This action might not be possible to undo. Are you sure you want to continue?

# Turbulent combustion modeling

Denis Veynante

a,

*

, Luc Vervisch

b

a

Laboratoire E.M2.C., CNRS et Ecole Centrale Paris, Grande Voie des Vignes, 92295 ChaÃtenay-Malabry Cedex, France

b

Institut National des Sciences AppliqueÂes de Rouen, UMR CNRS 6614/CORIA, Campus du Madrillet,

Avenue de l'UniversiteÂ ÐBP 8, 76801 Saint Etienne du Rouvray Cedex, France

Received 4 November 2000; accepted 12 October 2001

Abstract

Numerical simulation of ¯ames is a growing ®eld bringing important improvements to our understanding of combustion. The

main issues and related closures of turbulent combustion modeling are reviewed. Combustion problems involve strong coupling

between chemistry, transport and ¯uid dynamics. The basic properties of laminar ¯ames are ®rst presented along with the major

tools developed for modeling turbulent combustion. The links between the available closures are illuminated from a generic

description of modeling tools. Then, examples of numerical models for mean burning rates are discussed for premixed turbulent

combustion. The use of direct numerical simulation (DNS) as a research instrument is illustrated for turbulent transport

occurring in premixed combustion, gradient and counter-gradient modeling of turbulent ¯uxes is addressed. Finally, a review

of the models for non-premixed turbulent ¯ames is given. q 2002 Published by Elsevier Science Ltd.

Keywords: Premixed ¯ames; Non-premixed ¯ames; Turbulent combustion; Scalar turbulent transport; Direct numerical simulation; Reynolds

averaged Navier±Stokes modeling

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

2. Balance equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

2.1. Instantaneous balance equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

2.2. Reynolds and Favre averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

2.3. Favre averaged balance equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

2.4. Filtering and Large Eddy Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

3. Major properties of premixed, non-premixed and partially premixed ¯ames . . . . . . . . . . . . . . . . . . . . 200

3.1. Laminar premixed ¯ames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

3.2. Laminar diffusion ¯ames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

3.3. Partially premixed ¯ames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

4. A direct analysis: Taylor's expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

5. Scales and diagrams for turbulent combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.2. Turbulent premixed combustion diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.2.2. Combustion regimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

5.2.3. Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

5.3. Non-premixed turbulent combustion diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

5.3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

Progress in Energy and Combustion Science 28 (2002) 193±266 PERGAMON

www.elsevier.com/locate/pecs

0360-1285/02/$ - see front matter q 2002 Published by Elsevier Science Ltd.

PII: S0360-1285(01)00017-X

* Corresponding author. Tel.: 133-1-41-13-10-80; fax: 133-1-47-02-80-35.

E-mail address: denis@em2c.ecp.fr (D. Veynante).

6. Tools for turbulent combustion modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

6.2. Scalar dissipation rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

6.3. Geometrical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

6.3.1. G-®eld equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

6.3.2. Flame surface density description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216

6.3.3. Flame wrinkling description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

6.4. Statistical approaches: probability density function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

6.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

6.4.2. Presumed probability density functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

6.4.3. Pdf balance equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

6.4.4. Joint velocity/concentrations pdf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

6.4.5. Conditional moment closure (CMC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

6.5. Similarities and links between the tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

7. Reynolds-averaged models for turbulent premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

7.1. Turbulent ¯ame speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

7.2. Eddy-break-up model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

7.3. Bray±Moss±Libby model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

7.3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

7.3.2. BML model analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

7.3.3. Recovering mean reaction rate from tools relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

7.3.4. Reynolds and Favre averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

7.3.5. Conditional averagingÐcounter-gradient turbulent transport . . . . . . . . . . . . . . . . . . . . . 227

7.3.6. Extensions to partially premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

7.4. Models based on the ¯ame surface area estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

7.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

7.4.2. Algebraic expressions for the ¯ame surface density S . . . . . . . . . . . . . . . . . . . . . . . . . . 229

7.4.3. Flame surface density balance equation closures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

7.4.4. Analysis of the ¯ame surface density balance equation . . . . . . . . . . . . . . . . . . . . . . . . . 234

7.4.4.1. Turbulent transport. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234

7.4.4.2. Strain rate induced by the mean ¯ow ®eld, A

T

. . . . . . . . . . . . . . . . . . . . . . . . 234

7.4.4.3. Strain rate a

T

due to turbulent motions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

7.4.4.4. Propagation and curvature terms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236

7.4.5. Flame stabilization modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

7.4.6. A related approach: G-equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

8. Turbulent transport in premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

8.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

8.2. Direct numerical simulation analysis of turbulent transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238

8.2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238

8.2.2. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238

8.3. Physical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

8.3.1. Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241

8.4. External pressure gradient effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

8.5. Counter-gradient transportÐexperimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

8.6. To include counter-gradient turbulent transport in modeling . . . . . . . . . . . . . . . . . . . . . 243

8.7. Towards a conditional turbulence modeling? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

9. Reynolds averaged models for non-premixed turbulent combustion . . . . . . . . . . . . . . . . . . . . . . 245

9.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

9.2. Fuel/air mixing modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

9.2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

9.2.2. Balance equation and simple relaxation model for ~ x . . . . . . . . . . . . . . . . . . . 246

9.3. Models assuming in®nitely fast chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

9.3.1. Eddy dissipation model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

9.3.2. Presumed pdf: in®nitely fast chemistry model . . . . . . . . . . . . . . . . . . . . . . . . 247

9.4. Flamelet modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

9.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 194

9.4.2. Flame structure in composition space, Y

SLFM

i

(Z

p

; x

p

) . . . . . . . . . . . . . . . . . . . 249

9.4.3. Mixing modeling in SLFM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

9.4.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

9.5. Flame surface density modeling, coherent ¯ame model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

9.6. MIL model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

9.7. Conditional moment closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

9.8. Pdf modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

9.8.1. Turbulent micromixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255

9.8.2. Linear relaxation model, IEM/LMSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256

9.8.3. GIEM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256

9.8.4. Stochastic micromixing closures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

9.8.5. Interlinks PDF/¯ame surface modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

9.8.6. Joint velocity/concentrations pdf modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

10. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260

1. Introduction

The number of combustion systems used in power

generation and transportation industries is growing

rapidly. This induces pollution and environmental problems

to become critical factors in our societies. The accurate

control of turbulent ¯ames therefore appears as a real

challenge.

Computing is now truly on par with experiment and

theory as a research tool to produce multi-scale information

that is not available by using any other technique. Computa-

tional ¯uid dynamics (CFD) is ef®ciently used to improve

the design of aerodynamical systems, and today no real

progress in design can be made without using CFD. With

the same objectives, much work has been devoted to turbu-

lent combustion modeling, following a variety of

approaches and distinct modeling strategies. This paper is

intended to provide a generic review of these numerical

models.

A wide range of coupled problems are involved in turbu-

lent ¯ames:

² The ¯uid mechanical properties of the combustion system

must be well known to carefully describe the mixing

between reactants and, more generally, all transfer

phenomena occurring in turbulent ¯ames (heat transfer,

molecular diffusion, convection, turbulent transport,

etc.).

² Detailed chemical reaction schemes are necessary to esti-

mate the consumption rate of the fuel, the formation of

combustion products and pollutant species. A precise

knowledge of the chemistry is absolutely required to

predict ignition, stabilization or extinction of reaction

zones together with pollution.

² Two (liquid fuel) and three (solid fuel) phase systems may

be encountered. Liquid fuel injection is a common pro-

cedure and the three-dimensional spatial distribution of

gaseous reactants depends on complex interactions

between the breakdown of the liquid sheets, the vapor-

ization of the liquid, turbulent mixing, and droplet

combustion.

² Radiative heat transfer is generated within the ¯ame by

some species and carbon particles resulting from soot

formation and transported by the ¯ow motion. In

furnaces, walls also interact with combustion through

radiative transfer.

Turbulent combustion modeling is therefore a very

broad subject. All the aspects of the problem are not

addressed in the present review. We will only focus on

the closure schemes developed and used to understand

and calculate turbulent transport and mean burning rates

in turbulent ¯ames. The detail of chemistry, its reduc-

tion, tabulation, etc. are not considered. However, the

links existing between the models are identi®ed, showing

similarities which are sometimes much stronger than is

usually thought.

Numerical modeling of ¯ames is developed from the

following steps (Fig. 1):

² Under assumptions such as the high activation energy

limit, asymptotic analysis [1±4] allows the analytical

determination of ¯ame properties in well-de®ned model

problems (ignition, propagation of ¯ame front, instabil-

ities and acoustics, etc.). This approach, limited to simpli-

®ed situations, leads to analytical results exhibiting

helpful scale factors (dimensionless numbers) and

major ¯ame behaviors. Asymptotic analysis is particu-

larly well suited to perform quantitative comparison

between various phenomena.

² Simpli®ed experiments are useful to understand the basic

properties of combustion (laminar ¯ames, ¯ame/vortex

interactions, etc.) [5,6]. These experiments are accom-

panied by numerical simulations of laminar ¯ames incor-

porating complex chemistry and multi-species transports

along with radiative heat losses [7].

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 195

² For given chemistry and transport model, in direct

numerical simulation (DNS) all the scales of the turbu-

lence (time and length) are calculated without resorting to

closures for turbulent ¯uxes and mean burning rate.

Turbulent ¯ames are analyzed in simple con®gurations

to extract data impossible to measure in experiments, and

to isolate some speci®c phenomena (heat release, Lewis

number¼) [8±12].

Because of the large number of degrees of freedominvolved

in turbulent combustion, a full DNS of a practical system

cannot be performed and averaging techniques leading to

unclosed equations are necessary. Models for turbulent ¯ames

are then developed: closure techniques are proposed for

unknown terms found in exact averaged balance equations.

Once the models have been implemented in numerical

codes, validation procedures are required. The numerical

modeling is validated against measurements obtained from

experiments. Con®gurations as close as possible to actual

industrial systems are chosen for these tests. Then, the ultimate

step is the simulation of a real combustion device.

The decomposition discussed in Fig. 1 is quite formal.

Turbulent combustion modeling is actually a continuous

ring between theoretical studies to analyze combustion,

understand ¯ames and improve models, implementation of

these models into CFD, experimental measurements

and comparison between these experimental data and the

numerical results.

Following a short presentation of the balance equa-

tions for reactive ¯ows (Section 2), a ®rst part is

devoted to a brief description of laminar ¯ames (Section

3). After a presentation of the unsuccessful Taylor's

expansion for closing the mean burning rate (Section

4), the physical analysis leading to turbulent combustion

diagrams is developed (Section 5). Then, modeling tools

available to derive turbulent combustion models are

described and the relations between a priori quite differ-

ent formalisms are established (Section 6). The next

three sections are devoted to combustion modeling in

the context of Reynolds Averaged Navier±Stokes

(RANS) equations. For premixed turbulent combustion,

we review the available closures for the mean reaction

rate (Section 7) and turbulent transport (Section 8). In a

subsequent section (Section 9), the modeling of the

mean burning rate in non-premixed turbulent ¯ames is

addressed.

2. Balance equations

2.1. Instantaneous balance equations

The basic set of balance equations comprises the classical

Navier±Stokes, species and energy transport equations.

These instantaneous local balance equations are, using the

classical lettering [13±15]:

² Mass:

2r

2t

1

2ru

j

2x

j

= 0 (1)

² Momentum (i = 1,2,3):

2ru

i

2t

1

2ru

j

u

i

2x

j

= 2

2p

2x

i

1

2t

ij

2x

j

1F

i

(2)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 196

Fig. 1. Combustion modeling steps.

where t

ij

denotes the viscous force tensor and F

i

a body

force.

² Species (N species with k = 1; ¼; N):

2rY

k

2t

1

2ru

j

Y

k

2x

j

= 2

2J

k

j

2x

j

1 _ v

k

(3)

where J

k

j

is the molecular diffusive ¯ux of the species k

and _ v

k

the mass reaction rate of this species per unit

volume.

² Total enthalpy h

t

= h 1u

i

u

i

=2 :

2rh

t

2t

1

2ru

j

h

t

2x

j

=

2p

2t

1

2

2x

j

(J

h

j

1u

i

t

ij

) 1u

j

F

j

(4)

where u

i

t

ij

and u

i

F

j

denote respectively the power due to

viscous and body forces.

These equations are closed by expressions for the species

molecular ¯uxes and the viscous forces. In practical situa-

tions, all ¯uids are assumed to be Newtonian, i.e. the viscous

tensor is given by the Newton law:

t

ij

= m

l

2u

i

2x

j

1

2u

j

2x

i

_ _

2

2

3

m

l

d

ij

2u

k

2x

k

_ _

(5)

where the molecular viscosity m

l

depending on the ¯uid

properties is introduced. d

ij

is the Kronecker symbol.

Species molecular diffusivities are generally described

using the Fick law, assuming a major species:

J

k

j

= 2

m

l

Sc

k

2Y

k

2x

j

(6)

Sc

k

is the Schmidt number of the species k, de®ned as:

Sc

k

=

m

l

rD

k

(7)

D

k

is the molecular diffusivity of the species k relative to the

major species.

More complex expressions may be used to describe multi-

species molecular diffusion. Soret effect (species diffusion

under temperature gradients) and molecular transport due to

pressure gradients are usually neglected. Enthalpy diffusion

is described according to the Fourier law:

J

h

j

= 2

m

l

Pr

2h

2x

j

1

N

k=1

Pr

Sc

k

21

_ _

h

k

2Y

k

2x

j

_ _

(8)

The Prandtl number Pr compares the diffusive transport

of momentum (viscous forces) and temperature. In the

previous expressions, radiative heat transfer and Dufour

effect (enthalpy diffusion under mass fraction gradients)

are neglected. The Prandtl number is written as a function

of the thermal diffusivity l and the constant pressure

speci®c heat C

p

:

Pr =

m

i

C

p

l

_ _

(9)

Then, the Lewis number Le

k

of the species k, comparing

thermal and mass diffusivities is introduced:

Le

k

=

Sc

k

Pr

_ _

=

l

rC

p

D

k

_ _

(10)

Under the assumption of unity Lewis number, the enthalpy

diffusive ¯ux (Eq. (8)) is simpli®ed and mass fraction and

enthalpy balance equations are formally identical if 2P=2t;

u

i

t

ij

and u

i

F

j

are negligible (low Mach number assumption)

[16]. This assumption is generally made to simplify tur-

bulent ¯ame modeling, especially in premixed ¯ames

when species mass fractions and temperature are assumed

to be equivalent variables. Nevertheless, thermo-diffusive

instabilities occur in premixed systems when the Lewis

number is lower than unity (for example for hydrogen).

One direct consequence of these instabilities is an increase

of the premixed ¯ame area and of the global reaction rate

[14,17].

2.2. Reynolds and Favre averaging

Unfortunately, the full numerical solution of the instanta-

neous balance equations is limited to very simpli®ed cases

(DNS [9,11,12]), where the number of time and length

scales present in the ¯ow is not too great. To overcome

this dif®culty, an additional step is introduced by averaging

the balance equations to describe only the mean ¯ow ®eld

(local ¯uctuations and turbulent structures are integrated in

mean quantities and these structures have no longer to be

described in the simulation). Each quantity Q is split into a

mean

Q and a deviation from the mean denoted by Q

/

:

Q =

Q 1Q

/

with

Q

/

= 0 (11)

Then, the previous instantaneous balance equations may be

ensemble averaged to derive transport equations for the

mean quantity

Q: This classical Reynolds averaging tech-

nique, widely used in non-reacting ¯uid mechanics, brings

unclosed correlations such as u

/

Q

/

that are unknown and

must be modeled. The numerical procedure is called

Reynolds Averaged Navier±Stokes (RANS) modeling.

In turbulent ¯ames, ¯uctuations of density are observed

because of the thermal heat release, and Reynolds averaging

induces some additional dif®culties. Averaging the mass

balance equation leads to:

2 r

2t

1

2

2x

i

ru

i

1r

/

u

/

i

_ _

= 0 (12)

where the velocity/density ¯uctuations correlation r

/

u

/

i

appears. To avoid the explicit modeling of such correlations,

a Favre (mass weighted) [18] average

~

Q is introduced and

any quantity is then decomposed into Q =

~

Q 1Q

//

:

~

Q =

rQ

r

;

¯

Q

//

=

r Q 2

~

Q

_ _

r

= 0 (13)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 197

The Favre averaged continuity equation:

2 r

2t

1

2 r ~ u

i

2x

i

= 0 (14)

is then formally identical to the Reynolds averaged conti-

nuity equation for constant density ¯ows. This result is true

for any balance equations (momentum, energy, mass frac-

tions, etc.). Nevertheless, Favre averaging is only a math-

ematical formalism:

² There is no simple relation between Favre,

~

Q and

Reynolds,

Q; averages. A relation between

~

Q and

Q

requires the knowledge of density ¯uctuations correla-

tions r

/

Q

/

remaining hidden in Favre averaging (see

Section 7.3.4):

r

~

Q = r

Q 1r

/

Q

/

(15)

² Comparisons between numerical simulations, providing

Favre averaged quantities

~

Q; with experimental results

are not obvious. Most experimental techniques determine

Reynolds averaged data

Q and differences between

~

Q and

**Q may be signi®cant (Section 7.3.4 and Fig. 17).
**

2.3. Favre averaged balance equations

Averaging instantaneous balance equations yields:

² Mass:

2 r

2t

1

2 r ~ u

j

2x

j

= 0 (16)

² Momentum (i = 1,2,3):

2 r ~ u

i

2t

1

2 r ~ u

j

~ u

i

2x

j

= 2

2 r

¯

u

//

i

u

//

j

2x

j

2

2 p

2x

i

1

2 t

ij

2x

j

1

F

i

(17)

² Chemical species (for N species, k = 1; ¼; N):

2 r

~

Y

k

2t

1

2 r ~ u

j

~

Y

k

2x

j

= 2

2 r

¯

u

//

j

Y

//

k

2x

j

2

2J

k

j

2x

j

1

_ v

k

(18)

² Total enthalpy

~

h

t

:

2 r

~

h

t

2t

1

2 r ~ u

j

~

h

t

2x

j

= 2

2 r

¯

u

//

j

h

//

t

2x

j

1

2 p

2t

1

2

2x

j

J

h

j

1u

i

t

ij

_ _

1u

j

F

j

(19)

The objective of turbulent combustion modeling is to

propose closures for the unknown quantities appearing in

the averaged balance equations, such as:

² Reynolds stresses

¯

u

//

i

u

//

j

: The turbulence model provides

an approximation for this term. The closure may be done

directly or by deriving balance equations for these

Reynolds stresses. However, most combustion works

are based on turbulence modeling developed for non-

reacting ¯ows, such as k±1; simply rewritten in terms

of Favre averaging, and heat release effects on the

Reynolds stresses are generally not explicitly included.

² Species (

¯

u

//

j

Y

//

k

) and temperature (

¯

u

//

j

T) turbulent ¯uxes.

These ¯uxes are usually closed using a gradient transport

hypothesis:

r

¯

u

//

j

Y

//

k

= 2

m

t

Sc

kt

2

~

Y

k

2x

j

(20)

where m

t

is the turbulent viscosity, estimated from the

turbulence model, and Sc

kt

a turbulent Schmidt number

for the species k.

Nonetheless, theoretical and experimental works have

demonstrated that this assumption may be wrong in

some premixed turbulent ¯ames and counter-gradient

turbulent transport may be observed [19,20] (i.e. in an

opposite direction compared to the one predicted by Eq.

(20), see Sections 7.3.5 and 8).

² Laminar diffusive ¯uxes J

k

j

; J

h

j

; etc. are usually small

compared to turbulent transport, assuming a suf®ciently

large turbulence level (large Reynolds numbers limit).

² Species chemical reaction rates

_ v

k

: Turbulent combus-

tion modeling generally focuses on the closure of these

mean burning rates.

These equations, closed with appropriate models, allow

only for the determination of mean quantities, that may

differ from the instantaneous ones. Strong unsteady mixing

effects, resulting from the rolling up of shear layers, are

observed in turbulent ¯ames, and the knowledge of steady

statistical means is indeed not always suf®cient to describe

turbulent combustion. An alternative is to use large eddy

simulation (LES).

2.4. Filtering and Large Eddy Simulation

The objective of Large Eddy Simulation (LES) is to ex-

plicitly compute the largest structures of the ¯ow (typically

the structures larger than the computational mesh size),

while the effects of the smaller one are modeled. LES is

widely studied in the context of non-reactive ¯ows [21±

24], its application to combustion modeling is still at an

early stage [25]. As in RANS, the complex coupling

between micromixing and chemical reactions occurring at

unresolved scales needs models, however, LES possesses

some attractive properties:

² Large structures in turbulent ¯ows generally depend on

the geometry of the system. On the contrary, smaller

scales feature more universal properties. Accordingly,

turbulence models may be more ef®cient when they

have to describe only the smallest structures.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 198

² Turbulent mixing controls most of the global ¯ame prop-

erties. In LES, unsteady large scale mixing (between

fresh and burnt gases in premixed ¯ames or between

fuel and oxidizer in non-premixed burners) is simulated,

instead of being averaged.

² Most reacting ¯ows exhibit large scale coherent struc-

tures [26], which are also especially observed when

combustion instabilities occur. These instabilities result

from the coupling between heat release, the hydrody-

namic ¯ow ®eld and acoustic waves. They need to be

avoided because they induce noise variations of the

main properties of the system, large heat transfers

and, even in some extreme cases, the destruction of

the device.LES may be a powerful tool to predict the

occurrence of such instabilities [27] and consequently

improve passive or active control systems.

² With LES, large structures are explicitly computed and

instantaneous fresh and burnt gases zones, with different

turbulence characteristics (Section 8.7) are clearly iden-

ti®ed. This may help to describe some properties of the

¯ame/turbulence interaction.

In LES, the relevant quantities Q are ®ltered in the spec-

tral space (components greater than a given cut-off

frequency are suppressed) or in the physical space (weighted

averaging in a given volume). The ®ltered operation is

de®ned by:

Q(x) =

_

Q(x

p

)F(x 2x

p

) dx

p

(21)

where F is the LES ®lter. Standard ®lters are:

² A cut-off ®lter in the spectral space:

F(k) =

1 if k # p=D

0 otherwise

_

(22)

where k is the spatial wave number. This ®lter preserves

the length scales greater than the cut-off length scale 2D:

² A box ®lter in the physical space:

F(x) = F(x

1

; x

2

; x

3

) =

1=D

3

if ux

i

u # D=2; i = 1; 2; 3

0 otherwise

_

(23)

where (x

1

; x

2

; x

3

) are the spatial coordinates of the loca-

tion x. This ®lter corresponds to an averaging of the

quantity Q over a box of size D.

² A Gaussian ®lter in the physical space:

F(x) = F(x

1

; x

2

; x

3

)

=

6

pD

2

_ _

3=2

exp 2

6

D

2

x

2

1

1x

2

2

1x

2

3

_ _

_ _

(24)

All these ®lters are normalized:

_11

21

_11

21

_11

21

F(x

1

; x

2

; x

3

) dx

1

; dx

2

; dx

3

= 1 (25)

In combusting ¯ows, a mass-weighted, Favre ®ltering, is

introduced as:

r

~

Q(x) =

_

rQ(x

p

)F(x 2x

p

)dx

p

(26)

Instantaneous balance equations (Section 2) may be ®ltered

to derived balance equations for the ®ltered quantities

Q or

~

Q: This derivation should be carefully conducted:

Any quantity Q may be decomposed into a ®ltered

component

Q and a `¯uctuating' component Q

/

; according

to: Q =

Q 1Q

/

: But, in disagreement with classical

Reynolds averaging (ensemble average), Q

/

may be non-

zero:

Q

/

(x) =

_

Q(x

p

) 2

Q(x

p

)

_ _

F(x 2x

p

)dx

p

=

_

Q(x

p

)F(x 2x

p

)dx

p

2

_

Q(x

p

)F(x 2x

p

)dx

p

=

Q(x) 2Q(x) (27)

where

Q(x) =

_ _

Q(x

1

)F(x

p

2x

1

)dx

1

_ _

F(x 2x

p

)dx

p

=

__

Q(x

1

)F(x

p

2x

1

)F(x 2x

p

)dx

1

dx

p

±

Q(x)

(28)

To summarize:

Q ±

Q; Q

/

± 0;

~

~

Q ±

~

Q;

¯

Q

//

± 0 (29)

The relations used in RANS Q =

Q; Q

/

= 0;

~

~

Q =

~

Q;

¯

Q

//

= 0

are true when a cut-off ®lter in the spectral space is chosen

(Eq. (22)). Then, all the frequency components greater than

a cut-off wave number k

c

= p=D vanish.

The derivation of balance equations for the ®ltered quan-

tities

Q or

~

Q requires the exchange of ®ltering and differ-

entiation operators. This exchange is theoretically valid only

under restrictive assumptions and is wrong, for example,

when the ®lter size varies (®lter size corresponding to the

mesh size, depending on the spatial location). This point has

been carefully investigated [28]. In most simulations, the

uncertainties due to this operator exchange are neglected

and assumed to be incorporated in subgrid scale modeling.

Filtering the instantaneous balance equations leads to

equations formally similar to the Reynolds averaged balance

equations given in Section 2.3:

² mass:

2 r

2t

1

2 r ~ u

j

2x

j

= 0 (30)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 199

² momentum (for i = 1; 2; 3):

2 r ~ u

i

2t

1

2 r ~ u

j

~ u

i

2x

j

= 2

2

2x

j

r ¯ u

i

u

j

2 ~ u

i

~ u

j

_ _ _ _

2

2 p

2x

i

1

2 t

ij

2x

j

1

F

i

(31)

² Chemical species (N species, k = 1; ¼; N):

2 r

~

Y

j

2t

1

2 r ~ u

j

~

Y

k

2x

j

= 2

2

2x

j

r

¯

u

j

Y

k

2 ~ u

j

~

Y

k

_ _ _ _

1

_ v

k

(32)

² Total enthalpy h

t

= h 1u

i

u

i

=2

2 r

~

h

t

2t

1

2 r ~ u

j

~

h

t

2x

j

= 2

2

2x

j

r

¯

u

j

h

t

2 ~ u

j

~

h

t

_ _ _ _

1

2 p

2t

1

2

2x

j

J

h

j

1u

i

t

ij

_ _

1u

j

F

j

(33)

where

Q and

~

Q denote LES ®ltered quantities instead of

ensemble means.

The unknown quantities are:

² Unresolved Reynolds stresses

_

¯ u

i

u

j

2 ~ u

i

~ u

j

_

; requiring a

subgrid scale turbulence model.

² Unresolved species ¯uxes

_

¯

u

j

Y

k

2 ~ u

j

~

Y

k

_

and enthalpy

¯uxes

_

¯

u

j

h

t

2 ~ u

j

~

h

t

_

:

² Filtered laminar diffusion ¯uxes J

k

j

; J

h

j

:

² Filtered chemical reaction rate

_ v

k

:

These ®ltered balance equations, coupled to subgrid scale

models may be numerically solved to simulate the unsteady

behavior of the ®ltered ®elds. Compared to direct numerical

simulations (DNS), part of the information contained in the

unresolved scales is lost (and should be modeled).

