available at http://dx.doi.org/10.1063/1.4818981.
[32] S.B. Pope and R. Gadh, Fitting noisy data using cross-validated cubic smoothing splines,
Commun. Stat. Simul. Comput. 17 (1988), pp. 349–376.
[33] M. Muradoglu, S.B. Pope and D.A. Caughey, The hybrid method for the PDF equations of
turbulent reactive flows: consistency conditions and correction algorithms, J. Comput. Phys.
172 (2001), pp. 841–878.

Appendix A. Mesh-free particle method

A mesh-free particle method is used to obtain numerically-accurate solutions to the LES/PDF equa-
tions for the random-walk model in the DNS limit applied to the idealised, unstrained, 1D, plane
laminar flame. The equations are solved in the frame in which the reactants are at rest at x = −∞.
From a specified initial condition, the solution is advanced in time through small time steps t, and
eventually reaches a fully-developed state, with the flame moving to the left at a constant speed uF .
The solution domain extends from pure reactants to pure products, and is moved with the flame.
There are a large number N of particles, the nth having position X(n) (t) and composition φ (n) (t)
at time t. The solution is advanced from time t to time t + t through the following sequence of
operations (with some details to follow). Prior to the start of each step, the particles are sorted in
position so that X(n) ≥ X(n − 1) .

(1) The particle diffusivities are evaluated, and the mean diffusivity at the particle locations,
(n) , is estimated using a cross-validated smoothing linear spline.
D
(2) The mixing rate  (n) is determined from D (n) .
(3) The mixing sub-step is performed using the near-neighbour implementation of the IEM
model [31].
(4) The reaction sub-step is performed.
(5) The particles are moved by the diffusion term of the random walk, and then re-sorted.
(6) The particles are moved by the drift term in the random walk.
(7) As necessary, the domain is moved, and particles are removed and added as necessary at the
boundaries.

In order to estimate means, the particles are partitioned into odd and even-numbered particles.
Linear smoothing splines [32] (with smoothing parameter s) are generated based on the two sets of
particles, and the cross-validation error χ (s) is computed. An iteration is performed to determine the
value of the smoothing parameter that minimises the error. (For computational efficiency, if there are
a large number of particles, then this procedure is applied, not to individual particles, but to clusters
of adjacent particles. Here the number of clusters is limited to 1000.)
The particles move due to the drift term in the random walk (Equation 8) according to dX∗ /dt =
 +
[U  ∗ . Fortunately, however, this drift coefficient does not need to be evaluated, because
v ∇(ρD)]
the effect of the drift can be implemented implicitly by enforcing a consistency condition [33]. Every
 Combustion Theory and Modelling 865

particle is ascribed the same mass per unit cross-sectional area, m . The nth particle, with specific
volume v (n) , therefore has a length associated with it of (n) = m v (n) . The consistency condition is
that the particle length is equal to the geometric length that the particle occupies. This condition is
strongly enforced by (at the end of the time step) setting the particle positions by

1
X (n+1) = X(n) + [(n+1) + (n) ]. (A1)
2

The solutions reported in Section 4.5 are performed with N = 105 particles, which is more
than needed to produce very accurate results. Tests are performed to ensure that the time step t
is sufficiently small, and the simulation duration sufficiently long, so that the associated numerical
errors are negligible.

Appendix B. Cloud-in-cell particle/mesh method

The cloud-in-cell (CIC) particle/mesh method is applicable to both the random-walk and the mean-
drift models, and is similar to the method generally used in LES/PDF simulations.
As in the mesh-free method, there are N computational particles. There is also a mesh of uniform
spacing h. There is a linear-spline basis function associated with each mesh node. Means are estimated
at the nodes using kernel estimation with the basis functions as the kernels, and then the means are
represented by linear splines.
The differences compared to the mesh-free method are as follows.

(1) All means are estimated and represented using the CIC linear basis functions.
(2) The IEM mixing model is implemented directly using the CIC means   and φ by integrating
the equation

dφ ∗
∗ ).
= −∗ (φ ∗ − φ (B1)
dt

(3) The mean drift is implemented by

(t + t) − φ
φ ∗ (t + t) = φ ∗ (t) + [φ (t)]∗ , (B2)

(x, t + t) is the solution after time t of the partial differential equation
where φ

∂ρφ ∂ 
∂φ
= 
ρD , (B3)
∂t ∂x ∂x

(x, t) = φ
from the initial condition φ (x, t). Equation (B3) is integrated numerically at the
nodes, using a fully implicit three-point difference scheme.