Compared to RANS, LES provides valuable information

on the large resolved motions.

Either using RANS or LES, combustion occurs at the

unresolved scales of the computations. Then, the basic

tools and formalism of turbulent combustion modeling are

somehow the same for both techniques. Most of the RANS

combustion models can be modi®ed and adapted to LES

modeling. The scope of this review is limited to RANS.

3. Major properties of premixed, non-premixed and

partially premixed ¯ames

3.1. Laminar premixed ¯ames

The structure of a laminar premixed ¯ame is displayed in

Fig. 2. Fresh gases (fuel and oxidizer mixed at the molecular

level) and burnt gases (combustion products) are separated

by a thin reaction zone (typical thermal ¯ame thicknesses,

d

l

, are about 0.1±1 mm). A strong temperature gradient is

observed (typical ratios between burnt and fresh gases

temperatures are about 5±7). Another characteristic of a

premixed ¯ame is its ability to propagate towards the

fresh gases. Because of the temperature gradient and the

corresponding thermal ¯uxes, fresh gases are preheated

and then start to burn. The local imbalance between diffu-

sion of heat and chemical consumption leads to the propa-

gation of the front. The propagation speed S

L

of a laminar

¯ame depends on various parameters (fuel and oxidizer

compositions, fresh gases temperature, etc.) and is about

0.1±1 m/s. There is an interesting relation between the ther-

mal ¯ame thickness, d

l

, the laminar ¯ame speed, S

L

and the

kinematic viscosity of the fresh gases, n:

Re

f

=

d

l

S

L

n

< 4 (34)

where the thermal thickness d

l

corresponds to a temperature

jump of 98% of the temperature difference between fresh

and fully burnt products. The ¯ame Reynolds number, Re

f

,

is then almost constant. This relation, derived, for example,

from the Zeldovich/Frank-Kamenetskii (ZFK) theory

[14,15] is often implicitly used in theoretical derivation of

models for premixed turbulent combustion.

For a simple one-step irreversible chemical scheme:

reactants - products

the ¯ame is described using a progress variable c, such as

c = 0 in the fresh gases and c = 1 in the fully burnt ones.

This progress variable may be de®ned as a reduced tempera-

ture or a reduced mass fraction:

c =

T 2T

u

T

b

2T

u

or c =

Y

F

2Y

u

F

Y

b

F

2Y

u

F

(35)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 200

Fig. 2. Structure of a laminar plane premixed ¯ame.

where T, T

u

and T

b

are respectively the local, the unburnt gas

and the burnt gas temperatures. Y

F

, Y

u

F

and Y

b

F

are respec-

tively the local, unburnt gas and burnt gas fuel mass frac-

tions. Y

b

F

is non-zero for a rich combustion (fuel in excess).

For a unity Lewis number (same molecular and thermal

diffusivities), without heat losses (adiabatic combustion)

and compressibility effects, the two de®nitions (35) are

equivalent and mass and low Mach number energy balance

equations reduce to a single balance equation for the

progress variable:

2rc

2t

17´ ruc

_ _

= 7´ rD7c

_ _

1 _ v (36)

The previous Eq. (36) may be recast in a propagative form,

introducing the displacement speed v of the iso-c surface:

2c

2t

1u´7c =

1

r

7´ rD7c

_ _

1 _ v

u7cu

_ _

.........,,.........

displacement speed

7c } } = wu7cu (37)

Eq. (36) then describes the displacement of an iso-c surface

with the displacement speed w measured relative to the ¯ow.

Introducing the vector n normal to the iso-c surface and

pointing towards fresh gases (n = 27c=u7cu); the displace-

ment speed may be split into three contributions:

w =

1

r 7c } }

nn : 7 rD7c

_ _

2D7´n 1

1

r 7c } }

_ v (38)

w = 2

1

ru7cu

2

2n

rDu7cu

_ _

........,,........

w

n

2 D7´n

..,,..

w

c

1

1

ru7cu

_ v

...,,...

w

r

where 2=2n = n´7 denotes a normal derivative. w

n

corre-

sponds to molecular diffusion normal to the iso-c surface,

w

c

is related to the curvature 7´n of this surface and corre-

sponds to tangential diffusion. w

r

is due to the reaction rate

_ v: In a ®rst approximation, w

n

1w

r

may be modeled with

the laminar ¯ame speed, S

L

, whereas w

c

incorporates wrink-

ling surface effects and may be expressed using Markstein

lengths [29].

The propagation of reactive fronts has been the subject of

various developments and more discussion may be found in

Ref. [2] and references therein.

3.2. Laminar diffusion ¯ames

In laminar diffusion ¯ames, fuel and oxidizer are on both

sides of a reaction zone where the heat is released. The

burning rate is controlled by the molecular diffusion of the

reactants toward the reaction zone (Figs. 3 and 4). In a

counter-¯owing fuel and oxidizer ¯ame (Fig. 4), the amount

of heat transported away from the reaction zone is exactly

balanced by the heat released by combustion. A steady

planar diffusion ¯ame with determined thickness may be

observed in the vicinity of the stagnation point. Increasing

the jet velocity, quenching occurs when the heat ¯uxes

leaving the reaction zone are greater than the chemical

heat production. The structure of a steady diffusion ¯ame

therefore depends on ratios between characteristic times

representative of molecular diffusion and chemistry [30].

The thicknesses of the mixing zone and of the reaction

zone vary with these characteristic times. In opposition

with premixed ¯ames:

² Diffusion ¯ames do not bene®t from a self-induced propa-

gation mechanism, but are mainly mixing controlled.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 201

Fig. 3. Generic structure of a laminar diffusion ¯ame.

Fig. 4. Sketch of a counter-¯owing fuel and oxidizer diffusion

¯ame.

² The thickness of a diffusion ¯ame is not constant, but

depends on the local ¯ow properties.

Let us consider the irreversible single step chemical reac-

tion between fuel and oxidizer:

F 1sO - (1 1s)P

where s is the mass stoichiometric coef®cient.

In term of mass fraction, this chemical reaction may be

written:

n

P

dY

P

= n

F

dY

F

1n

o

dY

o

where dY

F

, dY

O

and dY

P

are the variations of fuel, oxidizer

and product mass fractions. n

i

are the stoichiometric molar

coef®cients of the reaction, W

i

denotes the species molar

weight and _ v is the reaction rate. The balance equations

for mass fractions and temperature are necessary to identify

the properties of the ¯ame:

2rY

F

2t

17´(ruY

F

) = 7´(rD

F

7Y

F

) 2n

F

W

F

_ v

2rY

O

2t

17´(ruY

O

) = 7´(rD

O

7Y

O

) 2n

O

W

O

_ v

2rT

2t

17´(ruT) = 7´

l

C

P

7T

_ _

1n

F

W

F

Q

C

p

_ _

_ v

The molecular diffusion is expressed using the Fick law,

the chemical rate of fuel and oxidizer are respectively _ v

F

=

n

F

W

F

_ v and _ v

O

= n

O

W

O

_ v: Q is the amount of heat released

by the combustion of an unit mass of fuel.

The internal structure of diffusion ¯ames is usually

discussed using the extent of mixing between fuel and oxidi-

zer. It is ®rst assumed that fuel and oxidizer molecular

diffusivities are equal (i.e. D

F

= D

O

= D). Combining the

transport equation for Y

F

and Y

O

, a conserved scalar (quan-

tity that is not in¯uenced by the chemical reaction, a

Schwab±Zeldovitch variable) w(Y

F

; Y

O

) = Y

F

2Y

O

=s is

introduced, with the mass stoichiometric coef®cient s =

(n

O

W

O

=n

F

W

F

): The mixture fraction Z is then de®ned by

normalizing w using values in the fuel and oxidizer streams.

Z evolves through the diffusive layer from zero (oxidizer) to

unity (fuel):

Z =

f

Y

F

Y

F;o

2

Y

O

Y

O;o

11

f 11

(39)

Y

F,o

is the fuel mass fraction in the fuel feeding stream.

Similarly, Y

O,o

is the oxidizer mass fraction in the oxidizer

stream (for instance, in air, Y

O,o

<0.23), f is the chemical

equivalence ratio:

f =

sY

F;o

Y

O;o

(40)

The mixture fraction follows the balance equation:

2rZ

2t

17´(ruZ) = 7´(rD7Z) (41)

Other Schwab±Zeldovitch variables w(Y

F

, T) and w(Y

O

, T)

(conserved scalars) may be derived by combining the

variables (Y

F

, T) and (Y

O

, T). The mixture fraction and these

additional conserved scalars are linearly related and one

may write:

Y

O

(x; t) = Y

O;o

(1 2Z(x; t))

.......,,.......

Mixing

1

n

O

W

O

n

F

W

F

C

p

Q

_ _

[Z(x; t)(T

F;o

2T

O;o

) 1(T

O;o

2T(x; t))]

.......................,,.......................

Combustion

(42)

Y

F

(x; t) = Y

F;o

Z(x; t)

....,,....

Mixing

1

C

p

Q

[Z(x; t)(T

F;o

2T

O;o

) 1(T

O;o

2T(x; t))]

...................,,...................

Combustion

(43)

where T

O,o

and T

F,o

are the temperatures of the fuel and

oxidizer streams respectively. Using these algebraic re-

lations, the diffusion ¯ame is fully determined when the

mixture fraction Z and any one of T, Y

F

, or Y

O

is known.

The conserved scalar approach may still be useful when

fuel and oxidizer molecular diffusivities differ, but an ad-

ditional mixture fraction:

Z

L

=

F

Y

F

Y

F;o

2

Y

O

Y

O;o

11

F11

(44)

should be introduced, satisfying [31]:

r

DZ

Dt

=

1

L

7´

l

C

p

7Z

L

_ _

(45)

where:

L= Le

O

(1 1f)=(1 1F) with F = (Le

O

=Le

F

)f

(46)

where Le

i

is the Lewis number of the species i. The relations

between Z and Z

L

are given in Table 1. When Le

O

= Le

F

;

Z

L

= Z: In experiments or in simulations involving complex

chemistry, the mixture fraction is de®ned from mass frac-

tions of atomic elements [32].

Mass fractions and temperature balance equations may be

reorganized into a new frame where Z is one of the coordi-

nates (see for instance Ref. [14] or Ref. [33]). A local or-

thogonal coordinate system attached to the surface of

stoichiometric mixture is introduced and the derivatives in

the stoichiometric plane are denoted '. For unity Lewis

number and using Eq. (41), the species transport equation

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 202

may be written:

r

2Y

i

2t

1ru

'

´7

'

Y

i

= rx

2

2

Y

i

2Z

2

17

'

´(rD7

'

Y

i

)

2rD7

'

(lnu7Zu)´7

'

Y

i

1 _ v

i

(47)

In Eq. (47), x is the scalar dissipation rate of the mixture

fraction Z:

x = D

2Z

2x

j

2Z

2x

j

_ _

= Du7Zu

2

(48)

measuring the inverse of a diffusive time t

x

= x

21

: As this

time decreases, mass and heat transfers through the stoichio-

metric surface are enhanced.

When iso-Z surface curvatures are not too strong, the

gradients measured along the stoichiometric surface are

smaller than the gradients in the direction Z perpendicular

to the stoichiometric surface, the balance equation for the

mass fractions reduces to:

r

2Y

i

2t

= rx

2

2

Y

i

2Z

2

1 _ v

i

(49)

Neglecting unsteady effects, the time derivative vanishes

and for unity Lewis numbers, the ¯ame structure is fully

described by:

rx

2

2

Y

i

2Z

2

1 _ v

i

= 0 and rx

2

2

T

2Z

2

1 _ v

T

= 0 (50)

showing that the chemical reaction rate is directly related to

the function T(Z, x). Under these hypothesis, the diffusion

¯ame is completely determined as a function of the mixture

fraction Z and the scalar dissipation rate x (or 7Z):

Y

i

= Y

i

(Z; x); T = T(Z; x)

An expression for x(Z,t) and full solutions for various lami-

nar ¯ames may be derived from asymptotic developments

[30,34], or solving Eq. (50) leading to Fig. 5.

A coordinate j is de®ned across the one-dimensional

¯amelet such as:

dZ

dj

= 2n

Z

´7Z = u7Zu (51)

where n

Z

denotes the normal vector to the iso-Z surfaces,

pointing towards Z = 0: The reaction rate may be integrated

in the direction j through the ¯ame, using Eq. (50):

_

V

i

=

_11

21

_ v

i

(j)dj =

_1

0

_ v

i

(x; Z)

u7Zu

dZ

= 2

_1

0

rDu7Zu

2

2

Y

i

2Z

2

dZ (52)

Assuming that r, D and u7Zu do not vary across the ¯ame

(this is typical of a ¯amelet assumption used in turbulent

combustion modeling where the ¯ame is assumed very

thin),

_

V

i

becomes:

_

V

i

< 2rDu7Zu

2Y

i

2Z

_ _

Z=1

Z=0

= 2 rD

2Y

i

2j

_ _

Z=1

Z=0

(53)

This last relation illustrates how the integrated reaction

rate of a species i is directly related to the molecular diffu-

sion ¯ux of that species through the ¯ame.

Diffusion combustion is limited by two regimes corre-

sponding to pure mixing of the reactants and in®nitely fast

chemistry (Fig. 5). When the chemistry is in®nitely fast, the

temperature depends on mixing through Z, but not on the

rate of mixing x [35]. Then, piecewise relationships exist

between Z, Z

L

, species mass fractions and temperature,

summarized in Table 1. Eq. (43) provides the maximum

¯ame temperature T

f

obtained when Y

F

= Y

O

= 0 and Z =

Z

st

= 1=(1 1f)

T

f

=

T

F;o

1T

O;of

1Y

F;o

Q

C

p

1 1f

In many combustion systems, the in®nitely fast chemistry

hypothesis cannot be invoked everywhere. For example in

ignition problems or in the vicinity of stabilization zones,

and more generally when large velocity gradients are found.

The characterization of diffusion ¯ames from the in®nitely

fast chemistry situation to the quenching limit is therefore of

fundamental interest for turbulent combustion. The counter-

¯ow diffusion ¯ame (Fig. 4) is a generic con®guration well

suited to reproduce and to understand the structure and the

extinction of laminar diffusion ¯ames. These extinction

phenomena have been theoretically described using asymp-

totic developments [30,34,36]. A diffusive time t

x

< x

21

st

=

(Du7Zu

2

)

21

Z=Z

st

and a chemical time t

c

are combined to build a

DamkoÈhler number Da

p

= (t

x

=t

c

) < (t

c

x

st

)

21

: The response

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 203

Table 1

Piecewise relations for in®nitely fast chemistry including non-unity Lewis number. Z

st

= 1=(1 1f) and Z

L

st

= 1=(1 1F) (see Eq. (46)). The

subscript

o

denotes a quantity measured in pure fuel or oxidizer, T

f

is the ¯ame temperature

Oxidizer side

Z , Z

st

and Z

L

, Z

L

st

Fuel side

Z . Z

st

and Z

L

. Z

L

st

Z = Z

L

(1 1F)=(1 1f) Z = (f(Z

L

(1 1F) 21)=F11)=(1 1f)

Y

F

= 0 Y

O

= 0

Y

O

= Y

O;o

(1 2Z(1 1f)) Y

F

= Y

F;o

(Z(1 1f) 21)=f

T = (T

f

2T

O;o

)Z(f 11) 1T

O;o

T = (T

F;o

2T

f

)(Z(f 11) 21)=f 1T

f

of the burning rate to variations of Da

p

leads to the so-called

`S' curve (Fig. 6) [14]. Starting from a situation where the

chemistry is fast, decreasing Da

p

(increasing x) makes the

burning rate and transport through the stoichiometric

surface greater, until chemistry cannot keep up with the

large heat ¯uxes. Then, extinction develops. The value of

the DamkoÈhler Da

p

q

at the extinction point may be estimated

by quantifying the leakage of fuel (or oxidizer) through the

stoichiometric surface [37].

Two limit cases are thus important for non-premixed

turbulent combustion modeling: pure mixing without

combustion (Da

p

- 0) and in®nitely fast chemistry

(Da

p

-1). These cases delineate the domain where ¯ames

may develop in planes (Z, Y

F

), (Z, Y

O

) and (Z, T) (Fig. 5).

Moreover, for a given location within a diffusion ¯ame, by

traveling along the normal to the stoichiometric surface,

T(Z) can be constructed and characterizes the combustion

regime (i.e. fast or slow chemistry, Fig. 5). Many turbulent

combustion models are based on this description of diffusion

¯ame; when the ¯ow is turbulent, T(Z) is replaced by the

mean temperature calculated for a given value of Z, i.e. for a

given state in the mixing between fuel and oxidizer.

3.3. Partially premixed ¯ames

In non-premixed combustion, some partial premixing of

the reactants may exist before the reaction zone develops.

Then, the pure diffusive/reactive layer, as observed in a

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 204

Fig. 5. Inner structure of non-premixed ¯ames. The distribution in mixture fraction space of fuel, oxidizer and temperature lies between the

in®nitely fast chemistry limit and the pure mixing case. The thickness of the diffusive zone l

d

is estimated from the scalar dissipation rate x at

the stoichiometric surface, whereas the characteristic thickness of the reaction zone l

r

depends on both l

d

and the DamkoÈhler number [12].

laminar diffusion ¯ame, may not be the unique relevant

model problem. Furthermore, many ¯ames in burners are

stabilized by the recirculation of burnt gases, leading to

stabilization mechanisms controlled by the mixing between

fuel, oxidizer, and burnt gases. The mixtures feeding the

reaction zone are then not always pure fuel and pure

oxidizer. There are situations where partial premixing is

clearly important:

² Auto-ignition in a non-homogeneous distribution of fuel

and oxidizer, where the reactants can be mixed before

auto-ignition occurs.

² Laminar or turbulent ¯ame stabilization, when combus-

tion does not start at the very ®rst interface between fuel

and oxidizer in the vicinity of burner exit, so that fuel and

oxidizer may mix without burning.

² After quenching of the reaction zone, the reactants may

mix leading to possibility of re-ignition and combustion

in a partially premixed regime [38].

The triple ¯ame is an interesting model problem to

approach partially premixed combustion. In a laminar

shear layer where the mixing between cold fuel and oxidizer

develops, a diffusion ¯ame may be stabilized at the splitter

plate by the combination of heat losses with viscous ¯ow

effects, or, further downstream [39]. In this latter case,

combustion starts in a region where fuel and oxidizer have

been mixed in stoichiometric proportion. The resulting

premixed kernel tends to propagate towards fresh gases

and contributes to the stabilization of the trailing diffusion

¯ame. In a mixing layer con®guration, the stoichiometric

premixed kernel evolves to a rich partially premixed ¯ame

in the direction of the fuel stream, while a lean partially

premixed ¯ame develops on the air side (Fig. 7). These

two premixed ¯ames are curved because their respective

propagation velocities decrease when moving away from

the stoichiometric condition. The overall structure,

composed of two premixed ¯ames and of a diffusion

¯ame, is usually called `triple ¯ame'. Such triple ¯ames

have been ®rstly experimentally observed by Phillips [40].

Since this pioneer work, more recent experiments have

con®rmed the existence of triple ¯ames in laminar ¯ows

[41±43]. Theoretical studies [44±48] and numerical simula-

tions [49±53] have been devoted to triple ¯ames. The propa-

gation speed of triple ¯ames is controlled by two

parameters: the curvature of the partially premixed front,

increasing with the scalar dissipation rate imposed in front

of the ¯ame, and the amount of heat release. The effect of

heat release is to de¯ect the ¯ow upstream of the triple

¯ame, making the triple ¯ame speed greater than the propa-

gation speed of a planar stoichiometric ¯ame. This de¯ec-

tion also induces a decrease of the mixture fraction gradient

in the trailing diffusion ¯ame. The triple ¯ame velocity

decreases when increasing the scalar dissipation rate at the

¯ame tip. Triple ¯ame velocity response to variations of

scalar dissipation rate may be derived by approximating

the ¯ame tip by a parabolic pro®le and using results from

expansions in parabolic-cylinder coordinates. This analysis

was used by Ghosal and Vervisch to include small but ®nite

heat release and gas expansion, the triple ¯ame velocity U

TF

may be written [47]:

U

TF

< S

L

(1 1a) 2

b

Z

st

(1 1a)

4n

F

22

_

l

rC

p

x

st

_

(54)

where a = (T

burnt

2T

fresh

)=T

burnt

is de®ned from the

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 205

Fig. 6. Generic response of the heat released by a one-dimensional strained diffusion ¯ame versus DamkhoÈler number. The dashed line denotes

in®nitely fast chemistry. Da

p

q

and Da

p

i

are the critical values of Da

p

= (t

c

x

st

)

21

at quenching and ignition, respectively. t

c

is a given chemical

time and x

st

= Du7Zu

2

Z=Z

st

is the scalar dissipation rate under stoichiometric conditions.

temperatures on both sides of a stoichiometric premixed

¯ame for the same mixture, b = a(T

A

=T

burnt

) is the Zeldo-

vitch number [14], T

A

is the activation temperature, n

F

the stoichiometric coef®cient of the fuel and x

st

is measured

far upstream in the mixing layer where the triple ¯ame

propagates. The value of the scalar dissipation rate at the

triple point is of the order of x

st

=(1 1a)

2

[47]. These re-

lations are valid for small values of a and moderate, but

non-zero, values of x

st

. The triple ¯ame velocity given by

Eq. (54) may be combined with Landau±Squire solution for

non-reacting laminar round jet to construct a stability

diagram for lift-off and blowout of jet laminar diffusion

¯ames [54].

A variety of studies suggest that ®nite rate chemistry and

quenching in non-premixed combustion are somehow

linked to partially premixed combustion [55].

4. A direct analysis: Taylor's expansion

A direct approach to describe turbulent combustion is ®rst

discussed in this section. This simple formalism, based on

series expansion, illustrates the dif®culties arising from the

non-linear character of chemical sources.

Consider a simple irreversible reaction between fuel (F)

and oxidizer (O):

F 1sO - (1 1s)P

where the fuel mass reaction rate _ v

F

is expressed from the

Arrhenius law as:

_ v

F

= 2Ar

2

T

b

Y

F

Y

O

exp 2

T

A

T

_ _

(55)

where A is the pre-exponential constant, and T

A

is the acti-

vation temperature.

As the reaction rate is highly non-linear, the averaged

reaction rate

_ v

F

cannot be easily expressed as a function

of the mean mass fractions

~

Y

F

and

~

Y

O

the mean density r

and the mean temperature

~

T: The ®rst simple idea is to

expand the mean reaction rate

_ v

F

as a Taylor series:

exp 2

T

A

~

T

_ _

= exp 2

T

A

~

T

_ _

1 1

11

n=1

P

n

T

//n

~

T

n

_ _

;

T

b

=

~

T

b

1 1

11

n=1

Q

n

T

//n

~

T

n

_ _

(56)

where P

n

and Q

n

are given by:

P

n

=

n

k=1

(21)

n2k

(n 21)!

(n 2k)![(k 21)!]

2

k

T

A

~

T

_ _

k

;

Q

n

=

b(b 11)¼(b 1n 21)

n!

(57)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 206

Fig. 7. Schematic of a freely propagation triple ¯ame.

The mean reaction rate,

_ v

F

becomes [56]:

_ v

F

= 2A r

2

~

T

b

~

Y

F

~

Y

O

exp

_

2

T

A

~

T

_

×

_

1 1

¯

Y

//

F

Y

//

O

~

Y

F

~

Y

O

1(P

1

1Q

1

)

_

¯

Y

//

F

T

//

~

Y

F

~

T

1

¯

Y

//

O

T

//

~

Y

O

~

T

_

1(P

2

1Q

2

1P

1

Q

1

)

_

¯

Y

//

F

T

//

2

~

Y

F

~

T

2

1

¯

Y

//

O

T

//

2

~

Y

O

~

T

2

_

1

¼

_

(58)

Eq. (58) leads to various dif®culties. First, new quantities

such as

¯

Y

//

k

T

//n

have to be closed using algebraic expressions

or transport equations. Because of non-linearities, large

errors exist when only few terms of the series expansion

are retained. Expression (58) is quite complicated, but is

only valid for a simple irreversible reaction and cannot be

easily extended to realistic chemical schemes (at least 9

species and 19 reactions for hydrogen combustion, several

hundred species and several thousand reactions for hydro-

carbon combustion, etc.). For these reasons, reaction rate

closures in turbulent combustion are not based on Eq. (58).

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 207

Fig. 8. Turbulent premixed combustion regimes as identi®ed by Borghi and Destriau [62]: (a) ¯amelet (thin wrinkled ¯ame), (b) thickened

wrinkled ¯ame regime, and (c) thickened ¯ame regime.

Models are rather derived from physical analysis as discuss

below.

Nevertheless, this approach is used in some simulations

of supersonic reacting ¯ows [57] or to describe reaction in

atmospheric boundary layer where the temperature T may

be roughly assumed to be constant [58]. In these situations,

only the ®rst two terms in the series expansion are kept.

A segregation factor, a

FO

, is then introduced:

a

FO

= 2

¯

Y

//

F

Y

//

O

~

Y

F

~

Y

O

= 2 1 2

¯

Y

F

Y

O

~

Y

F

~

Y

O

_ _

(59)

to characterize the mixing between the reactants F and O. If

they are perfectly separated

¯

Y

F

Y

O

= 0 and a

FO

= 21: On

the other hand, a perfect mixing (

¯

Y

//

F

Y

//

O

= 0) leads to a

FO

=

0: This segregation factor may be either postulated or

provided by a balance equation (see Ref. [59] in a large

eddy simulation context). Then, the mean reaction rate

becomes:

_ v = 2A(1 1a

FO

) r

2

~

T

b

~

Y

F

~

Y

O

exp 2

T

A

~

T

_ _

(60)

5. Scales and diagrams for turbulent combustion

5.1. Introduction

As the mean burning rate

_ v cannot be found from an

averaging of Arrhenius laws, a physical approach is required

to derive models for turbulent combustion. Turbulent

combustion involves various lengths, velocity and time

scales describing turbulent ¯ow ®eld and chemical reac-

tions. The physical analysis is mainly based on comparison

between these scales.

The turbulent ¯ow is characterized by a Reynolds number

comparing turbulent transport to viscous forces:

Re =

u

/

l

t

v

(61)

where u

/

is the velocity rms (related to the square root of the

turbulent kinetic energy k), l

t

is the turbulence integral

length scale and n the kinematic viscosity of the ¯ow.

The DamkoÈhler number compares the turbulent (t

t

) and

the chemical (t

c

) time scales:

Da =

t

t

t

c

(62)

In the limit of high DamkoÈhler numbers (Da q1); the

chemical time is short compared to the turbulent one, corre-

sponding to a thin reaction zone distorted and convected by

the ¯ow ®eld. The internal structure of the ¯ame is not

strongly affected by turbulence and may be described as a

laminar ¯ame element called a `¯amelet'. The turbulent

structures wrinkle and strain the ¯ame surface. On the

other hand, a low DamkoÈhler number (Da p1) corresponds

to a slow chemical reaction. Reactants and products are

mixed by turbulent structures before reaction. In this

perfectly stirred reactor limit, the mean reaction rate may

be expressed from Arrhenius laws using mean mass frac-

tions and temperature, corresponding to the ®rst term of the

Taylor's expansion (58).

In turbulent ¯ames, as long as quenching does not occur,

most practical situations correspond to high or medium

values of the DamkoÈhler numbers. It is worth noting that

various chemical time scales may be encountered: fuel

oxidation generally corresponds to short chemical time

scales (Da q1) whereas pollutant production or destruc-

tion such as CO oxidation or NO formation are slower.

5.2. Turbulent premixed combustion diagram

5.2.1. Introduction

The objective is to analyze premixed turbulent combus-

tion regimes by comparing turbulence and chemical char-

acteristic length and time scales. This analysis leads to

combustion diagrams where various regimes are presented

as function of various dimensionless numbers [14,23,29,60±

62]. These diagrams could be a support to the selection and

development of the relevant combustion model for a given

situation. A formalism combining recent analysis [29,62] is

retained here.

For turbulent premixed ¯ames, the chemical time scale,

t

c

, may be estimated as the ratio of the thickness d

l

and the

propagation speed S

L

of the laminar ¯ame.

1

Estimating the

turbulent time from turbulent integral scale characteristics

(t

t

= l

t

=u

/

); the DamkoÈhler number becomes:

Da =

t

t

t

c

=

l

t

d

l

S

L

u

/

(63)

where a velocity ratio (u

/

/S

L

) and a length scale ratio (l

t

/d

l

)

are evidenced.

5.2.2. Combustion regimes

For large values of the DamkoÈhler number (Da q1); the

¯ame front is thin and its inner structure is not affected by

turbulence motions which only wrinkle the ¯ame surface.

This ¯amelet regime or thin wrinkled ¯ame regime (Fig. 8a)

occurs when the smallest turbulence scales (i.e. the Kolmo-

gorov scales), have a turbulent time t

k

larger than t

c

(turbulent motions are too slow to affect the ¯ame structure).

This transition is described in term of the Karlovitz

number Ka:

Ka =

t

c

t

k

=

d

l

l

k

u

k

S

L

(64)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 208

1

This chemical time t

c

corresponds to the time required for the

¯ame to propagate over a distance equal to its own thickness. This

time may also be viewed as a diffusive time scale, using Eq. (34)

t

c

= d

l

=S

L

= (1=Re

f

)(d

2

l

=n):

The size l

k

and the velocity u

k

of Kolmogorov structures

are given by [64]:

l

k

=

v

3

1

_ _

1=4

; u

k

= (v1)

1=4

(65)

where 1 is the dissipation rate of the turbulent kinetic energy

k. The integral length scale lt may be written:

l

t

=

u

/3

1

_ _

(66)

using v = d

l

S

L

; corresponding to a unity ¯ame Reynolds

number Re

f

(Eq. (34)), yields

Ka =

u

/

S

L

_ _

3=2

l

t

d

l

_ _

21=2

(67)

Reynolds, Re, DamkoÈhler, Da, and Karlovitz, Ka, numbers

are related as:

Re = Da

2

Ka

2

(68)

and a set of two parameters (Re, Da), (Re, Ka) or (Da, Ka) are

necessary to discuss regimes in the case of premixed reactants.

The Karlovitz number also compares the ¯ame and the

Kolmogorov length scales according to:

Ka =

d

l

l

k

_ _

2

(69)

The Karlovitz number is used to de®ne the Klimov±

Williams criterion, corresponding to Ka = 1; delineating

between two combustion regimes. This criterion was ®rst

interpreted as the transition between the ¯amelet regime

(Ka , 1); previously described, and the distributed combus-

tion regime where the ¯ame inner structure is strongly modi-

®ed by turbulence motions. A recent analysis [29] has

shown that, for Karlovitz numbers larger than unity (Ka .

1); turbulent motions become able to affect the ¯ame inner

structure but not necessarily the reaction zone. This reaction

zone, where heat is released, has a thickness d

r

much lower

that the thermal thickness d

l

of the ¯ame (d

r

< 0:1d

l

): The

Karlovitz number based on this reaction thickness is:

Ka

r

=

d

r

l

k

_ _

2

=

d

r

d

l

_ _

2

d

l

l

k

_ _

2

<

1

100

d

l

l

k

_ _

2

<

Ka

100

(70)

Then, the following turbulent premixed ¯ame regimes are

proposed [29]:

² Ka ,1: Flamelet regime or thin wrinkled ¯ame regime

(Fig. 8a). Two subdivisions may be proposed depending

on the velocity ratio u

/

/S

L

:

Ð(u

/

/S

L

) ,1: wrinkled ¯ame. As u

/

may be viewed as

the rotation speed of the larger turbulent motions,

turbulent structures are unable to wrinkle the ¯ame

surface up to ¯ame front interactions. The laminar

propagation is predominant and turbulence/combus-

tion interactions remain limited.

Ð(u

/

/S

L

) .1: wrinkled ¯ame with pockets (`corrugated

¯ames'). In this situation, larger structures become able

to induce ¯ame front interactions leading to pockets.

² 1 , Ka # 100 (Ka

r

, 1) : Thickened wrinkled ¯ame

regime or thin reaction zone. In this case, turbulent

motions are able to affect and to thicken the ¯ame preheat

zone, but cannot modify the reaction zone which remains

thin and close to a laminar reaction zone (Fig. 8b).

² Ka .100 (Ka

r

.1): Thickened ¯ame regime or well-

stirred reactor. In this situation, preheat and reaction

zones are strongly affected by turbulent motions and no

laminar ¯ame structure may be identi®ed (Fig. 8c).

These various regimes are generally displayed on a log-

arithmic diagram (u

/

/S

L

; l

t

/d

l

), similar to the one presented

on Fig. 9.

5.2.3. Comments

This analysis, leading to a rough classi®cation of combus-

tion regimes as a function of characteristic numbers, has

been developed as a support to derive and choose turbulent

combustion models. Following this classi®cation, most

practical applications correspond to ¯amelet or thickened

wrinkled ¯ame regimes. Nevertheless, such analyses are

only qualitative and should be used with great care. A

diagram such as the one displayed on Fig. 9 cannot be

readily used to determine the combustion regime of a prac-

tical system from (u

/

/S

L

) and (d

l

/l

t

) ratios:

² The analysis is based on the assumption of a homo-

geneous and isotropic turbulence unaffected by heat

release, which is not the case in combustion systems.

² Some quantities used are not clearly de®ned. For

example, the ¯ame thickness d

l

may be based on the

thermal thickness or on the diffusive thickness. Accord-

ingly, the limits between the various regimes may notice-

ably change.

² All regime limits are based on order of magnitude esti-

mations and not on precise derivations. For example, the

¯amelet regime limit could correspond to a Karlovitz

number Ka = 0:1 or Ka = 10; rather than Ka = 1:

² Various effects are not taken into account. Unsteady and

curvature effects play an important role neglected here.

Turbulent premixed combustion diagrams were analyzed

using direct numerical simulations of ¯ame/vortex inter-

actions [65]. Results show that the ¯amelet regime seems

to extend above the Klimov±Williams criterion (see Fig.

9). DNS has revealed that small turbulent scales, which

are supposed in classical theories to have the strongest

effects on ¯ames, have small lifetimes because of viscous

dissipation and therefore only limited effects on combus-

tion, results recovered experimentally [66]. Peters [29]

shows that the criterion Ka = 100 (i.e. Ka

r

= 1) is in

quite good agreement with the transition proposed in

Ref. [65], at least when the length scale ratio, l

t

/d

l

, is

suf®ciently large.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 209

² Additional length scales have been introduced in the

literature. For instance the Gibson scale l

G

, to character-

ize the size of the smaller vortex able to affect the ¯ame

front was used [63]. This length was de®ned as the size of

the vortex having the same velocity as the laminar ¯ame

speed S

L

.

² All these analyses are implicitly based on a single step

irreversible reaction. In actual turbulent combustion, a

large number of chemical species and reactions are

involved (several hundred species and several thousand

reactions for propane burning in air). These reactions

may correspond to a large range of chemical time scales.

For example, the propane oxidation may assumed to be

fast compared to turbulent time scale. On the other hand,

the CO

2

formation from carbon monoxide (CO) and OH

radical in the burnt gases is much slower with a chemical

time of the same order as turbulent times.

5.3. Non-premixed turbulent combustion diagram

5.3.1. Introduction

Two numbers, a length ratio and a velocity ratio, have

been used to identify premixed turbulent combustion

regimes. The problem is more dif®cult in non-premixed

turbulent combustion because diffusion ¯ames do not propa-

gate and, therefore, exhibit no intrinsic characteristic speed.

In addition, the thickness of the ¯ame depends on the aero-

dynamics controlling the thickness of the local mixing

layers developing between fuel and oxidizer (Section 3.2)

and no ®xed reference length scale can be easily identi®ed

for diffusion ¯ames. This dif®culty is well illustrated in the

literature, where various characteristic scales have been

retained depending on the authors [33,67±71]. These

classi®cations of non-premixed turbulent ¯ames may be

organized in three major groups:

² The turbulent ¯ow regime is characterized by a Reynolds

number, whereas a DamkoÈhler number is chosen for the

reaction zone [72].

² The mixture fraction ®eld is retained to describe the

turbulent mixing using

¯

Z

//2

and a DamkoÈhler number

(ratio of Kolmogorov to chemical time) characterizes

the ¯ame [33].

² A velocity ratio (turbulence intensity to premixed lami-

nar ¯ame speed) and a length ratio (integral scale to

premixed laminar ¯ame thickness) may be constructed

[68] to delineate between regimes.

Additional lengths have also been introduced, using for

instance thicknesses of pro®les in mixture fraction space

[67].

A laminar diffusion ¯ame is fully determined from a

DamkoÈhler number Da

p

= (t

c

x

st

)

21

; where the value of

the chemical time t

c

depends on the fuel chemistry [30]

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 210

Fig. 9. Turbulent premixed combustion diagram [29,62]. Combustion regimes are identi®ed using the length scale (l

t

=d

l

) and the velocity

(u

/

=S

L

) ratios. The Klimov±Williams criterion (Ka = 1) corresponds to a ¯ame thickness d

l

equal to the Kolmogorov scale l

k

. Below this line,

the ¯ame is thinner than any turbulent scale. Below the line delineating the Peters criterion (Ka = 100 or Ka

r

= 1), the reaction zone thickness,

d

r

, is thinner than any turbulent scale and is not affected by turbulent motions (the criterion is plotted assuming d

r

< 0:1d

l

). The ¯amelet regime

limit devised by Poinsot et al. [65] from direct numerical simulations is also displayed. The criterion proposed in Ref. [140] to delineate

between gradient (above) and counter-gradient (below) turbulent transport is displayed assuming a heat release factor t = T

b

=T

u

21 = 6 where

T

u

and T

b

are, respectively, the fresh and the burnt gases temperature (see Section 8).

(Section 3.2). In this number, the scalar dissipation rate

under stoichiometric condition (Z = Z

st

); x

st

= Du7Zu

2

st

;

measures at the same time a mechanical time, t

x

= x

21

st

;

and, a characteristic mixing length, l

d

= (D=x

st

)

1=2

: Accord-

ing to asymptotic developments [30], the reaction zone

thickness is of the order of l

r

< l

d

(Da

p

)

21=(a11)

; where a is

the order of a global one-step reaction.

Because diffusion ¯ames do not feature a ®xed reference

length, a main dif®culty arises when effects of unsteadiness

need to be quanti®ed. In a steady laminar ¯ame the local rate

of strain is directly related to x

st

(and to a ¯ame thickness),

however, when the velocity ®eld ¯uctuates, unsteadiness in

diffusion ¯ames develops at two levels [73]:

² The mixture fraction ®eld Z does not immediately

respond to velocity ¯uctuations, leading to a distribution

of x

st

for given rates of strain. Because a strong correla-

tion exists between x

st

and velocity gradients taken along

the stoichiometric line [74], this effect is not the dominant

one when ®nite rate chemistry occurs.

² For ®nite rate chemistry, the burning rate does not im-

mediately follow variations of x

st

, leading to a second

level of unsteadiness, modifying the burning rate (Eq.

(71)).

u

/

- [Unsteadiness in mixing] - x

s

- [Burning rate unsteadiness(for Da

p

, 1)] - _ v

i

(71)

Summarizing these effects in a generic diagram is an

arduous task.

A diagram for laminar ¯ames submitted to curvature

associated to a time varying strain rate was obtained by

Cuenot and Poinsot from DNS results of ¯ame/vortex inter-

action [70]. In this diagram presented on Fig. 10, the ¯ame

thickness is d

i

<l

d

, whereas r and u

/

denote respectively the

characteristic size and velocity of the vortex pair. This

analysis identi®es two limiting DamkoÈhler numbers,

Da

LFA

and Da

ext

. When Da

p

is larger than Da

LFA

, the ¯ame

front may be viewed as a steady laminar ¯ame element and

its inner structure is not affected by vortices. On the other

hand, when Da

p

# Da

ext

; ¯ame extinction occurs. In the

intermediate DamkoÈhler number range (i.e. Da

ext

,Da

p

,

Da

LFA

), strong unsteadiness effects are observed.

In a non-premixed turbulent ¯ame, the reaction zones

develop within a mean mixing zone whose thickness l

z

is

of the order of the turbulent integral length scale l

t

(Fig. 11):

l

z

< u7

~

Zu

21

< l

t

<

k

3=2

1

_ _

(72)

Turbulent small scale mixing mainly depends both on

velocity ¯uctuations, transporting the iso-Z surfaces (stir-

ring), and diffusion between these iso-surfaces that compose

the mixing layer of thickness l

d

, with

l

d

<

D

~ x

st

_ _

1=2

(73)

where ~ x

st

denotes the conditional mean value of the scalar

dissipation rate x for Z = Z

st

:

When transport of species and heat by velocity ¯uctua-

tions is faster than transfer in the diffusion ¯ame, a departure

from laminar ¯amelet is expected. Also, when the Kolmo-

gorov scale l

k

is of the order of the ¯ame thickness, the inner

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 211

Fig. 10. Non-premixed ¯ame/vortex interaction regimes by Cuenot and Poinsot [70]. This diagram delineates the steady laminar ¯amelet

assumption (LFA) validity regions, the quenching limits and the zone where unsteady and curvature effects are important during ¯ame vortex

interaction. u

/

is the level of velocity ¯uctuations, d

i

= l

d

is the ¯ame thickness (l

d

< u7Zu

21

), t is a chemical time and r the characteristic size

of the vortices.

structure of the reaction zone may be modi®ed by the turbu-

lence. As diffusion ¯ame scales strongly depend on the local

¯ow motions, one may write:

l

d

< a

1

l

k

and ~ x

st

<

a

2

t

k

(74)

where a

1

$ 1 and a

2

# 1 (the maximum local strain rate

would correspond to l

d

= l

k

).

Then using t

t

=t

k

=

Re

_

; the DamkoÈhler number com-

paring turbulent ¯ame scale and chemical ¯ame scale is

recast as:

Da =

t

t

t

c

=

t

t

t

k

t

k

t

c

<

t

t

t

k

a

2

~ x

st

t

c

< a

2

Re Da

p

_

(75)

Constant DamkoÈhler numbers Da

p

correspond to lines of

slope 1/2 in a log±log (Da, Re) diagram. When the chemistry

is suf®ciently fast (large Da values), the ¯ame is expected to

have a laminar ¯ame structure. This condition may be

simply expressed as Da

p

$ Da

LFA

On the other hand, for

large chemical times (i.e. when Da

p

# Da

ext

), extinction

occurs. Laminar ¯ames are encountered for low Reynolds

numbers (Re , 1): Results are summarized in Fig. 12.

In a practical combustion devices, a

1

and a

2

would

evolve in space and time according to ¯ow ¯uctuations,

velocity and scalar energetic spectra. In a given burner, it

is likely that one may observe at different locations ¯amelet

behavior and strong unsteadiness, or even quenching.

As the classi®cation of premixed turbulent ¯ames, these

considerations are limited by the numerous hypothesis

necessary to derive the regimes.

6. Tools for turbulent combustion modeling

6.1. Introduction

The mean heat release rate is one of the main quantities of

practical interest that should be approximated by turbulent

combustion models. The simplest and most direct approach

is to develop the chemical rate in Taylor series as a function

of species mass fractions and temperature (Eq. (58)). This

analysis is limited by its low accuracy and by the rapidly

growing complexity of the chemistry (Section 4). It is then

concluded that the non-linear character of the problem

requires the introduction of new tools. These new tools

must be designed to describe turbulent ¯ames and have to

provide an estimation of mean production or consumption

rates of chemical species. They also need to be based on

known quantities (mean ¯ow characteristics, for example)

or on quantities that may be easily modeled or obtained from

closed balance equations. In this section, a generic descrip-

tion of the main concepts used to model turbulent combus-

tion is proposed. Relations between the various approaches

are also emphasized, but the discussion of the closure

strategy is postponed to subsequent sections.

The basic ingredients to describe turbulent ¯ames remain

the quantities introduced for laminar ¯ame analysis: the

progress variable c for premixed combustion (c = 0 in

fresh gases and c = 1 in burnt gases, see Section 3.2),

and, the mixture fraction Z for non-premixed ¯ames (Z is

a passive scalar, with Z = 0 in pure oxidizer and Z = 1 in

pure fuel, see Section 3.2). The ¯ame position would corre-

spond to values of the progress variable c lying between 0

and 1, or, to Z taking on values in the vicinity of Z = Z

st

:

Three main types of approaches are summarized in

Fig. 13:

The burning rate may be quanti®ed in terms of turbulent

mixing. When the DamkoÈhler number Da = t

t

=t

c

; com-

paring turbulent (t

t

) and chemical (t

c

) characteristic times,

is large (a common assumption in combustion modeling),

the reaction rate is limited by turbulent mixing, described in

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 212

Fig. 12. Schematic of non-premixed turbulent combustion regimes

as function of the DamkoÈhler number Da = t

t

=t

c

(constructed from

the turbulent integral time scale t

t

and chemical time t

c

) and Re the

turbulent Reynolds number.

Fig. 11. Sketch of a non-premixed turbulent ¯ame. Z is the mixture

fraction, l

d

the diffusive thickness, l

r

the reaction zone thickness, l

t

the turbulence integral length scale.

terms of scalar dissipation rates [75]. The small scale dissi-

pation rate of species controls the mixing of the reactants

and, accordingly, plays a dominant role in combustion

modeling, even for ®nite rate chemistry.

In the geometrical analysis, the ¯ame is described as a

geometrical surface, this approach is usually linked to a

¯amelet assumption (the ¯ame is thin compared to all

¯ow scales). Following this view, scalar ®elds (c or Z) are

studied in terms of dynamics and physical properties of iso-

value surfaces de®ned as ¯ame surfaces (iso-c

p

or iso-Z

st

).

The ¯ame is then envisioned as an interface between fuel

and oxidizer (non-premixed) or between fresh and burnt

gases (premixed). A ¯ame normal analysis is derived by

focusing the attention on the structure of the reacting ¯ow

along the normal to the ¯ame surface. This leads to ¯amelet

modeling when this structure is compared to one-dimen-

sional laminar ¯ames. The density of ¯ame surface area

per unit volume is also useful to estimate the burning rate.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 213

Fig. 13. Three types of analysis of premixed or non-premixed turbulent ¯ames.

The statistical properties of scalar ®elds may be collected

and analyzed for any location within the ¯ow. Mean values

and correlations are then extracted via the knowledge of

one-point probability density functions (pdf). The determi-

nation of these pdfs leads to pdf modeling. A one-point

statistical analysis restricted to a particular value of the

scalar ®eld is related to the study of conditional statistics.

Conditional statistics which are obviously linked to the

geometrical analysis and to ¯ame surfaces when the con-

ditioning value is c

p

or Z

st

.

6.2. Scalar dissipation rate

In a ®rst step, the transport equation for

¯

c

//2

or

¯

Z

//2

is

derived, these ¯uctuations characterize non-homogeneities

and intermittencies. In the case of the progress variable, the

variance

¯

c

//2

is de®ned as:

r

¯

c

//2

= r (c 2 ~ c)

2

= r(

¯

c

//2

2 ~ c

2

) = rc

2

2 r ~ c

2

(76)

Starting from the balance equation for the progress vari-

able (Eq. (36)), c is decomposed into c = ~ c 1c

//

; then the

new equation is multiplied by c

//

and averaged. After

straightforward manipulations, the exact transport equation

for

¯

c

//2

reads:

2 r

¯

c

//2

2t

17( r ~ u

¯

c

// 2

) 17( r

¯

u

//

c

// 2

) = (77)

7´(rD7c

// 2

) 12c

//

7´(rD7~ c)

22 r

¯

u

//

c

//

´7~ c

.....,,.....

Production

22rD7c

//

´7c

//

.....,,.....

Dissipation

1 2 _ vc

//

..,,..

Source

In addition to the two diffusive terms 7´(rD7c

// 2

) and

2c

//

7´(rD7~ c); which are non-zero, but expected small for

large Reynolds number ¯ows (especially the second one),

two important terms are found: The ¯uctuating part of the

scalar dissipation rate 2rD7c

//

´7c

//

and a correlation _ vc

//

involving the chemical source.

In the literature, various expressions have been associated

to the terminology scalar dissipation rate (in laminar ¯ame

theory, it actually quanti®es a diffusion speed, Section 3.2).

It may include the density r, a factor 2 and be written in

term of instantaneous (c) or ¯uctuating (c

//

) values of the

concentration species. Hereafter:

r ~ x = rD7c´7c = rD7~ c´7~ c 12rD7c

//

´7~ c 1rD7c

//

´7c

//

leading to, when mean gradients are neglected:

r ~ x < rD7c

//

´7c

//

(78)

Then, r ~ x is the dissipation rate of the ¯uctuations of the

scalar ®eld.

In the simpli®ed case of homogeneous ¯ames (no ~ c or

~

Z

gradient), the time evolution of the scalar variances are

governed by:

Premixed combustion:

2 r

¯

c

// 2

dt

= 22rD7c

//

´7c

//

12 _ vc

//

Non-premixed combustion:

2 r

¯

Z

// 2

dt

= 22rD7Z

//

´7Z

//

These equations have important implications:

The scalar dissipation rate directly measures the decaying

speed of ¯uctuations via turbulent micromixing. Since the

burning rate depends on the contact between the reactants,

in any models, the scalar dissipation rate enters directly or

indirectly the expression for the mean burning rate. For

instance, when assuming very fast chemistry and a combus-

tion limited by mixing, the mean burning rate is proportional

to the scalar dissipation rate of Z or c.

Within a premixed system, turbulent mixing occurs

between fresh and burnt gases. One may then expect a

very strong coupling between mixing phenomena and

chemical reaction. This is observed in the equation for

¯

c

// 2

where, at the same time, ~ x and the chemical source _ vc

//

are

involved.

In a non-premixed ¯ame, fresh fuel and fresh oxidizer

have to be mixed at the molecular level for reacting

and the ¯ame is mainly controlled by turbulent mixing

occurring between the fresh gases. In consequence, there

is no chemical source acting on the evolution of

¯

Z

//2

:

The mixture fraction Z is sensitive to chemistry only via

density change, making the coupling between chemistry

and mixing different than in the case of premixed

combustion.

This preliminary analysis shows that dissipation rate of

scalars is a very key concept of turbulent combustion and,

directly or indirectly, x appears in any tools used to model

¯ames. The main stumbling block in turbulent combustion

modeling and bridges between the various modeling

concepts emerge through the scalar dissipation rate.

6.3. Geometrical description

The ¯ame front is here described as a geometrical entity.

This analysis is generally linked to the assumption of a

suf®ciently thin ¯ame, viewed as an interface between

fresh and burnt gases in premixed combustion or as an inter-

face between fuel and oxidizer in non-premixed situations.

Two formalisms have been proposed: ®eld equation or ¯ame

surface density concept.

6.3.1. G-®eld equation

The balance equation for the progress variable may

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 214

be written:

2c

2t

1u´7c =

1

r

(7´(rD7c) 1 _ v) = wu7cu

or

2c

2t

1u

c

´7c = 0

(79)

where u

c

is the absolute propagation velocity of the progress

variable ®eld. This velocity is easily decomposed into the

¯ow velocity u and w the relative velocity of the iso-c

surface measured with respect to the ¯ow, leading to u

c

=

u 1wn; where n = 2(7c=u7cu) is the local normal vector to

the iso-c surface, pointing toward the fresh gases. The rela-

tive displacement speed w is given by Eq. (37).

In laminar ¯ames, a G-®eld whose level G = G

0

repre-

sents the ¯ame surface, was introduced to simulate the

propagation of fronts [76]. The `G-equation' may be written:

2G

2t

1u

c

´7G = 0 or

2G

2t

1u´7G = wu7Gu (80)

This kinematic description of premixed combustion

possesses some attractive aspects:

² The internal ¯ame front structure does not need to be

resolved on the computational mesh. Only the G-®eld,

generally much thicker than the ¯ame front, needs to be

resolved. Usually, G(x) is de®ned as the distance of the

given location x to the ¯ame front [77].

² Under the assumption of constant density (thermodiffu-

sive assumption), the G-equation may be used for low

cost direct numerical simulations. Each G iso-surface is

then related to a ¯ame front, and one may argue that a

single simulation corresponds to the computation of

several ¯ames.

In non-constant density ¯ows where thermal heat release

is included, the displacement speed w of the G-iso-surface is

affected by thermal expansion and should be corrected for

density variations, even in the case of a steady laminar ¯ame

propagating at the constant laminar ¯ame speed S

L

, de®ned

relatively to the unburnt gases. This correction is:

w =

r

u

r

w

u

(81)

where w

u

is the ¯ame displacement speed relative to the

unburnt gases of density r

u

.

A more dif®cult point is the coupling required between

the G-equation and the species or energy balance equations.

The G-equation provides a kinematic description of the

¯ame front and involves its displacement speed w. The

reactant consumption and the heat release rate are controlled

by the consumption speed S

c

:

r

u

S

c

=

_1

21

_ vdj (82)

where r

u

is the fresh gases density and j the spatial coordi-

nate along the normal direction to the ¯ame front. Of course,

w and S

c

are related but may be quite different, especially in

high ¯ame front curvature zones [78]. The displacement

speed w may also be quite different from the laminar

¯ame speed S

L

.

The coupling between the consumption speed S

c

and the

displacement speed w is a very key point in G-®eld model-

ing. Three approaches have been proposed to overcome this

dif®culty:

² Flame front tracking technique: The displacement of the

¯ame front is evaluated from the displacement speed w,

leading to an estimation of the volume of burnt gases

produced along with the thermal heat release [79,80].

Based on a purely geometric approach, this technique is

well suited to two dimensional simulations, but its exten-

sion to 3D cases may not be straightforward.

² Temperature (or energy) reconstruction: The temperature

®eld is directly estimated from the G-®eld as [81]:

T = T

u

1

Q

C

p

H(G 2G

p

) (83)

where T

u

is the temperature of unburnt gases, Q the reac-

tion heat release and H denotes the Heaviside function

(the Heaviside function is smoothed on the mesh of the

simulation). This approach does not require a balance

equation for the energy but is not applicable when heat

losses or compressibility effects (for example in an in-

ternal combustion engines) occur.

² Estimation of the heat release rate from the G-®eld: The

G-®eld is used to estimate the heat release rate to be

incorporated in the balance energy equation from a

formulation similar to Eq. (83) [82]. Accordingly, any

other effects (heat losses, compressibility) may be

included.

This formalism may be extended to turbulent ¯ames.

Averaging the progress variable Eq. (79) leads to:

2~ c

2t

1 ~ u´7~ c =

ru

r

27´( r

¯

u

//

c

//

) 1rwu7cu

u7~ cu

_

_

............,,............

S

T

u7~ cu (84)

where the turbulent ¯ame speed, S

T

, has to be modeled. The

G-equation becomes:

2

~

G

2t

1 ~ u´7

~

G =

ru

r

S

T

u7

~

Gu

The mean turbulent ¯ame brush is then located at the

points where

~

G = G

0

: The G-equation does not required

thin ¯ame elements per se. The overall turbulent ¯ame is

only viewed as a propagating surface without solving for the

internal ¯ame structure. This formalism is therefore a good

candidate for the numerical simulation of large systems

where the knowledge of the internal structure of the ¯ame

brush is not required [79]. Nevertheless, a model has to be

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 215

supplied for S

T

, when doing so, turbulent transport and

propagation may be separated to carefully model the effect

of turbulence on w, the instantaneous displacement speed of

the premixed front [29,33].

6.3.2. Flame surface density description

The ¯ame is identi®ed as a surface and the ¯ame surface

density S is introduced, S measures the available ¯ame area

d A per unit volume d V. The mean burning rate of a species

i is then modeled as:

_ v

i

=

_

V

i

S (85)

where V

Ç

i

is the mean local burning rate per unit of ¯ame area

integrated along the normal direction to the ¯ame surface.

V

Ç

i

is related to the properties of the local ¯ame front and is

generally estimated from a prototype laminar ¯ame, incor-

porating more or less complexity. For instance, one may

consider a planar laminar ¯ame, submitted or not to a steady

strain, a laminar ¯ame where curvature effects have been

introduced, or even a laminar unsteady strained and curved

¯ame. The main advantage of this formulation, summarized

on Fig. 14, is to decouple the chemical description (V

Ç

i

) from

the ¯ame/turbulence interaction (S). The ¯ame surface is

convected, diffused, curved and strained by the velocity

®eld [17,83].

The ¯ame surface density S may be estimated either from

algebraic relations (see Section 7.4.2) or as a solution of a

balance equation. Using a phenomenological analysis, this

balance equation was ®rst proposed by Marble and

Broadwell [84] for non-premixed turbulent combustion.

More rigorous derivations were obtained from geo-

metrical considerations [17,85] and from a statistical

description [86±88] leading to an exact, but unclosed,

balance equation. The derivation using statistical tools is

now brie¯y summarized.

In premixed combustion, the ¯ame surface density S of

the iso-c

p

surface is estimated from the conditional gradient

of the progress variable c [86]:

S(c

p

) = u7cud(c 2c

p

) = u7cuuc = c

p

_ _

P(c

p

) (86)

where d(c 2c

p

) is a local measure of the probability (see

Section 6.4.3),

_

u7cuuc = c

p

_

is the conditional average of

u7cu for c = c

p

and

P(c

p

) is the probability to ®nd c = c

p

at

the given location. From this de®nition and the balance

equation for the progress variable c, an exact equation for

the ¯ame surface density S may be derived according to the

following steps [87]:

1. Derivation of an equation for u7cu from the equation for

the instantaneous progress variable c.

2. Derivation of an equation for u7cu d(c 2c)

p

by condi-

tioning the previous u7cu balance equation.

3. Averaging the u7cu d (c 2c)

p

balance equation leading

to an exact equation for the ¯ame surface density S =

u7cud(c 2c

p

):

This derivation is valid for any iso-scalar surface (c

p

can

take any values between zero and unity) and S(c

p

; x; t) is also

called a surface density function, the derivation of its balance

equation is quite tedious [87] and similar to the derivation of

a balance equation for the probability density function

(Section 6.4.3), details are not given here. Two equivalent

forms of the progress variable equation may be used:

² A classical reaction/diffusion formulation:

2c

2t

1u´7c =

1

r

[7´(rD7c) 1 _ v] (87)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 216

Fig. 14. Flame surface density modeling.

² A propagative form:

2c

2t

1u´7c = wu7cu (88)

where the displacement speed w of the ¯ame front rela-

tively to the ¯ow was introduced (Eq. (37)). This equa-

tion corresponds to the G-equation [77] (Section 6.3.1).

Two different types of equation for S are associated to

these two formulations of the progress variable balance

equation.

² A reaction/diffusion formulation:

2S

2t

17´(kul

s

S) = k7´u 2nn : 7ul

s

S

2

_

1

u7cu

2

2n

_

1

r

[7´(rD7c) 1 _ v]

__

s

S

2

2

2c

p

__

1

r

[7´(rD7c) 1 _ v]

_

s

S

_

(89)

² A propagative formulation:

2S

2t

17´(kul

s

S)

= k7´u 2nn : 7ul

s

S 27´ kwnl

s

S

_ _

1kw7´nl

s

S

(90)

where n is the unit vector normal to the c = c

p

surface and

pointing towards the fresh gases (n = 27c=u7cu): 7´n corre-

sponds to the ¯ame front curvature. 2=2n = n´7 is a deriva-

tive normal to the ¯ame front. 2/2c

p

is a derivative in the

sample space c

p

. The surface average of Q, kQl

s

, is de®ned

as:

kQl

s

=

Qu7cud(c 2c

p

)

u7cud(c 2c

p

)

=

_

Qu7cu c } = c

p

_

_

u7cu c } = c

p

_ (91)

Here the notation kQl

s

implicitly indicates that the mean

is taken for c = c

p

; and kQl

s

is a function of c

p

.

The propagative formulation is often written in term of

total stretch A

c

p

of the iso-c

p

surface

1

S

2S

2t

17´(ku 1wnl

s

S)

_ _

= A

c

p

= k 7´u 2nn : 7u

......,,......

tangential strain rate

1 w7´n

,....,

Curvature

l

s

(92)

The LHS terms in the two balance Eqs. (89) and (90) corre-

spond to unsteady effects and to the ¯ame surface convec-

tion. The ®rst term in the RHS expresses the action of the

tangential strain rate on the ¯ame surface. The last two terms

in Eq. (89) describe respectively reaction/diffusion effects

along the ¯ame normal direction and ¯uxes in the sample

space c

p

. The last two terms in the RHS of the propagative

Eq. (90) correspond to front convection due to a normal

propagation and combined propagation/curvature effects.

These two formulations of the S balance equation induce

the following comments:

² The two balance Eqs. (89) and (90) are mathematically

strictly equivalent but the problem is not expressed in the

same way. In the propagative form (Eq. (90)), many

effects are incorporated in the ¯ame front propagation

speed, w, that may differ from the laminar ¯ame speed

S

L

[78]. On the other hand, a ¯ux term in the sample space

is found in the reaction/diffusion derivation. In Eq. (90)

the imbalance between diffusion and reaction is cast in

the form of the propagation velocity w, Eq. (89) recalls

that transfer phenomena between iso-surfaces are

involved in this propagation.

² The derivation of the balance Eqs. (89) and (90) impli-

citly forgets some mathematical singularities that may

become important in particular situations. For example,

the normal vector n is assumed to be well de®ned and

having a ®nite derivative. This may not be the case when

two close ¯ame fronts interacts.

² In combustion modeling, a single iso-c

p

surface is

assumed to correspond to the ¯ame front. This is a priori

true when the ¯ame is in®nitely thin, assuming that the

local reaction rate per unit ¯ame area, V

Ç

i

, describes

whether the ¯ame is actually burning or not. However

in real turbulent ¯ames, the local burning zone is not

in®nitely thin and the ¯ame front, identi®ed as the

location of the maximum reaction, may differ from the

c

p

iso-surface. This may be for instance the case when

analyzing data of direct numerical simulations where the

¯ame front has to be resolved on the computational grid.

² Using the Favre decomposition (u = ~ u 1u

//

); the

convection ¯ux term may be decomposed into mean

and turbulent components:

kul

s

S = ~ uS 1ku

//

l

s

S (93)

² The strain rate term is also split into a contribution due to

the mean ¯ow and a contribution due to turbulent velo-

city ¯uctuations:

k7´u 2nn : 7ul

s

S = (7´ ~ u 2knnl

s

: 7~ ul

s

........,,........

S

A

T

1k7´u

//

2nn : 7~ u)S

........,,........

a

T

(94)

All these de®nitions may be extended to ¯ame fronts which

are not in®nitely thin. Integrating Eq. (86) across iso-surface

levels leads to:

_1

0

S(c

p

)dc

p

=

_1

0

_

u7cuuc = c

p

_

P(c

p

)dc

p

= u7cu (95)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 217

that is a direct estimation of the inverse of the mean local

¯ame thickness d (d = (1=u7cu)

21

) and may be viewed as a

`generalized ¯ame surface density', u7cu follows a balance

equation similar to the S(c

p

) balance Eq. (90), where

surface averages kQl

s

are replaced by `generalized surface

averages' kQl

s

which do not depend on c

p

:

kQl

s

=

1

u7cu

_1

0

kQl

s

S(c

p

)dc

p

(96)

In ¯ame surface density models, under a ¯amelet assump-

tion, for any values of c

p

, the ¯ame surface is assumed to

behave such as S(c

p

) < u7cu = S:

All balance equations remain formally the same but have

to be closed. The same modeling issues emerge using the G-

equation formalism [89]. One needs to develop closures for

the turbulent ¯ux of ¯ame surface, for the propagation vel-

ocity of the surface, as well as for the effects of curvature

and strain rate.

6.3.3. Flame wrinkling description

The previous formalism may be recast in term of ¯ame

surface wrinkling. The basic idea is to introduce the ratio

J(c

p

) of the ¯ame surface to its projection in the direction of

propagation n

p

:

J(c

p

) =

S(c

p

)

n

p

´(nu7cud(c 2c

p

))

=

_

u7cuuc = c

p

_

P(c

p

)

n

p

´

_

nu7cud(c 2c

p

)

_

=

_

u7cuuc = c

p

_

n

p

´

_

nu7cuuc = c

p

_ =

1

n

p

´knl

s

(97)

where n and n

p

are the unit vectors normal to the instan-

taneous ¯ame front and to the mean propagating direction

respectively. These vectors are chosen pointing towards the

fresh gases and are given by:

n = 2

7c

u7cu

; n

p

= 2

_

7cuc = c

p

_

_

7cuc = c

p

_ _

_

_

_

_

_

(98)

Then:

J(c

p

) =

_

u7cuuc = c

p

_

_

7cuc = c

p

_ _

_

_

_

_

_

=

S(c

p

)

_

7cuc = c

p

_ _

_

_

_

_

_

P(c

p

)

(99)

As in Section 6.3.2, a generalized ¯ame surface wrinkling,

J; is introduced:

J =

_1

0

S(c

p

)dc

p

_1

0

_

7cuc = c

p

_ _

_

_

_

_

_

P(c

p

)dc

p

=

u7cu

u7 cu

=

S

u7 cu

(100)

For an in®nitely thin ¯ame front, J(c

p

) =

J for any c

p

value.

Flame surface density and ¯ame wrinkling factor are

closely related. Balance equations may also be derived

and closed for J(c

p

) or

J: These equations are considerably

more complicated than ¯ame surface density balance equa-

tions, but the wrinkling factor may be more convenient for

initial and/or boundary conditions (

J $ 1 everywhere).

This approach has been explored by Weller et al. [90,91].

6.4. Statistical approaches: probability density function

6.4.1. Introduction

Predictions of radicals and intermediate species such as

OH, or pollutants like CO, require the description of the

¯ame front internal structure, for intermediate states

between fresh and burnt gases in premixed ¯ames or

between fuel and oxidizer in non-premixed ¯ames. Even

though G-®eld and density of ¯ame surface S need some

statistical treatments, they are initially based on a geo-

metrical view describing the ¯ame as a thin interface. In

probability density function methods, one wishes to relax

this assumption by focussing on the statistical properties of

intermediate states within the ¯ame front.

The probability density function (pdf)

P(Y

p

; x; t) quanti-

®es the probability to ®nd, for a given location x and a time t,

a variable Y (mass fraction, temperature, velocity, etc.)

within the range [Y

p

2DY=2; Y

p

1DY=2]: This probability

is equal to

P(Y

p

; x; t)DY [92±96]. The pdf satis®es the

following simple relations:

_

Y

P(Y

p

; x; t)dY

p

= 1

_

Y

P(Y

p

; x; t)dY

p

=

Y(x; t)

_

Y

(Y

p

2

Y)

2

P(Y

p

; x; t)dY

p

= Y

/2

(x; t)

where Y

p

is the sample space variable corresponding to the

random variable Y. When more than one variable is required

to capture the ¯ame structure, a joint probability density

function

P(Y

p

1

; ¼; Y

p

N

; x; t) is introduced [97]. The mean

burning rate (or any mean quantity) is then estimated as:

_ v

Y

1

(x; t)

=

_

Y

1

¼

_

Y

N

_ v

Y

1

(Y

p

1

; ¼; Y

p

N

)

P(Y

p

1

; ¼; Y

p

N

; x; t)dY

p

1

¼dY

p

N

(101)

Conditional statistics have been used to de®ned the ¯ame

surface density (Section 6.3.2), these conditional means are

also useful in a pdf context. Consider a non-premixed ¯ame

where the chemistry is reduced to a single step reaction, and

where radiative heat losses are neglected. Laminar combus-

tion would be parameterized with two variables, for

example, fuel mass fraction Y

F

and mixture fraction Z (see

Section 3.2). The turbulent ¯ame is then fully described by

the joint pdf of mixture fraction and fuel mass fraction,

P(Y

p

F

; Z

p

; x; t): For such ¯ames, it is interesting to focus on

the statistical properties of the fuel mass fraction Y

F

for a

given value of the mixture fraction Z (Section 3.2 and

Fig. 5).The conditional pdf

P

c

(Y

p

F

uZ

p

; x; t) is introduced

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 218

and de®ned for a given value of Z

p

as:

P(Y

p

F

; Z

p

; x; t) =

P

c

(Y

p

F

uZ

p

; x; t)

P(Z

p

; x; t)

The conditional mean is readily de®ned from:

_

Y

F

uZ(x; t) = Z

p

_

=

_1

0

Y

p

F

P

c

(Y

p

F

uZ

p

; x; t)dY

p

F

In this decomposition,

P(Z

p

; x; t) describes the statistical

properties of fuel/air mixing, whereas

P

c

(Y

p

F

uZ

p

; x; t) and

_

Y

F

uZ(x; t) = Z

p

_

are linked to the internal structure of the

¯ame front.

Two main directions may be chosen to built numerical

models from pdf:

² To presume the pdf shape from available mean quantities.

² To solve a balance equation for the pdf.

6.4.2. Presumed probability density functions

The simplest approach to the mixing between fuel and

oxidizer in a non-premixed ¯ame is to presume the shape

of the pdf

~

P(Z

p

; x; t); usually with a Beta function [72]:

~

P Z

p

; x; t

_ _

=

(Z

p

)

a21

1 2Z

p

_ _

b21

_1

0

(Z

1

)

a21

1 2Z

1

_ _

b21

dZ

1

This presumed pdf should reproduce the mean of the

mixture fraction

~

Z and its variance

¯

Z

//2

:

~

Z =

_1

0

Z

p

~

P Z

p

; x; t

_ _

dZ

p

;

¯

Z

//2

=

_1

0

Z

p

2

~

Z

_ _

2

~

P Z

p

; x; t

_ _

dZ

p

and using the relation [98]:

_1

0

(Z

p

)

n

~

P Z

p

; x; t

_ _

dZ

p

=

a(a 11)¼(a 1(n 21))

(a 1b)(a 1b 11)¼(a 1b 1(n 21))

The two-parameters a and b are determined as:

a =

~

Z

~

Z(1 2

~

Z)

¯

Z

//2

21

_ _

$ 0; b = a

1

~

Z

21

_ _

$ 0

This technique requires the solving of a balance equation for

the mean and the variance of Z (Section 9.2).

¯

Z

//2

vanishes

when reactants are perfectly mixed and reaches its maxi-

mum value,

~

Z

_

1 2

~

Z

_

; when fuel and oxidizer are comple-

tely segregated (Section 5.3), then the pdf takes the formof a

double peak function, with peaks located at Z = 0 and Z = 1:

This approach is also used in premixed ¯ames replacing the

mixture fraction Z by the progress variable c [99].

6.4.3. Pdf balance equation

The pdf balance equation is ®rst written for the progress

variable c. The time evolution of a probability density func-

tion is easily derived by expressing the pdf as the average of

a function d(c

p

2c(x; t)) de®ned as [100]:

d c

p

2c(x; t)

_ _

= 1=Dc

p

if c

p

2Dc

p

=2 , c , c

p

1Dc

p

=2

d(c

p

2c(x; t)) = 0 otherwise

By de®nition:

P(c

p

; x; t) = d(c

p

2c(x; t)) (102)

The time evolution of the reactive species c(x; t) is given by:

2c

2t

= 2u´7c 1

1

r

[7´(rD7c(x; t)) 1 _ v] (103)

and d(c

p

2c(x; t)) satis®es:

2

2C

[d(c

p

2c(x; t))] =

2

2c

[d(c

p

2c(x; t))]

2c(x; t)

2C

= 2

2

2c

p

[d(c

p

2c(x; t))]

2c(x; t)

2C

(104)

2

2

2C

2

[d(c

p

2c(x; t))]

= 2

2

2c

p

[d(c

p

2c(x; t))]

2

2

c(x; t)

2C

2

1

2

2

2c

p2

[d(c

p

2c(x; t))]

×

_

2c(x; t)

2C

_

2

(105)

where C can either be time or any spatial coordinates. These

relations (104) and (105) and Eq. (102) are key relations,

useful to obtain all equations discussed in this section.

After simple manipulations combining Eqs. (102) and

(103) with Eq. (104) the transport equation for the pdf

may be written:

2

2t

[

P(c

p

; x; t)] = 2

2

2c

p

2c(x; t)

2t

uc(x; t) = c

p

_ _

P(c

p

; x; t)

_ _

(106)

where conditional averaging

Q(c)d(c

p

2c(x; t)) =

_

Q(c)uc(x; t) = c

p

_

P(c

p

; x; t)

is introduced.

Eq. (106) shows that the time evolution of the pdf is

controlled by a ¯ux in the sample space c

p

. This ¯ux is

driven by a velocity equal to the conditional mean of the

time evolution of the progress variable c. In other words,

when the mean of (2c(x; t)=2t) is non-zero for the value c =

c

p

; the probability of ®nding the occurrence c = c

p

is modi-

®ed. Because the probability to ®nd all the possible values is

constant and equal to unity:

_1

0

P(c

p

; x; t)dc

p

= 1

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 219

an increase or a decrease of the density of probability for

c = c

p

; implies a modi®cation of this density for other

values of c

p

, justifying the convective term in sample

space. This term insures that the density of probability is

transported from point to point in the sample space c

p

in a

conservative manner. The ¯ux depends on the conditional

mean of the time evolution of the progress variable given by

Eq. (103):

2c

2t

uc(x; t) = c

p

_ _

= 2u´7cuc(x; t) = c

p

_ _

1

1

r

[7´(rD7c(x; t))]uc(x; t) = c

p

_ _

1 _ vuc(x; t) = c

p

_ _

(107)

In Eq. (107), the conditional value of the source term is a

function of c de®ned in one-point and is exactly known

since:

_ v(c)uc(x; t) = c

p

_ _

= _ v(c

p

)

The main advantage of pdfs in turbulent combustion lies in

this availability to deal with chemistry, any term de®ned in

one-point (as chemical source) is closed. Nonetheless, this

advantage is offset by the fact that reactants are brought to

the reaction zone by diffusion, and the conditional mean of

2c(x; t)=2t also includes a conditional diffusive term:

1

r

(7´(rD7c))uc(x; t) = c

p

_ _

(108)

named the micromixing term and remaining unclosed (as

any term involving spatial derivatives). This micromixing

term may be rewritten with the scalar dissipation rate,

x = Du7cu

2

; using Eq. (102) with Eq. (105) and assuming

rD < cst :

2

2c

p

D7

2

cuc(x; t) = c

p

_ _

P(c

p

; x; t)

_ _

= 2D7

2

P(c

p

; x; t) 1

2

2

2c

p2

xuc(x; t) = c

p

_ _

P(c

p

; x; t)

_ _

The ®rst term in the RHS, D7

2

P(c

p

; x; t) is usually neg-

ligible compared to the transport due to velocity ¯uctua-

tions. The total dissipation rate, x is recovered as:

x =

_1

0

xuc(x; t) = c

p

_ _

P(c

p

; x; t)dc

p

Using diffusion or scalar dissipation rate, the probability

density function balance equation may be organized in

two different forms:

² In terms of molecular diffusion:

2

2t

[

P(c

p

; x; t)]

= 2

2

2c

p

___

2u´7c 1

1

r

7´(rD7c)uc(x; t) = c

p

_

1 _ v(c

p

)

_

P(c

p

; x; t)

_

(109)

² In terms of scalar dissipation rate:

2

2t

[

P(c

p

; x; t)]

= 2

2

2c

p

2u´7cuc(x; t) = c

p

_ _

1 _ v(c

p

)

_ _

P(c

p

; x; t)

_ _

2

2

2

2c

p2

xuc(x; t) = c

p

_ _

P(c

p

; x; t)

_ _

The convective term may be split into mean,

u´7

P(c

p

; x; t)

and ¯uctuating components. Using an eddy viscosity model,

the ¯uctuating part becomes:

2n

t

7

P(c

p

; x; t)

leading to the pdf balance equation:

2

2t

[

P(c

p

; x; t)] 1 u´7

P(c

p

; x; t)

= n

t

7

P(c

p

; x; t) 2

2

2c

p

[ _ v(c

p

)

P(c

p

; x; t)]

2

2

2

2c

p2

xuc(x; t) = c

p

_ _

P(c

p

; x; t)

_ _

(110)

Weighted, or Favre, averages are also introduced in pdfs, for

instance when r = r(c) :

r

~

P(c

p

; x; t) = r(c

p

)d(c(x; t) 2c

p

) = r(c

p

)

P(c

p

; x; t)

When more than one species is taken into account, the pdf

balance equation is derived using the same formalism with

r

~

P(Y

p

1

; ¼; Y

p

N

) = r(Y

1

; ¼; Y

p

N

)d(Y

1

2Y

p

1

)¼d(Y

N

2Y

p

N

) [93]:

r

2

2t

~

P(Y

p

1

; ¼Y

p

N

; x; t)

= 2 r

N

i=1

2

2Y

p

i

___

2u´7Y

i

uY(x; t) = Y

p

_

1 _ v

p

i

_

×

~

P

_

Y

p

1

; ¼; Y

p

N

; x; t

__

2 r

N

i=1

N

j=1

2

2Y

p

i

2

2Y

p

j

__

D

2Y

i

2x

k

2Y

j

2x

k

uY(x; t) = Y

p

_

×

~

P

_

Y

p

1

; ¼; Y

p

N

; x; t

_

_

(111)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 220

The same terms are identi®ed on the RHS: convection by the

conditional velocity and by the chemical source, and micro-

mixing. As was done in Eq. (110), the convective term

in physical space can be decomposed into mean and

¯uctuating parts.

6.4.4. Joint velocity/concentrations pdf

To avoid the gradient transport assumption for the

turbulent ¯ux, (u

/

´7cuc(x; t) = c

p

); the joint velocity/

concentration pdf is introduced. Once this joint pdf is

known, turbulence models, such as k±1, are, a priori, no

longer required for the mean ¯ow. However, an equation

for 1 is still needed to estimate a characteristic mixing time,

except when the frequency of mixing is also included in the

joint pdf [101]. The transport equation for this joint compo-

sition/velocity pdf is given below for the progress variable c,

where the RHS contains the unclosed terms.

2

2t

[

P(u

p

; c

p

; x; t)] 1u

p

´7

P(u

p

; c

p

; x; t)

1 n7

2

u

i

2

1

r

2 p

2x

i

_ _

2

2u

p

i

[

P(u

p

; c

p

; x; t)]

1

2

2c

p

[ _ v(c

p

)

P(u

p

; c

p

; x; t)]

=

2

2u

p

i

1

r

2p

/

2x

i

_

_

_

_

u = u

p

; c = c

p

_ _

P(u

p

; c

p

; x; t)

_ _

....................,,....................

Pressure fluctuations

2

2

2u

p

i

n7

2

u

/

i

uu = u

p

; c = c

p

_ _

P(u

p

; c

p

; x; t)

_ _

...................,,...................

Viscous dissipation

2

2

2c

p

D7

2

cuu = u

p

; c = c

p

_ _

P(u

p

; c

p

; x; t)

_ _

...................,,...................

Molecular diffusion

(112)

The LHS terms are closed and represent respectively accu-

mulation, convection in physical space by the random vel-

ocity ®eld (incorporating turbulent transport), convection of

the pdf in velocity space, here the convective velocity is the

mean of the viscous dissipation and the mean pressure gradi-

ent, and ®nally the closed chemical source. The unclosed

terms (RHS) are the pressure ¯uctuations, the ¯uctuating

part of the viscous dissipation and micromixing. All these

phenomena remain to be closed.

6.4.5. Conditional moment closure (CMC)

Conditional moment closure modeling was ®rst proposed

in Refs. [102,103]. As with the pdf, the idea is to focus on

particular states between fresh gases and fully burnt product

in premixed ¯ames, or, between fuel and oxidizer in non-

premixed combustion. However, here only conditional

moments (rY

i

uc = c

p

) are considered. In premixed ¯ames,

the conditional quantity is the progress variable c, whereas

for non-premixed combustion, the mixture fraction is used.

In premixed ¯ames, the mean value,

~

Y

i

of Y

i

may be esti-

mated as:

r

~

Y

i

=

_1

0

rY

i

uc = c

p

_ _

P(c

p

; x; t)dc

p

(113)

One may solve a balance equation for the conditional quan-

tities Q

i

(c

p

) de®ned as:

Q

i

(c

p

) =

rY

i

uc = c

p

_ _

ruc = c

p

_ _

This balance equation is [104,105]:

ruc = c

p

_ _

2Q

i

2t

= 2 ru

i

uc = c

p

_ _

2Q

i

2x

i

1 rxuc = c

p

_ _

2

2

Q

i

2c

p2

1 _ v

i

uc = c

p

_ _

1E

Q

i

1E

Y

i

(114)

The two last terms of Eq. (114) are usually neglected,

E

Q

i

appears from molecular diffusion along with differ-

ential diffusion effects across the iso-c surface, E

Y

i

represents the effects of turbulent ¯uctuations on the

deviation from the conditional mean. The three ®rst

terms on the right hand side are unclosed, they describe

convective transport, micromixing (x enters this term) and

chemical source. Closures for the conditional values of the

scalar dissipation rate, but also for the conditional value of

the chemical source of Y

i

, calculated for a given value of the

progress variable c

p

or the mixture fraction Z, are required

[106,107]. One equation has to be solved for each value of

c

p

retained. The number of these values is determined from

the accuracy required to estimate both the mean from Eq.

(113) and the second order derivative in the sample space

(i.e. 2

2

/2c

p2

) found in Eq. (114). On the other hand, the

probability density function entering expression (113),

P(c

p

; x; t) is presumed.

CMC may also be viewed as a multi-surface description,

any conditional quantity

_

rY

i

uc = c

p

_

corresponding to the

conditional average of Y

i

along the iso-surface c = c

p

:

6.5. Similarities and links between the tools

Major links between the tools used in turbulent combus-

tion modeling are now developed. Without loss of the

generic character of the discussion, we consider the case

of a turbulent premixed ¯ame represented with a progress

variable c. To describe this turbulent ¯ame, three quantities

are useful:

² The scalar dissipation rate of the progress variable

r ~ x = rD7c´7c.

² The pdf of the progress variable

P(c

p

; x; t):

² The ¯ame surface density S or the mean ®eld

G:

Simple links exist between these quantities:

² The conditional value of the scalar dissipation rate enters

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 221

the pdf transport equation (Eq. (110)), therefore there is a

direct connection between

P(c

p

; x; t) and ~ x:

² The surface density function S(c

p

; x; t); or ¯ame surface,

is related to the pdf via the conditional value of u7cu; S =

_

u7cuuc = c

p

_

P(c

p

) (Eq. (86)).

Since x is proportional to u7cu; the ¯ame surface S, the

pdf

P(c

p

) and the dissipation rate ~ x are very strongly

related. Using the joint pdf of c and x, the ¯ame surface

density may be written:

S(c

p

; x; t) =

x

D

_

uc = c

p

_ _

P(c

p

; x; t)

=

_

x

x

p

D

_

P(c

p

; x

p

; x; t)dx

p

(115)

and characteristic length scales d

c

=

D=x

_

of the iso-c

distribution are embedded within S(c

p

).

Combining Eq. (86) and Eq. (91) with Eq. (96), the mean

c-scalar dissipation rate is also a function of surface

densities:

r ~ x = rDu7cu

2

=

_1

0

krDu7cul

s

S(c

p

)dc

p

= krDu7cul

s

u7cu

(116)

where the generalized surface average kQl

s

is de®ned by

Eq. (96). Relations between ¯ame surface densities and

scalar dissipation rates were anticipated by Borghi [108].

In premixed combustion, using a G-®eld equation, the

¯ame front is identi®ed to a given level G = G

0

[29].

The ¯ame surface density is then:

S(G

0

) = u7GuuG = G

0

_ _

P(G

0

) (117)

Hence, previous relations may be recast in terms of the G-

equation.

The CMC formalism (see Section 6.4.5) may be re-

organized in term of ¯ame surface density:

r

~

Y

i

=

_1

0

rY

i

uc = c

p

_ _

P(c

p

; x; t)dc

p

=

_1

0

rY

i

u7cu

_ _

s

S(c

p

)dc

p

(118)

where the conditional mean Q

i

(c

p

) may be written:

Q

i

(c

p

) =

rY

i

uc = c

p

_ _

ruc = c

p

_ _

=

krY

i

=u7cul

s

kr=u7cul

s

(119)

and appears as directly related to c = c

p

surface averaged

quantities.

The links between the combustion modeling tools are

summarized in Table 2. The mean burning rate is given by:

_ v =

_1

0

_ v(c

p

)

P(c

p

)dc

p

(120)

Using relations (86) and (91) and Eq. (116), this expression

becomes:

_ v =

_1

0

_ vuc = c

p

_ _

P(c

p

)dc

p

=

_1

0

_ v

u7cu

_ _

s

S(c

p

)dc

p

=

_ v

u7cu

_ _

s

u7cu =

_ v

x=D

_

_ _

s

S =

1

krDu7cul

s

_ v

u7cu

_ _

s

r ~ x

(121)

Models based on probability density functions, conditional

means, ¯ame surface density function and generalized ¯ame

surface density are then related via the scalar dissipation rate.

These expressions are extended to burning rate depending

on many species using multi-dimensional pdfs and con-

ditional averaging. When the local reaction rate is a function

of various quantities (species mass fractions, temperature,

etc.), Y

i

and a sampling scalar c:

_ v =

_

Y

1

¼

_

Y

N

_1

0

_ v

i

(Y

p

1

; ¼; Y

p

N

; c

p

)

P(Y

p

1

; ¼; Y

p

N

; c

p

)dY

p

1

; ¼; dY

p

N

dc

p

=

_1

0

_

Y

1

¼

_

Y

N

_ v

i

(Y

p

1

; ¼; Y

p

N

; c

p

)

P

c

(Y

p

1

; ¼; Y

p

N

uc

p

)dY

p

1

; ¼; dY

p

N

_ _

..........................,,..........................

_ v;(Y

1

;¼;Y

N

;c)uc=c

p

_ _

P(c

p

)dc

p

(122)

decomposing the joint-pdf as

P(Y

p

1

; ¼; Y

p

N

; c

p

) =

P

c

(Y

p

1

; ¼; Y

p

N

uc

p

)

P(c

p

): Note that such a relation may be

used when in premixed ¯ames the progress variable, based

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 222

Table 2

Exact relations between mean scalar dissipation rate, ~ x, probability density function,

P(c

p

), or surface density, S(c

p

), of iso-surface c = c

p

and

generalized ¯ame surface density S = u7cu: Connections with the G-equation are readily obtained with Eq. (117). kQl

s

is a surface average (Eq.

(91)), kQl

s

corresponds to generalized surface average and is de®ned in Eq. (96). c is the conditional variable (progress variable in premixed

¯ames, mixture fraction in non-premixed combustion)

Scalar dissipation rate r ~ x Probability density function

P(c

p

)

Flame surface density S(c

p

)

r ~ x ±

_1

0

_

rDu7cu

2

uc = c

p

_

P(c

p

)dc

p

_1

0

krDu7cul

s

S(c

p

)dc

p

P(c

p

)

_

xuc = c

p

_

via PDF Eq: ± S(c

p

)=

_

u7cuuc = c

p

_

S = u7cu

r ~ x

krDu7cul

s

_1

0

_

u7cuuc

p

_

P(c

p

)dc

p

_1

0

S(c

p

)dc

p

on reactant mass fractions, and the temperature are indepen-

dent (compressibility effects, heat losses, non-unity Lewis

numbers, etc.).

The fundamentals of turbulent combustion modeling

clearly rely on pdf, ¯ame surface density, G-®eld and scalar

dissipation rate concepts. The previous relations may be

used to carefully compare the proposed closure schemes

and Table 3 summarizes the tools and their related modeling

issues. Various expressions for mean reaction rates are

displayed in Table 4. Major differences between the various

approaches only appear when closing the unknown quanti-

ties, but at the light of these relations, many closures are

essentially equivalent.

7. Reynolds-averaged models for turbulent premixed

combustion

7.1. Turbulent ¯ame speed

Turbulent premixed ¯ames may be described in terms of a

global turbulent ¯ame speed S

T

. From experimental data

[109,110] or theoretical analysis (Renormalization group

theory [111]), the following expression has been proposed:

S

T

S

L

= 1 1a

u

/

S

L

_ _

n

(123)

where a and n are two model constants of the order of unity.

u' is the turbulent velocity (i.e the RMS velocity).

Unfortunately, the turbulent ¯ame speed S

T

is not a fully

well de®ned quantity [112]. Experimental data exhibit a

large scatter because they depend on various parameters

(chemistry characteristics, turbulence scales, ¯ow geometry,

etc.). While this global approach is not particularly well suited

to close Favre averaged transport equations, it may be of

interest in the context of Large Eddy Simulation [79,113].

7.2. Eddy-Break-Up model

Devised in Ref. [114], this model is based on a phenom-

enological analysis of turbulent combustion assuming high

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 223

Table 3

Tools for turbulent combustion modeling

Description Tools Modeling issues

Geometrical G-®eld with G = G

p

at the ¯ame

2

~

G=2t 1 ~ u´7

~

G = (r

u

= r)S

T

u7

~

Gu

Propagation speed S

T

Flame surface S

Algebraic closure or transport equation

_ v

i

= Combustion

_

V

i

S

Turbulence

Total stretch = curvature 1strain rate

Small scales Scalar dissipation rate, r ~ x = rDu7Yu

2

Fast chemistry

~

_ v < ~ x

Algebraic closure or transport equation Provide the rate of micromixing

Statistical Probability density function

~

P(Y

p

F

; Z

p

; x; t) Micromixing

(1) Presumed 1¯amelets (1) Strategy to generate the pdf

(2) Transport equation (2) Solve 2

~

P=2t = ¼

Conditional mean, CMC, presumed PDF

_

Y

F

uZ(x; t) = Z

p

_

=

_

1

0

Y

p

F

P

c

(Y

p

F

uZ

p

; x; t)dY

p

F

Micromixing and conditional source

Fundamental links: S(x; t) =

_

u7cuuc = c

p

_

P(c

p

; x; t) and r ~ x = krDu7cul

s

u7cu

Table 4

Exact expressions for the averaged reaction rate as a function of the mean scalar dissipation, ~ x; probability density function,

P(c

p

) or surface

density, S(c

p

) of iso-surface c = c

p

: Connections with the G-equation are readily obtained with Eq. (117). kQl

s

is a surface average (Eq. (91)),

kQl

s

corresponds to generalized surface average and is de®ned in Eq. (96)

Tools Averaged reaction rate

Scalar dissipation r

_

1

0

_

_ v

u7cu

_

s

S(c

p

)dc

p

_

1

0

krDu7cul

s

S(c

p

)dc

p

~ x = r

_

_ v

u7cu

_

s

krDu7cul

s

~ x

Probability density function

_

1

0

_ v(c

p

)

P(c

p

)dc

p

Flame surface density

_

1

0

_

_ v

u7cu

_

s

S(c

p

)dc

p

=

_

_ v

u7cu

_

s

u7cu

Reynolds (Re @ 1) and DamkoÈhler (Da @ 1) numbers. The

reaction zone is viewed as a collection of fresh and burnt

gases pockets. Following the Kolmogorov cascade, turbu-

lence leads to a break down of fresh gases structures.

Accordingly, the mean reaction rate is mainly controlled

by the turbulent mixing time t

t

. When oxidizer is in excess,

the mean reaction rate is expressed as:

_ v

F

= 2C

EBU

r

¯

Y

//2

F

_

t

t

(124)

where Y

//

F

denotes the fuel mass fraction ¯uctuations and

C

EBU

is a model constant of the order of unity. The turbulent

mixing time, t

t

is estimated from the turbulence kinetic

energy k and its dissipation rate 1 according to:

t

t

=

k

1

as an approximation of the characteristic time of the integral

length scales of the turbulent ¯ow ®eld.

The reaction rate may be recast in terms of progress vari-

able, c, as:

_ v = 2C

EBU

r

¯

c

//2

_

t

t

(125)

Mass fraction ¯uctuations

¯

Y

//2

F

(or progress variable ¯uctua-

tions

¯

c

//2

) must be modeled and may be estimated from a

balance equation (see Eq. (77)). Assuming an in®nitely thin

¯ame front (i.e. c = 0 or c = 1),

¯

c

//2

is easily estimated

because c

2

= c :

r

¯

c

//2

= r(c 2 ~ c)

2

= r(

`

c

2

2 ~ c

2

) = r ~ c(1 2 ~ c) (126)

The square root has been introduced for dimensional reasons

in Eqs. (124) and (125) but, unfortunately, Eqs. (125) and

(126) lead to inconsistencies because the ~ c derivative of

_ v;

d

_ v=d~ c; is in®nite both when ~ c = 0 and when ~ c = 1 (Borghi,

1999, private communication). Then, a corrected version of

the Eddy-Break-Up model, without the square root, is used

for practical simulations:

_ v = C

EBU

r

1

k

~ c(1 2 ~ c) (127)

or, in terms of fuel mass fraction:

_ v

F

= 2C

EBU

r

1

k

~

Y

F

Y

0

F

1 2

~

Y

F

Y

0

F

_ _

(128)

where Y

0

F

is the initial fuel mass fraction in the reactant

stream, assuming excess of oxidizer.

The EBU model was found attractive because the reaction

rate is simply written as a function of known quantities

without any additional transport equation and is available

in most commercial CFD codes. The modeled reaction rate

does not depend on chemical characteristics and assumes a

homogeneous and isotropic turbulence. Some adjustments

of the model constant C

EBU

have been proposed to mimic

chemical features [115]. Eddy-Break-Up modeling tends to

overestimate the reaction rate, especially in highly strained

regions, where the ratio 1/k is large (¯ame-holder wakes,

walls, etc.).

7.3. Bray±Moss±Libby model

7.3.1. Introduction

Known under the initials of its authors, Bray, Moss and

Libby, or, from the involved physical hypothesis, BiModal

Limit, this model, ®rst proposed in 1977 [116], has been the

subject of a large amount of work leading to many improve-

ments (see papers by Bray, Moss and Libby, and then by

Bray, Champion and Libby). Combining a statistical

approach using probability density functions and a physical

analysis, this model has evidenced some special features of

turbulent premixed combustion (counter-gradient turbulent

transport, ¯ame turbulence generation, etc.). The presenta-

tion is mainly limited here to basic concepts of the Bray±

Moss±Libby (BML) formulation.

A one-step, irreversible chemical reaction between two

reacting species, fresh gases (R) and combustion products

(P) is considered:

R - P

Classical assumptions are made to simplify the formula-

tion: perfect gases, incompressible ¯ows, constant chemical

properties, unity Lewis numbers, etc. A progress variable, c,

of the chemical reaction is introduced where c = 0 in fresh

gases and c = 1 in fully burnt gases, as described in Section

3.1.

The basic idea of the BML formulation is to presume the

probability density function of the progress variable c at a

given location (x; t) as a sum of fresh, fully burnt and burn-

ing gases contributions (Fig. 15):

P(c

p

; x; t) = a(x; t)d(c

p

)

....,,....

fresh gases

1b(x; t)d(1 2c

p

)

......,,......

burnt gases

1g(x; t)f (c

p

; x; t)

......,,......

burning gases

(129)

where a, b and g respectively denote the probability to

have, at location (x; t); fresh gases, burnt gases and reacting

mixture. d(c

p

) and d(1 2c

p

) are respectively the Dirac delta

functions corresponding to fresh gases (c = 0) and fully

burnt ones (c = 1):

Normalization of the probability density function:

_1

0

P(c

p

; x; t)dc

p

= 1 (130)

leads to the following relations:

a 1b 1g = 1 (131)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 224

_1

0

f (c

p

; x; t)dc

p

= 1 (132)

with f (0) = f (1) = 0:

The balance equation for the progress variable c may be

written:

2rc

2t

17´(ruc) = 7´(rD7c) 1 _ v (133)

This equation is averaged and the mean reaction rate, at the

location (x; t) is:

_ v(x; t) =

_1

0

_ v(c

p

)

P(c

p

; x; t)dc

p

(134)

leading to Eq. (129):

_ v(x; t) = g(x; t)

_1

0

_ v(c

p

)f (c

p

; x; t)dc

p

(135)

All studies devoted to this line of models are based on such a

formulation. The objective is now to determine the unknown

functions a, b, g and the probability density function f.

Using the DamkoÈhler number, Da, comparing turbulence

and chemical time scales, and, the turbulent Reynolds

number Re, we will focus only on the case where

Re @Da @1. In this situation, the combustion is controlled

by the turbulent transport and the reaction zone may be

assumed to be in®nitely thin. Accordingly, g !1 (i.e.

a @g and b @g).

7.3.2. BML model analysis

This model is developed under the assumption that

Re @Da @1, corresponding to g !1. At a given location

in the ¯ow, an intermittency between fresh gases (c = 0)

and fully burnt ones (c = 1) is observed and the probability

density function of the progress variable c reduces to:

P(c

p

; x; t) = a(x; t)d(c

p

) 1b(x; t)d(1 2c

p

) 1O(1=Da)

(136)

which is `quasi-bimodal' for large DamkoÈhler numbers.

Then at the point (x; t) inside the reaction zone, c looks as

a telegraphic signal as displayed in Fig. 16. This signal also

satis®es:

c

2

= c 1O(1=Da) and c(1 2c) = O(1=Da) (137)

Under this assumption, a and b are easily determined as a

function of the Favre average progress variable ~ c :

r ~ c = r ~ c =

_1

0

rc

p

P(c

p

)dc

p

= r

b

b (138)

where r

b

is the burnt gases density.

b =

r ~ c

r

b

and a = 1 2

r ~ c

r

b

(139)

The mean density r may be written:

r =

_1

0

r

P(c

p

)dc

p

= ar

u

1br

b

= 1 2

r`c

r

b

_ _

r

u

1

r`c

r

b

r

b

(140)

One may also introduce the reaction heat release factor t,

de®ned as:

t =

r

u

r

b

21 =

T

b

T

u

21 (141)

leading to:

r

u

= (1 1t)r

b

= r(1 1t~ c) (142)

corresponding to the perfect gases state law, assuming a

constant pressure P. Then, the probabilities a and b

become:

a =

1 2 ~ c

1 1t~ c

; b =

(1 1t) ~ c

1 1t~ c

(143)

The probability density function

P(c) is determined and

depends only on the mean progress variable ~ c (and on the

heat release factor, t, which is ®xed for a given chemical

reaction). The BML model involves a presumed pdf (Eq.

(135)), however, the mean reaction rate cannot be calculated

from the pdf since g was neglected in Eq. (129)).

Starting from the conservative and non-conservative

forms of the progress variable c balance equations:

2rc

2t

17´(ruc) = 7´(rD7c) 1 _ v

r

2c

2t

1ru´7c = 7´(rD7c) 1 _ v

and multiplying by c and adding these equations:

2rc

2

2t

17´(ruc

2

) = 7´(rD7c

2

) 22rD7c´7c 12c _ v (144)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 225

Fig. 15. Probability density function in premixed turbulent

combustion.

Fig. 16. Intermittency between fresh and fully burnt gases at a given

location in the reaction zone. This signal corresponds to a bimodal

(c = 0 and c = 1) probability density function.

Subtracting the balance equation for c

2

(Eq. (144)) to the

balance equation for the progress variable c (Eq. (133))

leads to a balance equation for c(1 2c) :

2

2t

[rc(1 2c)] 17´[ruc(1 2c)]

= 7´(rD7[c(1 2c)]) 12rD7c´7c 22c _ v 1 _ v (145)

Under the assumption of the BML model, the progress vari-

able c is equal to zero or to unity. Accordingly, c(1 2c) = 0

and the balance Eq. (145) reduces to:

2rD7c´7c = 2c _ v 2 _ v (146)

leading to, after averaging:

2rD7c´7c = (2c

m

21)

_ v (147)

where a progress variable c

m

, de®ned as:

c

m

=

_1

0

c _ vf (c)dc

_1

0

_ vf (c)dc

(148)

has been introduced and characterizes the chemical reaction.

The mean reaction rate

_ v becomes:

_ v = 2

rx

2c

m

21

(149)

where rx is the scalar dissipation rate of the progress

variable c.

rx = r` x = rD7c´7c = rD

2c

2x

i

2c

2x

i

(150)

The mean reaction rate

_ v is then related to the dissipation

rate ~ x; describing the turbulent mixing, and to c

m

, character-

izing the chemical reaction. A transport equation for the

scalar dissipation rate may be written and solved [117], or

one may postulate a linear relaxation of the ¯uctuations

generated by micromixing (Section 9.2.2), leading to:

rx =

rc

//2

t

t

(151)

where a turbulent time scale, t

t

is introduced. Assuming an

intermittency between fresh and burnt gases (c = 0 or

c = 1),

¯

c

//2

is given by Eq. (126). Then:

_ v =

2

2c

m

21

r ~ c(1 2 ~ c)

t

t

(152)

The Eddy-Break-Up model expression (Eq. (127)) is

recovered (Section 7.2). The BML model may then be

viewed as a theoretical derivation, where the assumptions

made are clearly stated, of the Eddy-Break-Up (EBU)

model, initially based on a phenomenological approach.

7.3.3. Recovering mean reaction rate from tools relations

The analysis developed in Section 7.3.2 is the usual

derivation of the BML model. However, expression (149)

linking mean reaction rate and scalar dissipation may also be

recovered from the general relations between modeling

tools (Eq. (121)):

_ v =

1

krDu7cul

s

_ v

u7cu

_ _

s

r ~ x (153)

To link mean reaction rate and scalar dissipation rate using

the ¯amelet analysis, Eq. (153) requires estimates for _ v=u7cu

and rDu7cu; averaging of these quantities along iso-surface

c = c

p

and integration over all the possible c

p

values.

Assuming that in®nitely thin ¯ame elements may be viewed

as one-dimensional, steady state, premixed laminar ¯ame

propagating at a given laminar ¯ame speed S

L

, the balance

equation for the progress variable c in this 1D ¯ame is:

r

0

S

L

7´c = 7´(rD7c) 1 _ v (154)

where r

0

denotes the density in the fresh gases. Integrating

Eq. (154) across all the ¯ame front from fresh gases to burnt

gases (21 # x # 11) and up to a given location x

0

, corre-

sponding to a progress variable c

0

(21 # x # x

0

); leads to:

r

0

S

L

=

_11

21

_ vdx =

_1

0

_ v

u7cu

dc (155)

(rD7c)

x

0

= r

0

S

L

c

0

2

_x

0

21

_ vdx

=

_1

0

_ v

u7cu

dc

_ _

c

0

2

_c

0

0

_ v

u7cu

dc (156)

As the ¯ame front is supposed to be thin and planar, all c

p

iso-surfaces have the same surface and all quantities are

constant along these surfaces, moreover the ¯ame is

supposed locally one-dimensional and u7cu = 7c: Then:

_1

0

k _ v=u7cul

s

dc

p

_1

0

krDu7cul

s

dc

p

=

_1

0

_ v=u7cudc

_1

0

rDu7cudc

=

_11

21

_ vdx

_11

21

_ vdx

_1

0

c dc 2

_1

0

_c

0

0

( _ v=u7cu)dc dc

0

=

2

_11

21

_ v dx

2

_11

21

c _ vdx 2

_11

21

_ v dx

=

2

2c

m

21

(157)

where c

m

is de®ned as:

c

m

=

_11

21

c _ vdx

_11

21

_ vdx

(158)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 226

Then, from Eq. (153):

_ v = 2

rx

2c

m

21

(159)

recovering Eq. (149). This second analysis leads to the

following comments:

² The de®nition of c

m

is slightly different in Eqs. (149) and

(159). In the ®rst one, the integration is performed over

all possible values of the progress variable c at a given

location (x; t): In the second analysis, the integration is

performed along a normal direction to the ¯ame front,

assumed to be a laminar one-dimensional premixed

¯ame. The expressions (148) and (158) are identical

when:

f (c) =

1

d

l

dc

dx

_ _

21

f

(160)

where (dc=dx)

f

describes the inner structure of the ¯ame-

let.

² This second derivation clearly exhibits how a ¯amelet

model works (the turbulent ¯ame is viewed as a collec-

tion of thin laminar ¯ame elements). The c

m

parameter is

related to the inner structure of the ¯ame front and to the

properties of the chemical reaction.

² Making use of the ¯ame surface density S as de®ned by

Eq. (116), the expression for the mean burning rate

becomes:

_ v = r

0

S

L

S (161)

where

S =

2

2c

m

21

1

r

0

S

L

r ~ x <

2

2c

m

21

1

r

0

S

L

r

~ c(1 2 ~ c)

t

t

(162)

² The BML model has been developed starting from a

statistical analysis (two peaks pdf), the mean reaction

rate was written in terms of the scalar dissipation rate

of the progress variable c, which may be recast in terms

of ¯ame surface density, exhibiting simple links between

modeling tools.

7.3.4. Reynolds and Favre averaging

Assuming a bimodal distribution of the progress variable

c, Reynolds ( c) and Favre ( ~ c) averages are easily related.

From:

c =

_1

0

c

P(c)dc = b (163)

rc = r ~ c =

_1

0

rc

P(c)dc = r

b

b = r

b

c (164)

together with expressions (142) and (143), one easily

obtains:

c =

(1 1t) ~ c

1 1t~ c

(165)

corresponding to a model for the density/progress variable

correlations (Eq. (15)):

r

/

c

/

= 2 r

t~ c(1 2 ~ c)

1 1t~ c

(166)

Reynolds ( c) and Favre ( ~ c) averages of the instantaneous

progress variable c are compared on Fig. 17 for various

values of the heat release factor t. The discrepancy between

the two quantities strongly increases with t. r

/

c

/

is plotted

as a function of ~ c on Fig. 18.

7.3.5. Conditional averagingÐcounter-gradient turbulent

transport

The analysis of intermittency between fresh gases (c = 0)

and fully burnt gases (c = 1) leads to the introduction of

conditional averaging. The Favre average

~

Q of any quantity

Q may be expressed as a function of the conditional

averages for fresh gases (

Q

u

) and fully burnt products (

Q

b

) :

r

~

Q = rQ =

_1

0

rQ

P(c)dc = ar

u

Q

u

1br

b

Q

b

(167)

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 227

Fig. 17. The Reynolds average c of the progress variable cis plotted as a function of the Favre average ~ c for various values of the heat release

factor t, assuming a bimodal distribution (c = 0 or c = 1) of c (Eq. (165)).

using Eqs. (142) and (143):

~

Q = (1 2 ~ c)

Q

u

1 ~ c

Q

b

(168)

where

Q

u

and

Q

b

are de®ned as:

Q

u

=

_11

21

Q

P

c

(Quc = 0)dQ

Q

b

=

_11

21

Q

P

c

(Quc = 1)dQ

P

c

(Quc) is the pdf of Q for the given value c of the progress

variable (conditional pdf, see Section 6.4.1).

The components ~ u

k

of the mean velocity vector ~ u may be

written as a linear combination of their conditional fresh and

burnt gases averages:

~ u

i

= (1 2 ~ c) u

u

i

1 ~ c u

b

i

(169)

Then:

r

¯

u

//

i

c

//

= r( ¯ u

i

c 2 ~ u

i

~ c) = r( ~ c u

b

i

2 ~ u

i

~ c) = r ~ c(1 2 ~ c)( u

b

i

2 u

u

i

)

(170)

which is the scalar turbulent ¯ux, generally modeled using a

gradient assumption:

r

¯

u

//

i

c

//

= 2

m

t

S

c

2~ c

2x

i

_ _

(171)

The two expressions (170) and (171) may describe opposite

¯uxes: consider a left-traveling one dimensional turbulent

¯ame, because of thermal expansion the conditional velo-

city in the burnt gases, u

b

i

; is expected to be larger than the

conditional velocity in the fresh gases, u

u

i

: According to Eq.

(170), the turbulent ¯ux,

¯

u

//

i

c

//

is expected to be positive. On

the other hand, as the mean progress variable gradient is also

positive, Eq. (171) leads to a negative value of

¯

u

//

i

c

//

: This

situation, called `counter-gradient turbulent transport', is a

key point of the BML analysis and will be further discussed

in Section 8.

The Reynolds stresses

¯

u

//

i

u

//

j

may also be decomposed

using the same formalism:

¯

u

//

i

u

//

j

= (1 2 ~ c)u

/

i

u

/

j

u

1 ~ cu

/

i

u

/

j

b

1 ~ c(1 2 ~ c)( u

b

i

2 u

u

i

)( u

b

j

2 u

u

j

)

............,,............

intermittency

(172)

where one may note a weighted mean between the Reynolds

stresses in the fresh

_

u

/

i

u

/

j

u

_

and in the burnt gases

_

u

/

i

u

/

j

b

_

representative of turbulent motions. The additional term

represents the intermittency between fresh and burnt gases.

7.3.6. Extensions to partially premixed combustion

Some attempts have been made to extend BML modeling

concepts to partially premixed combustion, i.e. when reac-

tants are not perfectly premixed before burning [118]. Two

dif®culties are then encountered. First, in the de®nition of

the progress variable c (Eq. (35)), unburnt and burnt gas

temperatures and fuel mass fractions T

u

, T

b

, Y

u

F

and Y

b

F

are

no longer constant and depend on local equivalence ratio. A

balance equation may be still derived (see, for example, Ref.

[119]) but it incorporates additional terms and is not obvious

to close, especially when rich and lean zones coexist in the

same ¯ow ®eld. Mixing should also be taken into account

through a mixture fraction Z. The challenge is then to model

the joint probability density function

P(c

p

; Z

p

; x; t): Lahjaily

et al. [118] write:

P(c

p

; Z

p

; x; t)

= a(x; t)d(c

p

)

P

u

(Z

p

; x; t) 1b(x; t)d(1 2c

p

)

P

b

(Z

p

; x; t)

1g(x; t)F(c

p

; Z

p

; x; t)H(Z

p

2Z

min

)

..............,,..............

burning zones

1g

m

(x; t)F

m

(c

p

; Z

p

; x; t)[1 2H(Z

p

2Z

min

)]

...................,,...................

mixing without burning

(173)

where

P

u

(Z

p

; x; t) and

P

b

(Z

p

; x; t) represent the mixture frac-

tion distributions in fresh and burnt gases respectively. F

and F

m

are the distributions within the reaction zones,

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 228

Fig. 18. Correlation 2r

/

c

/

= r as a function of the Favre averaged progress variable ~ c for various values of the heat release factor t, assuming a

bimodal distribution (c = 0 or c = 1) of c (Eq. (166)).

assuming to be ¯amelets, and within the non-reactive

mixing zone between fresh and burnt gases. Combustion

is supposed to occur when Z

p

$ Z

min

(H is the Heaviside

function verifying H(z , 0) = 0 and H(z $ 0) = 1). Under

BML assumptions, the two last contributions are neglected

(g p1 and g

m

p1) and dilution effects are incorporated

through

P

u

(Z

p

; x; t) and

P

b

(Z

p

; x; t):

7.4. Models based on the ¯ame surface area estimation

7.4.1. Introduction

Several ¯ame surface density models are now discussed.

Their derivation and their histories differ, but theyare all based

on similar concepts, described in Section 6.3.2. These models

assume that the chemical reaction occurs in thin layers

separating fresh gases from fully burnt ones (high Dam-

koÈhler number limit). The reaction zone may then be viewed

as a collection of laminar ¯ame elements called ¯amelets.

The ¯ame surface density is here introduced at the light of

experimental data from Refs. [120,121]. The experimental

burner is displayed on Fig. 19. A turbulent premixed

propane/air ¯ame is stabilized behind a small cylinder

(blockage ratio of 6%). Flow rates are about 35±100 g/s,

corresponding to inlet velocities between 10 and 30 m/s

(turbulence levels from 5 to 10%). Equivalence ratios f

are in the range 0.7±1.1. Velocity (laser Doppler veloci-

metry), CH and C

2

radical emission (reaction rate estima-

tion) and high-speed laser tomography (¯ame front

characteristics) measurements have been performed and

are described in Refs. [120±123].

In Fig. 20 (half burner), ¯ame surface density pro®les are

plotted as a function of the transverse coordinate for various

downstream locations and for two equivalence ratios.

7.4.2. Algebraic expressions for the ¯ame surface density S

Assuming intermittency between fresh and burnt gases

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 229

Fig. 19. Experimental burner. A propane/air premixed ¯ow is

injected in a rectangular burner through a grid. The turbulent

¯ame is stabilized behind a small cylinder (blockage ratio of 6%)

[120,121].

Fig. 20. Transverse ¯ame surface density (S) pro®les (m

21

) plotted as a function of the transverse location for various downstream locations

(mm downstream the rod). (a) f = 0:78; (b) f = 0:9: Flow rate: 35 g/s [121].

(Section 7.3), Bray, Moss, Libby and their coworkers have

proposed to describe the mean reaction rate

_ v as the product

of a ¯ame crossing frequency n and a local reaction rate per

¯ame crossing, w

F

:

_ v = w

F

n (174)

Because c(t) may be viewed as a telegraphic signal (Fig. 16),

this crossing frequency is derived from a statistical analysis

of the telegraph equation, leading to:

n =

2 c(1 2 c)

^

T

(175)

where

^

T is time scale for the ¯uctuations of c.

This analysis is attractive because the crossing frequency

n may be easily obtained in experiments, for example from

time-resolved local temperature measurements (thermo-

couple). On the other hand, the ¯ame surface density S

and the local reaction rate per ¯ame crossing (w

F

) are not

obvious to estimate. Eq. (175) is generally closed estimating

^

T from a characteristic turbulent time t

t

. The reaction rate

per crossing ¯ame, w

F

(Eq. (174)), is usually modeled as:

w

F

=

r

0

S

L

d

t

=t

t

(176)

where r

0

is the unburnt gases density, S

L

and d

l

are respec-

tively the speed and the thickness of the laminar ¯ame. The

¯ame transit time t

t

measures the averaged time spend by a

point in the ¯ow to cross a ¯ame front and corresponds to the

mean transition time between c = 0 and c = 1 levels of the

progress variable, as shown in Fig. 21 (in practice, the c-

signal is not exactly bimodal).

This model has been latter rewritten in term of ¯ame

surface density, leading to the algebraic expression [124]

S = g

c(1 2 c)

s

y

L

y

=

g

s

y

L

y

1 1t

(1 1t~ c)

2

~ c(1 2 ~ c) (177)

where g is a constant of order unity. s

y

is a ¯amelet orienta-

tion factor measuring the mean angle of the instantaneous

¯ame front with the c surface and assumed to be an universal

model constant (s

y

< 0:5): L

y

is a ¯ame front wrinkling

length scale and Eq. (165) has been used to replace c as a

function of ~ c: A submodel is required to describe the wrink-

ling length scale L

y

, generally assumed to be proportional to

the integral length scale l

t

:

L

y

= C

l

l

t

S

L

u

/

_ _

n

(178)

where C

l

and n are two constants of the order of unity [75].

According to Gouldin [125,126], the ¯amelet orientation

factor s

y

is directly linked to the vectors normal to the

instantaneous ¯ame front, n, and to the mean ¯ame brush,

n

p

(see Section 6.3.3):

1

s

y

=

1

un

p

´nu

_ _

s

(179)

Despite of the same normal vectors n and n

p

involved, the

¯amelet orientation factor s

y

(Eq. (179)) and the ¯ame

surface wrinkling factor J (Eq. (97)) are different quanti-

ties. These relations, combined with Eqs. (100) and (177),

may relate the wrinkling length scale L

y

to the thickness d

B

of the mean ¯ame brush, under the assumption of a in®nitely

thin ¯ame front:

S = g

c(1 2 c)

s

y

L

y

= Ju7 cu < aJ

c(1 2 c)

d

B

(180)

where a is a model constant of the order of unity. Then:

L

y

<

g

a

_ _

1

un

p

´nu

_ _

s

n

p

´ n ( )

s

d

B

(181)

Assuming that a, g and s

y

are constant, d

B

/J measures the

wrinkling length scale L

y

of the ¯ame front.

The agreement between this BML model and the exper-

imental data is very good, as displayed on Fig. 22 where the

ratio S=[ c(1 2 c)]; corresponding to g=s

y

L

y

and assumed to

be constant, is displayed. Nevertheless, a submodel is

required for the wrinkling length scale which increases

with the downstream location in our experiment.

Estimating the local reaction rate per unit ¯ame area

_

V

c

from the laminar ¯ame speed S

L

(

_

V

c

= r

0

S

L

where r

0

is the

fresh gases density) the mean reaction rate becomes, when

n = 1 :

_ v = r

0

g

s

y

C

l

u

/

l

t

c(1 2 c) = r

0

g

s

y

C

l

u

/

l

t

(1 1t) ~ c(1 2 ~ c)

(1 1t~ c)

2

(182)

As t

t

= l

t

=u

/

is a turbulence characteristic time scale, the

mean reaction rate

_ v is proportional to the intermittency

between fresh and burnt gases, determined from ~ c(1 2 ~ c)

or c(1 2 c); and inversely proportional to t

t

. The physical

analysis leading to the Eddy-Break-Up model is recovered

and a similar expression for the reaction rate is found. The

previous expression is slightly different than the one

proposed in Section 7.3.3. The discrepancies are easily

explained by the crude model used for the scalar dissipation

of the progress variable to derive Eq. (162).

In Eq. (182), u

/

=l

t

is generally modeled from the

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 230

Fig. 21. De®nition of the ¯ame transit time t

t

is the ¯ame crossing

frequency BML model.

turbulence k±1 model as 1=k: The ITNFS ef®ciency function

G

k

[127] may be introduced in the turbulent time scale to

account for the reduced ability of small turbulent structures

to wrinkle the ¯ame front. ITNFS is one of the closure

schemes for the ¯ame surface density balance equation

(see Section 7.4.3 and Table 5). The mean reaction rate is

then [118,128,129]:

_ v = ar

0

G

k

l

t

d

t

;

u

/

S

L

_ _

1

k

c(1 2 c) (183)

where a is a model constant and the ef®ciency function G

K

depends on the length scale (l

t

/d

l

) and the velocity (u

/

=S

l

)

ratio comparing the turbulence and the laminar ¯ame char-

acteristics. The ef®ciency function has been ®tted from DNS

data [127,130]:

log

10

(G

k

) = 2

1

s 10:4

exp[2(s 10:4)]

1

_

1 2exp[2(s 10:4)]) s

1

u

/

S

L

_ _

s 20:11

_ _

(184)

where

s = log

10

l

t

d

l

_ _

and

s

1

u

/

S

L

_ _

=

2

3

1 2

1

2

exp 2

u

/

S

L

_ _

1=3

_ _ _ _ (185)

When the length scale ratio l

t

/d

t

tends towards zero, G

K

also

decreases, reducing the effective ¯ame strain rate, as

displayed on Fig. 23. G

K

only slightly depends on the vel-

ocity ratio u

/

=S

L

: The ef®ciency function does not reach a

constant level when l

t

/d

l

increases since increasing l

t

/d

l

,

keeping u

/

=S

L

constant, corresponds to an implicit increase

of the turbulent Reynolds number Re.

Because the ef®ciency function G

K

tends to counter-

balance the known trend of Eddy-Break-Up modeling to

overestimate the mean reaction rate in highly strained

regions, this simple approach improves the accuracy of the

numerical predictions.

The BML model proposes a simple algebraic expression

to estimate the ¯ame surface density s and the correspond-

ing reaction rate, but Bray, Moss, Champion and Libby

have mainly focused their attention on a careful description

of the turbulent ¯uxes using the balance equations for the

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 231

Fig. 22. S=( c(1 2 c)) transverse pro®les in mm

21

corresponding to g=(s

y

L

y

) in the Bray, Champion, Libby model (Eq. (177)), plotted as the

function of the mean progress variable c for several locations downstream from the rod: (a) f = 0:78; (b) f = 0:9 [121].

Reynolds stresses r

¯

u

//

i

u

//

j

and the scalar ¯uxes r

¯

u

//

i

c

//

to take

into account the possible occurrence of counter-gradient

transport and ¯ame turbulence generation [131,132].

The ¯ame surface density may also be derived from frac-

tal theories, leading to [133]

S =

1

L

outer

L

outer

L

inner

_ _

D22

(186)

where L

inner

and L

outer

are respectively the inner and outer

cut-off length scales (the ¯ame surface is assumed to be

fractal between these two scales). D is the fractal dimension

of the ¯ame surface. The cut-off scales are generally esti-

mated from the turbulence Kolmogorov l

k

and the integral l

t

length scales.

7.4.3. Flame surface density balance equation closures

The previously described balance equations for S (Eqs.

(89) and (90)) are unclosed and require modeling. In Table

5, various closures found in the literature are compared

where S

1

is the strain rate acting on the surface and induced

by the mean ¯ow ®eld. S

2

is the strain rate due to the turbu-

lent motions (Eq. (94)) and the third source term, S

3

, occurs

only in the derivation proposed in Ref. [117]. D describes

the consumption of ¯ame area. The modeled balance equa-

tion is rewritten as:

2S

2t

17´(

~

US) = 7´

n

t

s

S

7S

_ _

1S

1

1S

2

1S

3

2D (187)

In this expression, the turbulent ¯ux of ¯ame surface

density is expressed using a classical gradient assumption,

n

t

is the turbulent viscosity and s

S

a ¯ame surface turbulent

Schmidt number. Five main closures are summarized:

² The CPB model [134] is derived from the exact transport

equation for the ¯ame surface density. The strain rate due

to the turbulent ¯uctuations is estimated from the time

scale

n=1

_

of the Kolmogorov structures. The turbulent

strain rate is probably overestimated. Despite the fact that

the Kolmogorov structures contain the highest energy,

their lifetime is too short (because of viscous effects) to

actually affect the ¯ame front [65].

² The coherent ¯ame model (CFM), developed by Candel

and his coworkers following the initial work of Marble of

Broadwell [84]. Three versions are presented in Table 5.

In the initial version (CFM1), the strain rate due to the

turbulent ¯uctuations is estimated from the character-

istic time of the integral length scale (k/1). In the two

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 232

Table 5

Comparison of source (S

i

) and consumption (D) terms in the ¯ame surface density balance equation in different turbulent premixed combustion

models (details in text). k and 1 denote, respectively, the turbulent kinetic energy and its dissipation. Re is the turbulent Reynolds number. a

0

,

b

0

, g, l, C

A

, a, d, C, E and K are model constants, G

k

is the ef®ciency function in the ITNFS model [127] and depends on the length scale (l

t

/d

l

)

and the velocity (u

/

/S

L

) ratios comparing the turbulence and the laminar ¯ame characteristics. In the Choi model (CH) [137], u

/

denotes the rms

turbulent velocity and l

tc

is an arbitrary length scale introduced for dimensional consistency and combined to a

0

as a single arbitrary constant

Model S

1

= k

m

S S

2

= k

t

S S

3

D

CPB [134] A

ik

2 u

k

2x

i

S a

0

C

A

_

1

n

_

1=2

S a

0

S

L

2 1e

2aR

3(1 2 c)

S

2

R =

(1 2 c)1

SS

L

k

CFM1 A

ik

2 u

k

2x

i

S a

0

1

k

S b

0

S

L

1C

k

_

1 2 c

S

2

CFM2-a A

ik

2 u

k

2x

i

S a

0

G

K

1

k

S b

0

S

L

1C

k

_

1 2 c

S

2

CFM2-b [130] A

ik

2 u

k

2x

i

S a

0

G

K

1

k

S b

0

S

L

1C

k

_

c(1 2 c)

S

2

MB [117] E

u

/

i

u

/

k

k

2 u

k

2x

i

S a

0

Re

_

1

k

S

F

S

L

1

k

u

/

i

c

/

2 c

2x

i

b

0

S

L

Re

_

c(1 2 c)

_

1 1d

S

L

k

_

_

2g

S

2

CD [136] a

0

l

1

k

S for k

t

# a

0

K

S

L

d

L

b

0

S

L

1 2 c

S

2

CH1 a

0

_

1

15n

_

1=2

S b

0

S

L

c(1 2 c)

S

2

CH2 [137] a

0

u

/

l

tc

S b

0

S

L

c(1 2 c)

S

2

succeeding formulations (CFM2), the expression of the

turbulent strain rate acting on the ¯ame front is improved

from direct numerical simulations and multi-fractal

analysis (ITNFS model, Eq. (184) [127]). The destruction

term differs in these two last formulations (CFM2a and

CFM2b).

² The MB model [117] is based on an exact equation for the

scalar dissipation rate x :

r ~ x = rD

2c

//

2x

i

2c

//

2x

i

(188)

assuming a constant density r. The transport equation is

rewritten as a ¯ame surface density transport equation

under the ¯amelet assumption (see Section 6.5, Eq.

(116)). This approach leads to a different expression for

the source term S

1

and an additional source term S

3

is

found. In a ®rst analysis, this term S

3

, which does not

depend on the available ¯ame surface density S; might

be viewed as an ignition term [135] involving a gradient

of the mean fuel mass fraction or of the mean temperature

(i.e. fresh gases are ignited by heat transfers). However,

this analysis does not hold because the ¯ame surface

density balance equation is derived assuming an

established ¯ame surface. In fact, as shown below, S

3

corresponds to an anisotropic contribution of the tur-

bulent strain rate a

T

(see Eq. (205)). S

3

seems to be negli-

gible in practical simulations, at least to describe the

¯ame front propagation in a homogeneous and isotropic

turbulent ¯ow ®eld.

² The CD model [136] is similar to the ®rst version of the

coherent ¯ame model (CFM1). An additional term is

proposed to take into account ¯ame extinction under

excessively high strain rates. Such a term was tested in

the coherent ¯ame model but the choice of the critical

strain rate is somewhat arbitrary. Moreover, this critical

strain rate cannot be deduced from planar strained

laminar ¯ame studies because, due to curvature and

unsteady effects, a ¯ame is able to sustain higher strain

rates than expected.

² The CH model [137] has been devised for spark-ignited

engines to recover experimental data obtained in a closed

vessel [138]. The consumption term D is similar to the

one proposed in CFM2-b model whereas two formula-

tions of the strain rate induced by turbulent motions are

proposed. The ®rst expression (CH1) corresponds to the

closure in CPB model, based on the Kolmogorov turbu-

lent time scale. In CH2, the strain rate is only propor-

tional to the turbulence intensity u

/

and an arbitrary

length scale, l

tc

, is incorporated in the model constant.

Despite these comments, all these closures exhibit strong

similarities. For example, the consumption term D is always

proportional to S

2

. A comparison between these models to

predict turbulent ¯ame speed S

T

may be found in Ref. [130].

In the case of a statistically one-dimensional turbulent ¯ame

propagating in a frozen turbulence, a KPP (Kolmogorov±

Petroski±Piskunov) analysis was used to analytically deter-

mine the turbulent ¯ame speed S

T

as a function of the model

parameters.It was found that only the CFM-2 formulation is

able to predict the so-called bending phenomenon, where

the turbulent ¯ame speed decreases before the occurrence of

¯ame extinction when the turbulence level increases, as

experimentally evidenced [110].

A recent work, [135] has compared CPB, CFM1, MB and

CD models to predict a turbulent premixed jet ¯ame. The

CD predicts extremely high temperatures whereas CFM1,

MB and CPB provide reasonable predictions of mean vel-

ocities and temperatures. A slight overestimate (respectively

underestimate) of temperature is pointed out for CPB

(respectively MB) and is probably due to the expression

for the strain rate. The MB closure [117] is found to be

more sensitive to the inlet turbulent quantities than CFM1

but CFM2 models, incorporating an ef®ciency function

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 233

Fig. 23. ITNFS ef®ciency function G

K

(Eq. (184)) as a function of the length scale ratio l

t

/d

l

for several values of the velocity ratio u

/

/S

L

(0.1;

1.0; 10; 11).

including the inability of small vortices to wrinkle the ¯ame

surface, have not been tested. Flame surface density models

are extended to non-isenthalpic ¯ows (premixed reactants

injected in a co-¯ow of air [139]). A pioneering work has

been ®rst conducted [118] extending the algebraic ¯ame

surface density BML model to non-isenthalpic ¯ows,

using similar concepts. A presumed pdf for a mixture frac-

tion Z, determined from mean and rms values of Z, is

incorporated to account for dilution phenomena.

7.4.4. Analysis of the ¯ame surface density balance equation

Following the description of the exact balance equation

for the ¯ame surface density S (Section 6.3.2) and the

summary of the most popular closures (Section 7.4.3), the

aim of this section is to carefully analyze this balance

equation. This analysis may be based on direct numerical

simulations [17] or on experimental data [120,121].

Starting from the `propagative' form (Eq. (90)):

2S

2t

17´( ~ uS) 17´(ku

//

l

s

S)

= (7´ ~ u 2knnl

s

: 7 u)

........,,........

A

T

S 1k 7´u

//

2nn : 7u

//

l

s

........,,........

a

T

S

27´[kwnl

s

S]

....,,....

propagation

1kw7´nl

s

S

....,,....

curvature

(189)

each unclosed term may be investigated as follows:

7.4.4.1. Turbulent transport. The turbulent ¯ux of ¯ame

surface density is generally expressed using gradient

transport:

ku

//

l

s

S = 2

n

t

s

S

7S (190)

Bidaux and Bray (1994, unpublished work, see for example

[75,140]) have shown from a simple BML-type approach

that the turbulent ¯uxes of the mean progress variable ( ~ c)

and of the ¯ame surface density (S) are closely related.

Assuming that the conditional velocity on the ¯ame

surface, k ~ ul

s

is a linear function of the conditional fresh

( u

u

) and burnt gases velocities:

k ~ ul

s

= u

u

1c

p

( u

b

2 u

u

) (191)

where c

p

denotes the c-level chosen to de®ne the ¯ame front.

The BML relation (Eq. (169)):

~ u = (1 2 ~ c) u

u

1 ~ c u

b

(192)

leads to:

ku

//

l

s

= kul

s

2 ~ u = (c

p

2 ~ c)( u

b

2 u

u

) (193)

From Eq. (170) the ¯ame surface density turbulent ¯ux

becomes:

k

¯

u

//

l

s

S =

(c

p

2 ~ c)

~ c(1 2 ~ c)

¯

u

//

c

//

S (194)

This relation, con®rmed by direct numerical simulations

[140], shows that turbulent ¯uxes of ~ c and S are closely

related. A counter-gradient turbulent transport will be

observed simultaneously for these two scalar ®elds.

7.4.4.2. Strain rate induced by the mean ¯ow ®eld, A

T

. The

only unclosed quantities in the strain rate due to the mean

¯ow ®eld, A

T

, are the orientation factors knnl

s

. Following

[134], the vector normal to the ¯ame front, n may be split

into a mean component, M, and a ¯uctuation, m:

n = M1m with knl

s

= M and kml

s

= 0 (195)

Then, the orientation factors become:

knnl

s

= MM1kmml

s

(196)

Using the de®nition nu7cu = 27c and assuming an

in®nitely thin ¯ame front leads after averaging to [134,141]:

knl

s

S = MS = 27 c (197)

where c and ~ c are related using the BML relation r ~ c = r

b

c

(Eq. (138)). Then, only the ¯uctuation cross products kmml

s

remain unclosed. Several closure schemes have been

proposed:

² In Ref. [134] one assumes an isotropic distribution of the

¯uctuating components of the normal vector n:

km

i

m

j

l

s

=

d

ij

3

(1 2M

k

M

k

) (198)

² In Ref. [117], a relation between orientation factors and

Reynolds stresses is indirectly proposed:

kn

i

n

j

l

s

=

¯

u

//

i

u

//

j

2k

(199)

where k is the turbulent kinetic energy.

² From experimental data [120], a much more complicated

model is proposed that will not be described here because

its practical implementation is probably not so easy.

These authors have also shown that the isotropic

assumption made in Ref. [134] is clearly wrong for the

turbulent ¯ame stabilized downstream a small rod. On the

other hand, a slight modi®cation of Ref. [117] leads to

very good results [121]:

kn

i

n

i

l

s

=

k±i

¯

u

//

k

2

4k

; kn

i

n

j±i

l

s

=

¯

u

//

i

u

//

j

2k

(200)

These orientation factors have very important effects and

may lead to surprising results. In the ¯ame-holder stabilized

turbulent ¯ame investigated in Ref. [121], the main velocity

gradient is the transverse gradient of the mean axial velocity,

corresponding to the mixing layer shear stress. But, because of

the low value of the corresponding orientation factor, its

contribution to the strain rate A

T

induced by the mean ¯ow

®eld is not that important. The strain rate A

T

is dominated by

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 234

the axial gradient of the mean axial velocity, induced by the

thermal expansion due to the combustion heat release, as

illustrated in Fig. 24. Notice also that the strain rate A

T

cannot be reduced to the simpli®ed expression ®rstly

proposed [84]:

A

T

= u7

~

Uu (201)

Recent works [142] have shown that u7

~

Uu may probably be

viewed as a model for the total strain rate A

T

1a

T

and not only

for the strain rate A

T

due to the mean ¯ow. This ®nding

explains why, in previous versions of the coherent ¯ame

model where the orientation factors kn

i

n

j

l

s

were not

incorporated, the mean strain rate A

T

was not included

because the simple expression (201) clearly overestimates A

T

.

7.4.4.3. Strain rate a

T

due to turbulent motions. The source

term due to the strain rate a

T

is:

a

T

S = k7´u

//

2nn : 7u

//

l

s

S = d

ij

2u

//

i

2x

j

2n

i

n

j

2u

//

i

2x

j

_ _

s

S

(202)

In most models, this term is generally assumed to be

proportional to the inverse of a turbulent time scale, either

the Kolmogorov time scale (CPB model) or the integral time

scale 1/k (CFM, MB and CD models). This turbulent time

may corrected with the ef®ciency function, G

k

, of the ITNFS

closure (Eq. (184)). Nevertheless, a

T

is always modeled

being isotropic despite the orientation factors nn occurring

in expression (202).

Using the previous splitting of the normal vector to the

¯ame front, n, combined with the geometrical relation

(197), leads to:

a

T

S =

_

d

ij

2u

//

i

2x

j

_

s

S 2

1

S

_

2u

//

i

2x

j

_

s

2 c

2x

i

2 c

2x

j

1

_

m

j

2u

//

i

2x

j

_

s

2 c

2x

i

1

_

m

i

2u

//

i

2x

j

_

2 c

2x

j

2

_

m

i

m

j

2u

//

i

2x

j

_

s

S (203)

This interesting relation decomposes the source term due to

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 235

Fig. 24. (a) Transverse pro®les of the mean velocity gradients and the corresponding strain rate A

T

. (b) Components of the source term A

T

S due

to the strain rate induced by the mean ¯ow ®eld. Data are plotted as a function of the transverse coordinate y for x = 80 mm and f = 0:9 [121].

the strain rate induced by turbulent motions, a

T

S, into two

parts: a contribution depending only on turbulent character-

istics that may be assumed to be isotropic (®rst and last

terms) and an anisotropic contribution, where the mean

¯ame surface orientation occurs through the gradient of

the mean progress variable (second, third and fourth

terms). A priori, this last part is not proportional to the

available ¯ame surface density S.

When velocity and normal vector ¯uctuations are

supposed uncorrelated (this assumption is well veri®ed in

direct numerical simulation data), the previous expression

may be reduced using:

m

j

2u

//

i

2x

j

_ _

s

= km

j

l

s

2u

//

i

2x

j

_ _

s

= 0 (204)

from the m

j

de®nition. Then:

a

T

S = d

ij

2u

//

i

2x

j

_ _

s

S 2 m

i

m

j

2u

//

i

2x

j

_ _

s

S2

...............,,...............

isotropic contribution

1

S

2u

//

i

2x

j

_ _

s

2 c

2x

i

2 c

2x

j

........,,........

anisotropic contribution

(205)

This simple analysis explains the third source term, S

3

found

in Ref. [117] from the derivation of a balance equation for

the scalar dissipation rate ~ x of the progress variable c. This

term S

3

corresponds to an anisotropic contribution depend-

ing on 7 c and on the strain rate due to turbulent motions.

Assuming a gradient type closure for the turbulent transport

of the fresh gases, S

3

is proportional to 7 c´7 c and appears as

a model for the anisotropic contribution of the turbulent

strain rate. Nonetheless, Mantel and Borghi [117] have

shown that this term is negligible in numerical simulations

of a ¯ame propagating in an homogeneous and isotropic

turbulent ¯ow ®eld. This ®nding is probably questionable

in other con®gurations and further investigations are

required.

7.4.4.4. Propagation and curvature terms. These two terms

are analyzed together because they derive from the laminar

¯ame propagation and are related to the ¯ame front

displacement speed w. The modeling of these terms

requires the description of the displacement speed w that

may have values far from the laminar ¯ame speed S

L

[78].

This stands as the main dif®culty of the propagative

approach for the ¯ame surface density balance equation

because various effects are incorporated in w (strain rate,

curvature effects, etc.).

Assuming a constant displacement speed w equal to the

laminar ¯ame speed S

L

, the normal propagation term

becomes:

7´[kwnl

s

S] = S

L

7´[knl

s

S] = S

L

7´[27 c] = 2S

L

7

2

c (206)

where the geometrical relation (197) is used.

This term is generally neglected in models (see Table 5).

Experimental data from Refs. [120,121] in a V-shape turbu-

lent ¯ame show that it is not always negligible and is of the

same order as the curvature term.

The propagation/curvature term becomes:

kw7´nl

s

S = S

L

k7´nl

s

S (207)

where the only unknown is the mean ¯ame front curvature

k7´nl

s

The curvature is positive when the ¯ame front is

convex towards the fresh gases, which is, a priori, the case

when the mean progress variable c is close to zero (see Fig.

25). On the other hand, curvatures are probably negative

(¯ame convex towards the burnt gases) when c < 1: As-

suming that the mean curvature is of the order of the inverse

of the wrinkling length scale L

y

, we have:

lim

c-0

k7´nl

s

=

1

L

y

and lim

c-1

k7´nl

s

= 2

1

L

y

Then, a simple linear model may be proposed:

k7´nl

s

<

c

p

2 c

L

y

(208)

Replacing L

y

by its value as a function of the ¯ame surface

density S from BML modeling (Eq. (177)) leads to the

following model for the curvature term in the S-equation:

kw7´nl

s

S < bS

L

c

p

2 c

c(1 2 c)

S

2

(209)

where b is a model constant.

This term is positive in the fresh gas side of the turbulent

¯ame brush and becomes negative in the burnt gas side. This

trend is in agreement with the ®ndings of direct numerical

simulations [17] and experimental measurements [120,121],

as displayed in Fig. 26. The expression differs from classical

closure where kw7´nl

s

is always negative (term D in

Table 5).

A linear relaxation is retained to model the mean curva-

ture k7´nl

s

(Eq. (208)). This type of closure is common in

turbulent reacting ¯ows and possesses similarities with the

relaxation model used for the scalar dissipation rate (Section

9.2) or for the pdf balance equation (Section 9.8.2). The

links between these closures are further discussed in Section

9.8.5.

The previous analysis was based on both a theoretical

analysis of the exact S-balance equation and experimental

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 236

Fig. 25. Flame front curvature analysis.

data [120,121]. An improved version of the closed balance

equation may be proposed:

2S

2t

17´( ~ uS) 17´(ku

//

l

s

S)

= (7´ ~ u 2knnl

s

: 7~ u)S 1G

k

u

/

S

L

;

l

t

d

l

_ _

1

k

S 1S

L

7

2

c

1bS

L

(c

p

2 c)

S

2

c(1 2 c)

(210)

where G

k

is the ITNFS ef®ciency function [127]. Orientation

factors, kn

i

n

j

l

s

may be described using a formulation

previously discussed. The turbulent ¯ux ku

//

l

s

S is described

using a classical gradient expression.

7.4.5. Flame stabilization modeling

The ¯ame surface density ®elds obtained from laser

tomography and the mean reaction rate estimated from

CH radical emission are compared in Fig. 27. According

to Eq. (85), reaction rate and ¯ame surface density are

roughly proportional, excepted close to the stabilization

rod, the ¯ame surface density S is high, whereas the mean

reaction rate remains low. In this zone, fresh and burnt gases

are separated by an interface (high surface densities) where

combustion has started but is not yet fully established. This

®nding displays one of the dif®culties of ¯ame surface

density models. In their present formulations, these models

are unable to describe the ¯ame stabilization because the S-

equation (Eq. (189)) is derived assuming that the ¯ame does

exist. All source terms in this equation are proportional to S

or to S

2

and the equation cannot generate ¯ame surface

when there is no initial ¯ame surface. In addition, an initia-

tion effect must also be incorporated to account for ignition

time delay in the local reaction rate per unit surface, V

Ç

i

(Eq.

(85)), to recover the observed results.

7.4.6. A related approach: G-equation

As described in Section 6.3.1, the premixed turbulent

¯ame may be described using a level set approach. Most

of the modeling issues discussed above are then recast in

terms of G-equation and modeling for the turbulent burning

velocity S

T

[89].

8. Turbulent transport in premixed combustion

8.1. Introduction

Turbulent ¯uxes of the progress variable c, r

¯

u

//

i

c

//

; are

generally modeled using a gradient transport hypothesis as

for inert scalars (Section 7):

r

¯

u

//

i

c

//

= 2

m

t

Sc

t

2~ c

2x

i

_ _

(211)

where m

t

is the turbulent viscosity given by the turbulence

model and Sc

t

is a turbulent Schmidt number. Theoretical

[143] and experimental studies [19,20] have evidenced in

some turbulent premixed ¯ames counter-gradient turbulent

transport where the turbulent ¯uxes r

¯

u

//

i

c

//

and the mean

progress variable ~ c gradient, 2~ c=2x

i

; have the same sign in

some regions and cannot be approximated with Eq. (211).

This phenomenon is known as counter-gradient turbulent

transport or counter-gradient turbulent diffusion.

This surprising ®nding was explained by the work of

Bray, Moss and Libby discussed in Section 7.3. In their

formulation, the turbulent ¯uxes of the progress variable c

are directly connected to the conditional mean velocities in

fresh ( u

u

i

) and burnt gases ( u

b

i

) (Eq. (170) Section 7.3.5):

r

¯

u

//

i

c

//

= r ~ c(1 2 ~ c) u

b

i

2 u

u

i

_ _

(212)

Even though conditional velocities are not obvious quanti-

ties, this expression is useful to explain counter-gradient

turbulent transport: Because of the thermal expansion due

to combustion heat release, the burnt gas conditional vel-

ocity, u

b

i

; is likely to be greater than the fresh gas conditional

velocity, u

u

i

: Then, in opposition with the modeled expres-

sion (Eq. (211)), the turbulent ¯uxes of the progress variable

c have the same sign as the mean gradient (2~ c=2x

i

): Counter-

gradient transport also increases when the heat release factor

t, de®ned by Eq. (141), increases [144].

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 237

Fig. 26. Transverse pro®les of propagation (27´[knl

s

S]), curvature

(k7´nl

s

S) and combined (k7´nl

s

S 27´[knl

s

S]) terms (m

22

) in the

¯ame surface density balance equation plotted as a function of the

mean progress variable c: The consumption term modeled as 2S

2

/

(1 2 c) and the new proposed closure, 7

2

c 1b(c

p

2 c)S

2

= c(1 2 c)

where b = 0:4 and c

p

= 0:5 are also displayed. f = 0:90; x =

80 mm [121].

Fig. 27. Flame surface density (half top) and mean reaction rate,

estimated form CH radical emission (half bottom) ®elds are

compared for f = 0:90: Flame surface density data, extracted

from two different data sets, are not available from 30 to 70 mm

downstream the rod [121].

The various versions of this BML model propose an alge-

braic closure for the mean reaction rate of the progress

variable c, but focus attention on the scalar turbulent trans-

port (closure schemes for the turbulent ¯uxes r

¯

u

//

i

c

//

balance

equations), whereas other models may lead to more sophis-

ticated reaction rate formulations, retaining a simple gradi-

ent closure (Eq. (211)) for these ¯uxes.

The occurrence of counter-gradient turbulent transport

have been analyzed using direct numerical simulation

(DNS) [140]. The results demonstrate the power of DNS

to help in the modeling of turbulent combustion. This

section is devoted to these results, obtained from a direct

solution without any closure models, of the instantaneous

balance equations.

8.2. Direct numerical simulation analysis of turbulent

transport

8.2.1. Introduction

Gradient and counter-gradient turbulent scalar transport

was observed in DNS of premixed ¯ame/three-dimensional

turbulence interactions [145,146]. Then, simulations of two-

dimensional ¯ame/turbulence interactions were reported

[140]. The reduced costs of 2D simulations allow the inves-

tigation of a large range of ¯ame and turbulence parameters.

In these simulations, a planar laminar premixed ¯ame is

®rstly superimposed on a homogeneous isotropic turbulent

¯ow ®eld (Fig. 28). The ¯ame front is progressively

wrinkled by turbulent motions, and the turbulence decays

in time. After a time of the order of the eddy-turnover time

of the largest turbulence structures, the ¯ame may be

assumed to be in equilibrium with the turbulent ¯ow ®eld

and relevant modeling information is extracted from DNS.

As the numerical con®guration is statistically one-

dimensional in the propagating direction, quantities such

as mean progress variable or mean turbulent ¯uxes may

be extracted from averaging in the perpendicular direction.

8.2.2. Results

Turbulent ¯uxes ru

//

c

//

extracted from the direct numer-

ical simulations conducted by TrouveÂ [145] (denoted CTR)

and Rutland [146] are displayed in Fig. 29. The ®rst data-

base clearly exhibits gradient turbulent transport, whereas

the second one corresponds to a counter-gradient situation.

The main discrepancy between the two databases lies in the

initial turbulence level which is higher in the CTR simula-

tion (u

/

0

=S

L

= 10) than in the Rutland database (u

/

0

=S

L

= 1):

Mean and conditional average velocities across the turbu-

lent ¯ame brush are displayed in Fig. 30 for the two DNS

simulations. These results lead to the following comments.

² As expected, in the Rutland database, the burnt gas con-

ditional velocity, u

b

; is higher than the fresh gas condi-

tional velocity, u

u

; leading to a counter-gradient turbulent

transport in agreement with expression (212).

² On the other hand, in the CTR database, the fresh gas

conditional velocity is higher than the burnt gas con-

ditional velocity. This result is ®rst surprising: u

b

is

expected to be larger than u

u

because of thermal expan-

sion due to the heat release. However, this observation is

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 238

Fig. 28. Numerical con®guration. A plane laminar premixed ¯ame

is initially superimposed to an homogeneous and isotropic turbu-

lence.

Fig. 29. Turbulent ¯ux ru

//

c

//

displayed as a function of the mean progress variable ~ c (bold lines) in the tri-dimensional direct numerical

simulations from CTR (Ð) and Rutland (- - -). Positive (respectively negative) values of ru

//

c

//

denote a counter-gradient (respectively

gradient) turbulent transport. The corresponding turbulent ¯uxes estimated using the BML expression (Eq. (212)) are displayed for comparison

using thin lines. Velocities are made dimensionless using the laminar ¯ame speed S

L

[140].

in agreement with expression (212) and a gradient turbu-

lent transport. Note that Eq. (212) overestimates the

turbulent ¯ux

¯

u

//

c

//

(see Fig. 29) because, due to the

¯ame thickness that must be resolved in the numerical

simulation, the progress variable c is not fully bimodal

(c = 0 or c = 1). In both cases, the Bray±Moss±Libby

model remains able to predict the turbulent ¯ux type

(gradient or counter-gradient).

The two-dimensional simulations conducted in Ref.

[140], referenced as CRCT in Fig. 31 have been used to

analyze the occurrence of counter-gradient turbulent

transport. The gradient turbulent transport is clearly

enhanced by an increase in the turbulence level u

/

/S

L

and

decreasing values of the heat release factor t. The

`increase', in terms of velocity ratio u

/

/S

L

, of the line

delimiting gradient and counter-gradient turbulent transport

when l

t

/d

l

decreases is due to the reduced ability of small

scale turbulence motions to wrinkle the ¯ame front. This

phenomenon has already been discussed [65] and included

in the ITNFS model [127]. Three- and two-dimensional

direct numerical simulations lead to very similar results

for this problem.

8.3. Physical analysis

A simple physical analysis is now reported to derive a

criterion predicting the occurrence of counter-gradient

turbulent diffusion [140].

Following Bidaux and Bray (1994) (unpublished work

already presented in the BML model context Section

7.4.4), turbulent ¯uxes of the progress variable c,

¯

u

//

i

c

//

;

are directly connected to the surface-averaged ¯uctuating

velocity, ku

//

i

l

s

(Eq. (194)). Thus, a model for

¯

u

//

i

c

//

may be

deduced from a model for ku

//

i

l

s

involving the conditional

fresh and burnt gases mean velocities:

ku

//

i

l

s

= ku

i

l

s

2 ~ u

i

= (c

p

2 ~ c) u

b

i

2 u

u

i

_ _

(213)

In the following, the ¯ow ®eld is assumed to be statistically

one-dimensional and only the turbulent transport in the

propagating direction,

¯

u

//

c

//

will be described. Our analysis

is based on the two limiting cases pictured in Fig. 32:

Low turbulence level: The ¯ame front remains smooth

and the velocity jump between fresh and burnt gases, u

b

2

u

u

; is determined primarily by thermal expansion and its

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 239

Fig. 30. Conditional velocities across the turbulent ¯ame brush displayed as a function of the mean progress variable ~ c in the Rutland (top) and

the CTR (bottom) databases. Favre averaged velocity, ~ u (Ð Ð

Ð), fresh gases conditional velocity, u

u

(´ ´ ´), burnt gases conditional velocity, u

b

(±)

and ¯ame front conditional velocity, kuls

(- - -). Velocities are made dimensionless using the laminar ¯ame speed S

L

[140].

value is close to the one obtained in a plane and laminar

¯ame ( u

b

2 u

u

< tS

L

): Eq. (213) becomes:

ku

//

l

s

= (c

p

2 ~ c)tS

L

(214)

High turbulence level: Due to strong viscous dissipation

of turbulent eddies in the hot burnt gas, the ¯ame front

motions are assumed to be dominated by the turbulence

properties taken upstream of the ¯ame.

² At the leading edge of the turbulent ¯ame (near ~ c = 0),

the ¯ame front is convected towards the fresh gas with a

mean velocity estimated by 2u

/

(see Fig. 32). Then,

u

b

2 u

u

< 2u

/

(215)

where u

/

denotes the rms velocity in the fresh gases.

² At the trailing edge of the ¯ame brush ( ~ c < 1); the ¯ame

front is convected by turbulent motions towards the burnt

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 240

Fig. 31. Premixed turbulent combustion diagram. The DNS ¯ame-¯ow conditions are plotted as a function of the velocity ratio, u

/

/S

L

, and

length-scale ratio, l

t

/d

L

. The Classical Klimov±Williams criterion and the criterion from Ref. [65] are given to show the domain of validity of

¯amelet combustion. Also plotted are the DNS conditions of the Rutland (CR, t = 2:3) and CTR (t = 3) simulations. As the turbulence is

decaying in the CTR simulation, CTR conditions are displayed as an almost vertical line. The symbols W (t = 3) and A (t = 6) correspond to

the CRCT DNS. In 2D DNS, the turbulence decay is smaller and is not represented. Filled (open) symbols denote gradient (counter-gradient)

turbulent diffusion. The transition criterion, N

B

; ts

L

=2au

/

= 1 (Eq. (220)), separating CGD (below) from GD (above) is plotted for (t = 3)

and (t = 6) [140].

Fig. 32. Two limiting cases: counter-gradient transport promoted by thermal expansion (left); gradient transport due to turbulent motions

(right).

gases with a mean speed estimated by 1u

/

(see Fig. 32):

u

b

2 u

u

< 1u

/

(216)

leading to the simple linear model:

ku

//

l

s

= 22(c

p

2 ~ c)au

/

(217)

where a is an ef®ciency function, similar to the ITNFS

model [127], accounting for the weak ability of small

scale turbulent motions to wrinkle and convect the

¯ame front. a is expected to be of order unity for large

turbulent length scales and vanishes when turbulent

eddies are too small to affect the ¯ame front. The factor

2 has been introduced assuming c

p

< 0:5:

Then, modeling ku

//

l

s

as a sum of these two contributions

leads to:

ku

//

l

s

= (c

p

2 ~ c)(tS

L

22au

/

) (218)

and the turbulent ¯ux becomes:

¯

u

//

c

//

= ~ c(1 2 ~ c)(tS

L

22au

/

) (219)

This simple model is well veri®ed in direct numerical simu-

lations [140] and has also been recovered when applying a

second order modeling (i.e. balance equations for turbulent

scalar ¯uxes) to stagnating ¯ames in the limit of small turbu-

lence intensities [131]. The turbulent ¯ux may be viewed as

the sum of two contributions acting in opposite directions,

one induced by turbulent motions and the other by thermal

expansion. Then, the turbulent transport is analyzed as

follows: for a suf®ciently high turbulence level, the ¯ame

is unable to impose its own dynamics to the ¯ow ®eld and

the turbulent transport is of the gradient type, as for any

inert scalar. On the contrary, when the turbulence level

remains low, the thermal expansion due to heat release

dominates and the ¯ame is able to impose its own dynamics

leading to a counter-gradient turbulent transport. Counter-

gradient turbulent diffusion occurs when

¯

u

//

c

//

is positive

and expression (219) may be used to derive a criterion deli-

neating gradient and counter-gradient regimes. The Bray

number:

N

B

=

tS

L

2au

/

(220)

is greater (respectively lower) than unity when a counter-

gradient (respectively gradient) turbulent transport is

expected. This criterion is well veri®ed by direct numerical

simulation results, as shown in Fig. 31. The ef®ciency func-

tion a has been estimated from DNS (Fig. 33). Recent

experimental results [147] have con®rmed these ®ndings.

8.3.1. Comments

² Following Fig. 31 and criterion (220), in practical appli-

cations turbulent transport may be counter-gradient, or

close to the transition between gradient and counter-

gradient regimes. In many cases, the heat release factor

t is about 5±7 leading to a transition between gradient

and counter-gradient situations when u

/

=S

L

is of the order

of 3. Nevertheless, the possible occurrence of counter-

gradient transport is generally neglected in modeling.

² The mean progress variable gradient may be estimated

as:

2~ c

2x

<

~ c(1 2 ~ c)

d

b

(221)

where a length scale d

b

characterizing the ¯ame brush is

introduced. Then, the gradient type contribution in Eq.

(219) corresponds to a Prandt±Kolmogorov turbulence

modeling:

2~ c(1 2 ~ c)d

b

au

/

< 2ad

b

k

_

2~ c

2x

(222)

where k is the turbulent kinetic energy and u

/

=

k

_

.

² Recent works [148,149] have reported regimes, corre-

sponding to low turbulence levels, where combustion

instabilities occur and wrinkle the ¯ame front, acting in

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 241

Fig. 33. DNS estimated ef®ciency function a (Eq. (21)) as a function of the length scale ratio l

t

/d

l

comparing the integral turbulent length scale

and the laminar ¯ame thickness [140].

an opposite direction than counter-gradient diffusion

reducing the ¯ame front wrinkling. These instabilities

should also be included in combustion models [150].

² Expression (219) has been derived to analyze the occur-

rence of counter-gradient transport, but is not suited for

numerical simulations. For instance, such a simpli®ed

formulation is unable to predict a gradient transport in

the vicinity of the leading edge ( ~ c < 0) always observed,

even in counter-gradient situations.

8.4. External pressure gradient effects

The previous analysis is completed by investigating the

effects of externally imposed pressure gradients on turbulent

premixed ¯ames. Counter-gradient turbulent transport was

®rst explained [19,151] by differential buoyancy effects of

pressure gradients on pockets of heavy and cold fresh gases

and on pockets of light and hot burnt gases. In many

combustion systems, ¯ames are ducted and submitted to

strong pressure gradients due to thermal expansion and lead-

ing to ¯ow accelerations. In Ref. [20], it is experimentally

shown that counter-gradient transports are enhanced in

ducted ¯ames.

In Refs. [148,152], the same type of DNS as [140] was

conducted, introducing an externally imposed pressure

gradient (in fact, due to technical reasons, a constant

acceleration). The main conclusions, displayed in Figs. 34

and 35, are:

² A favorable pressure gradient, i.e. a pressure decrease

from unburnt to burnt gases, is found to decrease the

¯ame wrinkling (see Fig. 34), the ¯ame brush thickness,

and the turbulent ¯ame speed S

T

(Fig. 35). It also

promotes counter-gradient turbulent transport.

² On the other hand, adverse pressure gradients tend to

increase the ¯ame brush thickness and turbulent ¯ame

speed (Fig. 35), and enhance classical gradient turbulent

transport. As proposed in Ref. [153], the turbulent ¯ame

speed is modi®ed by a buoyancy term linearly dependent

on both the imposed pressure gradient and the integral

length scale l

t

.

² A corrected Bray criterion (Eq. (220)) has been proposed

to account for the pressure gradient effects. We do not

give more details because this modi®ed criterion needs to

be validated and improved (length scale effects are not so

clear and various analyses are possible, see [152]).

All these results suggest that counter-gradient turbulent

transport should be expected in most ducted turbulent ¯ames

and some experimental observations are now reported.

8.5. Counter-gradient transportÐexperimental results

The V-shape turbulent ¯ame was described in Section

7.4.1. The thermal expansion modi®es the ¯ame dynamics

as is clearly apparent on high speed tomography ®lms

(scheme in Fig. 36, see also Ref. [121]): in the ®rst part

of the chamber (region 1), because of the rod wake, the

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 242

Fig. 34. Superimposed instantaneous temperature and vorticity

®elds at the same time. (a) No imposed pressure gradient±gradient

turbulent transport; (b) favorable imposed pressure gradient±

counter-gradient turbulent transport. A planar laminar ¯ame

separating fresh gases (left) from burnt gases is initially super-

imposed to an homogeneous and isotropic turbulent ¯ow ®eld

(u

/

0

=S

L

= 5) [152].

Table 6

Geometrical analysis of the scalar turbulent transport in a V-shaped premixed ¯ame (see Fig. 36). G and CG denote, respectively, a gradient and

a counter-gradient turbulent diffusion

Zone u

b

2 u

u

2~ c

2x

¯

u

//

c

//

v

b

2 v

u

(when y .0)

2~ c

2y

(when y .0)

¯

v

//

c

//

1 ,0 .0 G .0 ,0 G

2 .0 .0 CG .0 ,0 G

coherent structures embedding the ¯ame front turn clock-

wise (counter-clockwise) in the upper (lower) ¯ame sheet,

as in classical Von KaÂrmaÂn vortex streets, except for

their symmetry due to pressure waves. When the center-

line velocity increases, because of thermal expansion in

burnt gases (region 2), the upper (lower) coherent struc-

tures start to turn counter-clockwise (clockwise). This

phenomenon may be recast in terms of turbulent trans-

port using a simple geometrical analysis based on the

Bray±Moss±Libby relation (212) and summarized in

Table 6. Accordingly, the transverse turbulent ¯ux

¯

v

//

c

//

is always of gradient type, but the change in rota-

tion of the coherent structures corresponds to a transi-

tion between gradient and counter-gradient transport for

the downstream turbulent ¯ux

¯

u

//

c

//

: The turbulent ¯uxes

are of gradient type just behind the rod, as expected to

ensure the stabilization of the ¯ame, and becomes of

counter-gradient type further downstream.

8.6. To include counter-gradient turbulent transport in

modeling

Except for the work of Bray, Moss, Libby and their

co-workers [19,128,129,132,151], very few works have

been devoted to the effective modeling of counter-

gradient transport. The approach of Bray et al. is mainly

based on second order modeling, using balance equations

for the turbulent ¯uxes

¯

u

//

i

c

//

: These equations are easily

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 243

Fig. 35. Reduced turbulent ¯ame speed S

T

/S

L

plotted as a function of the reduced time t=t

f

= S

L

t=d

L

; where t

f

= d

L

=S

L

is a ¯ame time, for

different values of the externally imposed pressure gradient. (a) initial turbulent level u

/

0

=S

L

= 5; no pressure gradient (Ð) and increasing

favorable pressure gradient (´ ´ ´ and - - -); (b) initial turbulent level u

/

0

=S

L

= 2; no pressure gradient (Ð) and increasing adverse pressure

gradient (´ ´ ´ and - - -). Markers correspond to the Libby theory [152,153].

Fig. 36. A rough scheme of the coherent structures dynamics

observed using laser tomography [121].

derived from momentum and progress variable balance

equations:

2 r

¯

u

//

i

c

//

2t

...,,...

(I)

1

2 r ~ u

j

¯

u

//

i

c

//

2x

j

....,,....

(II)

= 2

2 r

¯

u

//

j

u

//

i

c

//

2x

j

....,,....

(III)

2 r

¯

u

//

i

u

//

j

2~ c

2x

j

....,,....

(IV)

2r

¯

u

//

j

c

//

2~ u

i

2x

j

....,,....

(V)

2 c

//

2 p

2x

i

..,,..

(VI)

2 c

//

2p

/

2x

i

..,,..

(VII)

2u

//

i

2J

k

2x

k

...,,...

(VIII)

c

//

2t

ik

2x

k

..,,..

(IX)

1 r

¯

u

//

i

_ v

..,,..

(X)

(223)

where the RHS terms correspond, respectively, to turbulent

transport of r

¯

u

//

i

c

//

(III), ~ c-gradient effects (IV), mean vel-

ocity gradient effects (V), the action of mean (VI) and ¯uc-

tuating (VII) pressure gradients, r

¯

u

//

i

c

//

turbulent dissipation

(VIII and IX) and reaction rate (X). This balance equation is,

of course, unclosed and each term may be extracted from

direct numerical simulations [132,140,152].

The discussion of the closure schemes for this equation is

beyond the scope of the present review and the reader may

®nd relevant information in Refs. [128,132,151]. Some

comments may be made:

² The

¯

u

//

i

c

//

turbulent transport (III) is generally modeled

using a classical gradient expression (gradient turbulent

transport at the third order).

² Mean progress variable gradient terms (IV) needs

Reynolds stress modeling for

¯

u

//

i

u

//

j

: Then, new balance

equations are derived, and closed, for these quantities

(second order turbulence model).

² The mean velocity gradient term (V) is closed because

turbulent ¯uxes

¯

u

//

j

c

//

are provided by their balance

equations.

² The mean pressure gradient term (VI) is easily known

under a BML assumption (Section 7.3). Making use of

Eq. (165) yields:

c

//

= c 2 ~ c = t

~ c(1 2 ~ c)

1 1t~ c

(224)

² The ¯uctuating pressure term (VII) is more dif®cult to

understand and to model. In Ref. [151], this term is

neglected, however this assumption is not supported by

DNS results [152]. The mean pressure term (VI) and the

¯uctuatingpressureterm(VII) probablyneedtobe modeled

together as c

//

2p=2x

i

: In Ref. [128], the pressure gradient

termclosure is carefullydiscussed. Recently, a model based

on a partitioning of each pressure ¯uctuation covariance

into contributions from reactants, products and thin ¯ame-

lets was proposed [132]. The comparison of this new

model with DNS results are encouraging and con®rm the

importance of the intermittency between the conditional

mean pressure gradients in reactants and products.

² Turbulent dissipation terms (VIII and IX) are generally

expressed together using small scale dissipation rate

assumptions [151].

Second order closures for turbulent ¯uxes and Reynolds

stresses require nine additional balance equations in 3D simu-

lations (three for progress variable ¯uxes,

¯

u

//

i

c

//

; and six for

Reynolds stresses,

¯

u

//

i

u

//

j

). Because of very high computa-

tional costs, model closures and implementation dif®culties,

very few simulations have been conducted using the second

order formulation. Recently [129], very promising results

were obtained, particularly in predicting experimental results

[20].

8.7. Towards a conditional turbulence modeling?

The BML formulation (Section 7.3.5) directly distin-

guishes between the properties of fresh and burnt gases,

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 244

Fig. 37. Axial velocity u histograms for four transverse locations 80 mm downstream the rod in the V-shape turbulent ¯ame (Fig. 19): burnt

gases (y = 0 mm; -´-); reaction zone (y = 5 mm; ´ ´ ´ and y = 12 mm; Ð Ð

Ð); unburnt gases (y = 18 mm; Ð) [121].

this was achieved using conditional averaging (Eq. (168)).

This approach is very attractive because turbulence charac-

teristics may differ in fresh and in burnt gases as shown in

Fig. 37, where velocity histograms obtained using laser

Doppler velocimetry in the V-shape turbulent ¯ame

(Section 7.4.1) are displayed.

The velocity distribution is almost Gaussian in the

unburnt and the burnt gases, but becomes clearly bimodal

in the reaction zone denoting an intermittency between fresh

and burnt gases, according to Eq. (172). This bimodal distri-

bution does not correspond to the assumptions made in the

derivation of most turbulence models, such as k±1.

A conditional approach to determine the conditional

averages of a quantity Q in fresh (

Q

u

) and burnt (

Q

b

)

gases is probably a promising way leading to a straight-

forward description of turbulent transport (Eqs. (170) and

(172)). The objective is then to derive balance equations for

quantities such as:

rcQ = r ~ c

Q

b

; r(1 2c)Q = r(1 2 ~ c)

Q

u

which is easy from c and Q balance equations [141]. These

equations remain to be closed and a few attempts have been

conducted in this direction [91,148,154], but no clear

conclusions can be found in the literature.

All these works devoted to turbulent transport suggest

that turbulent combustion modeling might probably be

greatly improved by advancing the description of turbulent

transport itself. This point motivates large eddy simulation

(LES) for turbulent combustion modeling [155].

9. Reynolds averaged models for non-premixed

turbulent combustion

9.1. Introduction

Much work has been devoted to the numerical modeling

of non-premixed combustion systems, mainly assuming a

chemistry much faster than mixing and molecular diffusion.

This `mixed is burnt' regime is easily described from the

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 245

Table 7

Modeling strategies for non-premixed turbulent ¯ames. Z is the mixture fraction, x = Du7Zu

2

its scalar dissipation rate,

~

P denotes pdfs, V

Ç

, an

integrated amount of heat release, S, the density of ¯ame surface and

Y

C

i

; a conditional moment. PaSR stands for partially stirred reactor,

turbulent micromixing was expressed with a linear relaxation (IEM-LMSE) D7

2

Y

i

< (

~

Y

i

2Y

i

)=t

t

, where t

t

< (k=1) is an eddy break-up mixing

time; other formulation exist [96] (from Ref. [55])

Major assumptions Flame structure in mixture

fraction space

Turbulent combustion model

In®nitely fast chemistry Species are function of Z:

Y

i

= Y

i

(Z

p

)

Presumed Pdf:

~

P(Z

p

) from

~

Z and

¯

Z

//2

Chemical equilibrium [35]

Y

i

=

_

Z

Y

i

(Z

p

)

~

P(Z

p

)dZ

p

1D strained reaction zones Species are functions of Z and x:

Y

i

= Y

i

(Z

p

; x

p

(t))

Flamelet modeling:

~

Y

i

=

_

Z

_

x

Y

i

(Z

p

; x

p

)

~

P(Z

p

; x

p

)dZ

p

dx

p

[165]

Diffusion balances chemistry

as in a 1D counter-¯ow ¯ame

Steady [63] or unsteady [177]

Single-step or complex

chemistry

Total heat release is a function of x:

_

V(x

st

) =

_

11

21

_ v

T

(j; x

st

)dj

Coherent ¯ame model:

Solve for the density of ¯ame surface S

Constant (Le

i

= 1) or

complex transport properties

(Le

i

± 1)

j coordinate across the ¯amelet [88]

_ v

T

=

_

V( x)S [84]

Diffusion is captured via

micromixing modelling

Solve for conditional moments:

Y

C

i

(Z

p

) =

_

Y

i

uZ = Z

p

_

Conditional moment closure:

presumed pdf

P(Z

p

) [102,103]

~

Y

i

=

_

1

0

Y

C

i

(Z

p

)

~

P(Z

p

)dZ

p

Simple transport properties

(Le

i

= 1)

Solve for representative PaSR:

dY

i

=dZ = (

~

Y

i

2Y

i

1t

t

_ v

PaSR

i

)=(

~

Z 2Z)

to get _ v

i

(Z; t

t

)

PaSR

Presumed pdf 1PaSR modelling:

MIL/PEUL [68,191]

_ v

i

=

k

_

Z

( _ v

PaSR

i

(Z

p

; t

p

t

))

k

P(Z

p

; t

p

t

)dZ

p

dt

t

Single-step or complex

chemistry

Solve for Monte-Carlo particles:

dY

k

i

=dt = (

~

Y

i

2Y

k

i

)=t

t

1 _ v

k

i

to get

~

P(Y

p

1

; ¼; Y

p

n

)

Pdf methods:

~

Y

i

=

_

Y

1

¼

_

Y

n

Y

p

i

~

P(Y

p

1

; ¼; Y

p

n

)dY

p

1

¼dY

p

n

[93]

Chemical source is closed

(only for pdf methods)

turbulent mixing of conserved scalars (Section 3.2) and

mixing problems have been the subject of many discussions

[156±158].

There exist strong motivations for improving non-premixed

and partially premixed turbulent combustion modeling:

² The development of new combustion technologies for

aircraft engines, and more generally for gas turbines

operating in the non-premixed regime, implies the accu-

rate determination of the position in the ¯ow where

combustion starts and the control of pollutants emission.

Crucial points which cannot be addressed using the in®-

nitely fast chemistry hypothesis.

² Many practical systems include liquid injection of the

fuel, followed by non-premixed and partially premixed

combustion.

² Even in burners operating in the premixed regime, the

premixing of the reactants is not always complete at the

molecular level and some partial premixing may be

observed. Sometimes, partial premixing is even desirable

to limit pollutant emissions (strati®ed charge engines).

As for premixed combustion (Sections 7 and 8), the

modeling of turbulent diffusion ¯ames relies on simplifying

assumptions for both chemistry and transport. Depending on

the simpli®cations made for these mechanisms, various

approaches for laminar ¯ames and models for turbulent

¯ames are obtained (see Table 7). Hypotheses formulated

to construct models for non-premixed turbulent ¯ames may

be organized into three major groups:

² Assumption of in®nitely fast chemistry (mixed is burnt).

² Finite rate chemistry assuming a local diffusive±reactive

budget similar to the one observed in laminar ¯ames

(¯amelet assumption).

² Finite rate chemistry with treatment of molecular and

heat transport separated from chemical reaction (CMC,

pdf method; Section 6). Diffusion is then addressed using

turbulent micromixing modeling, while chemical sources

can be dealt with in an exact and closed form (only for

pdf).

The proposed closures are all based on a particular

description of turbulent mixing, hence the basic concepts

useful to capture fuel/air turbulent mixing are ®rst

presented.

9.2. Fuel/air mixing modeling

9.2.1. Introduction

The mean value of the mixture fraction

~

Z gives an indica-

tion of the local mean fuel/air mixing in turbulent ¯ows. In

addition, the structure and the properties of the ¯ame depend

on

¯

Z

//2

; measuring the degree of mixing between reactants.

A simple description of turbulent mixing is thus obtained

from the two ®elds:

~

Z and

¯

Z

//2

: Introducing the classical

gradient transport closure for the turbulent ¯uxes (Section

6), one may write:

r

2

~

Z

2t

1 r ~ u´7

~

Z = rn

t

7

2

~

Z (225)

r

2

¯

Z

//2

2t

1 r ~ u´7

¯

Z

// 2

= rn

t

7

2 ¯

Z

//2

12 rn

t

u7

~

Zu

2

....,,....

Production

2 2 r ~ x

.,,.

Dissipation

(226)

The ®rst term on the RHS is the turbulent transport, the

second in Eq. (226) is the production of ¯uctuations by

the mean gradients, the last is the scalar dissipation rate ~ x

remaining unclosed. r ~ x = rDu7Zu

2

was discussed for

premixed turbulent ¯ames, in the EBU, BML and S models

(Section 7).

9.2.2. Balance equation and simple relaxation model for ~ x

A transport equation may be derived for ~ x from the Z

balance equation (Eq. (41)):

2 r ~ x

2t

1

2 r ~ u

j

~ x

2x

j

1

2

2x

j

rDu

//

j

2Z

//

2x

i

2Z

//

2x

i

_ _

= 22

2

2x

j

rDu

//

j

2Z

//

2x

i

2

~

Z

2x

i

_ _

.........,,.........

(I)

22

2

~

Z

2x

i

rD

2Z

//

2x

j

2u

//

j

2x

i

_ _

........,,........

(II)

22

2

~

Z

2x

j

rD

2Z

//

2x

i

2u

//

j

2x

i

_ _

........,,........

(III)

22

2~ u

j

2x

i

rD

2Z

//

2x

i

2Z

//

2x

j

_ _

.........,,.........

(IV)

22 rD

2u

//

j

2x

i

2Z

//

2x

i

2Z

//

2x

j

.......,,.......

(V)

22 D

2

r

2

2x

j

2Z

//

2x

i

_ _

2

2x

j

2Z

//

2x

i

_ _

............,,............

(VI)

2

2

r

D

2Z

2x

i

2r

2x

i

2

2x

j

rD

2Z

2x

j

_ _

...........,,...........

(VII)

(227)

The term (I) on the RHS describes curvature effects of the

mean mixture fraction ®eld, and as (II) and (III), contains

correlations between the Z ®eld and the velocity ®eld.

Compared to other terms, these correlations vanish for suf®-

ciently large Reynolds numbers. (IV) corresponds to the

correlations between mean ¯ow motion and ¯uctuations.

(V) is the strain rate of the scalar Z ®eld, already discussed

for ¯ame surface density modeling in premixed ¯ames

(Section 7.4.4). (VI) is the dissipation rate of the scalar

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 246

dissipation rate and (VII) is negligible when density gradi-

ents are kept small.

To derive a linear relaxation closure for the scalar dissi-

pation rate, a homogeneous Z steady ®eld in equilibrium

(production = dissipation) is ®rst considered. Then,

Eq. (227) reduces to a balance between the straining rate

of Z and the dissipation rate of the scalar dissipation rate:

22rD

2u

//

j

2x

i

2Z

//

2x

i

2Z

//

2x

j

= 2D

2

r

2

2x

j

2Z

//

2x

i

_ _

2

2x

j

2Z

//

2x

i

_ _

(228)

the ¯uctuating velocity gradient is assumed to be of the

order of the inverse of the small scales characteristic time,

i.e. (2u

//

j

=2x

i

) , (1=t

k

) , (1=n)

1=2

: The remaining part of the

strain rate term is proportional to ~ x; then,

22rD

2u

//

j

2x

i

2Z

//

2x

i

2Z

//

2x

j

,

r ~ x

t

k

, R

l

r ~ x

(k=1)

where R

l

is a Reynolds number based on the Taylor

microscale. A quadratic behavior in ~ x; with a direct depen-

dence on (

¯

Z

//2

)

21

is anticipated for the dissipation rate of the

scalar dissipation rate:

2rD

2

2

2x

j

2Z

//

2x

i

_ _

2

2x

j

2Z

//

2x

i

_ _

, R

l

r ~ x

2

¯

Z

//2

The equilibrium condition (Eq. (228)) leads to:

~ x =

¯

Z

//2

(k=1)

(229)

recovering the widely used linear relaxation model.

Eq. (229) is a very simpli®ed description of micromixing.

The scalar dissipation rate depends on the detail of the char-

acteristics of the turbulence (velocity and scalar energetic

spectrum), the linear relaxation is thus a ®rst approximation

that may be improved, for instance, by solving a modeled

balanced equation for ~ x: Such a closed equation may be

found in Ref. [159].

The knowledge of

~

Z and

¯

Z

//2

from their closed transport

equation may be used to presume the pdf of Z,

~

P(Z

p

; x; t);

with a b-shape (see Section 6.4.2). Because the internal

structure of diffusion ¯ames is easily obtained from condi-

tional statistics as

_

Y

i

uZ = Z

p

_

(Section 3.2), many models

(¯amelet, CMC) incorporate a b-pdf to calculate means

quantities:

~

Y

i

(x; t) =

_1

0

(Y

i

uZ = Z

p

)

....,,....

Flame structure

~

P(Z

p

; x; t)

,........,

Mixing

dZ

p

(230)

The simplest of these models invokes the in®nitely fast

chemistry assumption.

9.3. Models assuming in®nitely fast chemistry

9.3.1. Eddy dissipation model

The eddy dissipation model (EDC) is a direct extension to

non-premixed ¯ame of the eddy break up (EBU) closure,

initially devoted to turbulent premixed combustion [160]

(Section 7.2). The fuel burning rate is calculated according

to:

r _ v

F

= a r

1

k

min

~

Y

F

;

~

Y

O

s

; b

~

Y

P

(1 1s)

_ _

(231)

where a and b are adjustable parameters of the closure. In

Eq. (231), the reaction rate is limited by a de®cient species.

To account for the existence of burnt gases bringing the

energy to ignite the fresh reactants, this species may be

the reaction products. A priori, this model does not respect

the response of diffusion combustion in mixture fraction

space and may generate mean fuel mass fraction values

lower than

~

Y

IFCM

F

: Actually, for large a, Eq. (231) is dif®cult

to justify.

9.3.2. Presumed pdf: in®nitely fast chemistry model

One of the ®rst descriptions of non-premixed combustion

was given in Ref. [35] which assumed an in®nitely fast

single step chemical reaction (Section 3.2). Considering

the piecewise relations (Fig. 5, Table 1):

Y

F

= Y

IFCM

F

(Z); Y

O

= Y

IFCM

O

(Z); T = T

IFCM

(Z) (232)

relating the fuel and oxidizer mass fractions, and the

temperature to the mixture fraction in the case of in®nitely

fast chemistry, mean quantities may be directly obtained

with a b-pdf (Section 6.4.2) and Eq. (230):

~

Y

IFCM

F

=

_1

0

Y

IFCM

F

(Z

p

)

~

P(Z

p

; x; t)dZ

p

(233)

~

Y

IFCM

O

=

_1

0

Y

IFCM

O

(Z

p

)

~

P(Z

p

; x; t)dZ

p

(234)

~

T

IFCM

=

_1

0

T

IFCM

(Z

p

)

~

P(Z

p

; x; t)dZ

p

(235)

The mean ¯ame structure is then known from an in®nitely

fast chemistry model (IFCM) without solving balance equa-

tions for mean thermochemical quantities. The inputs of this

`mixed is burnt' closure are

~

Z and

¯

Z

//2

determining the b-

pdf chosen to capture the detail of fuel/air mixing. IFCM is

therefore a two-equations model for non-premixed turbulent

combustion. It is very popular and is usually coupled with a

low Mach number solution of the turbulent ¯ow [161], so

that the thermodynamics of the ¯ow is fully known from Eq.

(235). IFCM is an interesting ®rst guess to provide the

global ¯ame structure for the maximum of heat that can

be released, a valuable result for certain aspects of design.

Note that the piecewise relations are exact when the chemi-

cal time is in®nitely small. Therefore, for very large

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 247

DamkoÈhler numbers (Section 3.2), IFCM is an accurate

description of a turbulent diffusion ¯ame. Unfortunately

such a zero order model does not exist for premixed turbu-

lent combustion. To handle multi-step chemistry, in®nitely

fast chemistry may be replaced by a chemical equilibrium

condition [162].

This model lacks any prediction capacities when ignition,

quenching or even small ®nite rate chemistry effects exist.

9.4. Flamelet modeling

9.4.1. Introduction

Experiments in jets ¯ames and direct numerical simula-

tions suggest that there exist situations in burners where the

chemistry is fast, but not in®nitely fast (see for example

[163,164]). In these measurements and calculations, the

response of the ¯ame in mixture fraction space lies in the

vicinity of the curves given by Y

IFCM

i

(Fig. 38).

Models have been proposed for such ¯ames [63,165]. For

a given state of mixing in the turbulent ¯ow, thus given

values of Z and x, ¯amelet models are derived assuming

that the local balance between diffusion and reaction is

similar to the one found in a prototype laminar ¯ame for

the same values of Z and x. Flamelet models are therefore

constructed from an asymptotic view of diffusive-reaction

layers as given by Fig. 6. The two control parameters of

planar and steady laminar strained ¯ames are used: the

mixture fraction Z and its scalar dissipation rate x (Section

3.2). In a turbulent ¯ow, these two quantities ¯uctuate in

space and time, but when the joint pdf

~

P(Z

p

; x

p

; x; t) is

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 248

Fig. 38. Scatter plots of major species mole fractions and temperature as functions of mixture fraction: (a) Data from three-dimensional direct

numerical simulations of Montgomery et al. [164]; (b) Raman scattering measurements of Barlow et al. [163] in H

2

/argon±air ¯ame. Solid lines

indicate chemical equilibrium.

known, the mean properties of the ¯ame may be calculated

as:

~

Y

i

=

_

Z

p

_

x

p

Y

SLFM

i

(Z

p

; x

p

)

~

P(Z

p

; x

p

; x; t)dx

p

dZ

p

(236)

Y

SLFM

i

(Z

p

; x

p

) is the local ¯ame structure in mixture fraction

space and

~

P(Z

p

; x

p

; x; t) captures the statistics of fuel/air

mixing. SLFM stands for steady laminar ¯amelet model.

This model may be viewed as a direct improvement of the

in®nitely fast chemistry model (IFCM), since it uses the

same formalism, but with an additional parameter: the scalar

dissipation rate x, thereby including ®nite rate chemistry

effects. For a given chemistry, and therefore a given chemi-

cal time t

c

, when the DamkoÈhler number Da

p

= (t

c

x

st

)

21

is

large, IFCM is recovered. An increase in x is followed by

®nite rate chemistry effects, or even quenching when x

becomes too large (Fig. 6), then, Y

SLFM

i

(Z

p

; x

p

) features

mixing without reaction (Fig. 5).

The inputs of SLFM are similar to those of IFCM:

~

Z;

¯

Z

//2

to which ~ x is added. Two issues emerge:

1. Y

SLFM

i

(Z

p

; x

p

) must be determined and tabulated under

particular hypothesis, choosing a given laminar ¯ame

prototype.

2.

~

P(Z

p

; x

p

; x; t) must be presumed using the mean values

available to quantify fuel/air mixing (i.e.

~

Z;

¯

Z

//2

and ~ x).

9.4.2. Flame structure in composition space, Y

SLFM

i

(Z

p

; x

p

)

Y

SLFM

i

(Z

p

; x

p

) may be tabulated from solutions of coun-

ter-¯ow diffusion ¯ames (Fig. 4 [63]), a ¯ame con®guration

widely studied experimentally [166,167]. Assuming within

the turbulent ¯ow thin quasi-one-dimensional structures

convected and stretched by the ¯uid motions, and neglecting

higher order terms, the equations for the species and

temperature become (Section 3.2):

2Y

i

2t

= _ v

i

1

x

Le

i

_ _

2

2

Y

i

2Z

2

_ _

2T

2t

= 2

N

n=1

h

n

_ v

n

C

p

1x

2

2

T

2Z

2

_ _

where x = (l=rC

p

)u7Zu

2

:

These equations were used to discuss laminar diffusion

¯ames (see Eqs. (47) and (50) and Section 3.2). Omitting the

time derivative (steady ¯amelet), for a given value of x

corresponding to local micromixing conditions, one has to

solve for:

_ v = 2

x

Le

i

2

2

Y

i

2Z

2

(237)

The solution of this equation for given concentrations and

temperatures boundary conditions, and various x provides a

Flamelet Library Y

i

(Z,x). A variety of techniques are avail-

able to build these libraries [168].

In SLFM, the characteristic time required to balance diffu-

sion and reaction is assumed to be much smaller than any

other ¯owtime scale of the problem. The hypotheses involved

in SLFM can be disputed, leading to various improvements:

² Eq. (237) has been obtained neglecting diffusion in the

direction tangential to the iso-Z

st

surface, arguing that

when the mixing element is suf®ciently thin and features

weak curvature, the gradients measured along the stoi-

chiometric surface are much smaller than those in the

perpendicular direction (Section 3.2). Thus, when using

Eq. (237) to describe a prototype ¯ame of turbulent

¯amelets, one supposes that the mixing ®eld may be

reduced to a steady one-dimensional structure. In conse-

quence, the validity of Eq. (237) in a turbulent ¯ow also

depends on the properties of micromixing, and up to now,

it is not obvious how to draw conclusions since one

would need to measure the scalar dissipation rate in

turbulent ¯ames. Some experimental results are available

[169], but more works are required to conclude on the

dimensionality of scalar micromixing in ¯ames.

² There are other issues related to the multi-dimensional

character of diffusion ¯ames. Straining cannot be

uniformly distributed along the ¯ame sheet, leading to

¯amelet interactions when the distribution of x is non-

uniform on the iso-Z

st

. This transverse loss or gain of heat

modi®es the structure of the ¯amelet in the normal direc-

tion to the stoichiometric surface [55].

² Reference states at in®nity used to tabulate the ¯amelets

may have to account for partial premixing [170,171].

Consider the simple case of a jet ¯ame, where close to

the nozzle inlet, pure fuel and pure air react to form

products. Moving downstream, turbulent diffusion

mixes these products with air, on the air side, and with

fuel, on the fuel side. Further downstream, the reactants

feeding the reaction zone are not likely to be either pure

fuel or pure air. This situation is strongly enhanced in

¯ows where recirculation zones are found to stabilize

combustion. In Ref. [171], this shortcoming of SLFM is

overcome introducing transient ¯amelets, for which

reference states at in®nity vary according to the value

of a progress variable ~ c:

² Flamelet libraries can be calculated in physical space or

in mixture fraction space. It is usually observed that

the decay of OH towards equilibrium is predicted by

the ¯amelet solution in mixture fraction space using the

scalar dissipation rate as a control parameter sensitive to

species boundary conditions, a trend that is not fully

reproduced by the ¯amelet solution in physical space

where the input parameter is the strain rate [171].

² Unsteadiness is also an important aspect. Time-depen-

dent ¯amelets have been used to include unsteady effects

[172±174]. As shown in Ref. [175], when a high value of

the scalar dissipation rate is imposed to the ¯ame for a

suf®ciently short period of time, extinction may not be

completed, with a limiting frequency at which the ¯ame

almost behaves like a steady state ¯ame. Unsteady ¯ame-

lets were used [176] to simulate extinction and re-ignition

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 249

in a turbulent jet ¯ame, history effects were included

using a Lagrangian time measured along the stoichio-

metric line. This work was pursued [177] introducing

as an additional control parameter, the diffusion time

needed to exchange mass and energy over a distance

DZ in mixture fraction space.

This is more generally related to questions arising

concerning the determination of quenching limits along

with the accuracy of quenching predictions using ¯amelet

theory a problem which can be addressed numerically,

for instance by post-processing DNS databases to study

the reactive/diffusive layer in terms of ¯amelets

[70,74,178±181]. Numerical simulations of ¯ame/vortex

interaction, used to construct a combustion diagram

([70], Fig. 10), show that when strong unsteadiness

and/or curvature effects appear, the laminar ¯amelet

assumption does not always predict quenching (Fig.

10). Two critical DamkoÈhler numbers, Da

LFA

and Da

ext

,

are easily derived from the simulations to ®nd the limit

conditions where unsteady (Da

LFA

) and extinction (Da

ext

)

effects become important. These numbers are then

compared with Da

p

q

; the quenching DamkoÈhler derived

from asymptotic analysis [30]. It is found that

(Da

LFA

=Da

p

q

) < 2 while (Da

ext

=Da

p

q

) < 0:4: Therefore,

Eq. (237) is an interesting approximation for Da

p

larger

than Da

p

q

; twice the asymptotic quenching value in these

simulations. Below this value, even in the simple con®g-

uration of ¯ame/vortex interaction, the time evolution of

the reaction zone is fully dominated by unsteady effects.

² Another related point of interest in ¯amelet theory is the

response of the turbulent ¯ame when quenching zones

develop. In other words, one may discuss the assumption

that the occurrence of local extinction at some points

does not prevent the use of a ¯amelet model for the

remaining part of the turbulent ¯ame. Again using

DNS, constant density ¯ames near extinction were

studied [182]. A critical DamkoÈhler number at which

extinction occurs is determined as a function of a ¯ame

thickness parameter, de®ned as the ratio between the rms

mixture fraction and the reaction zone width in mixture

fraction space. As expected for these ¯ames featuring

strong unsteadiness effects, the value of the critical

extinction DamkoÈhler number was different from the

one predicted by laminar ¯ame theory. This observation

was explained by statistical variability. Consequently, an

extinction may be observed with DamkoÈhler numbers

larger or smaller than Da

p

q

: In this last study, the lower

value of Da

ext

differs from that of steady ¯amelets and

exceeds the value Da

p

q

given by ¯amelet theory.

Other DNS results have shown that the scalar dissipation

rate controlling the growth of the ¯ame hole is lower than

the one that should be applied to ®rst quench the ¯ame

[38]. In Fig. 39, a DNS database is further analyzed [53]

in which a diffusion ¯ame is pinched by a pair of vortices.

The value of x when the ¯ame extinguishes is in perfect

agreement with x

q

measured in a laminar ¯amelet library

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 250

Fig. 39. Three successive times of two-dimensional unsteady quenching of a planar diffusion ¯ame pinched by a pair of vortices, con®guration

studied in Ref. [53]. Left, bold: iso-reaction rate, dotted: vorcity, zoom: region of the ¯ow where the centerline response of the ¯ame versus

inverse mixture fraction dissipation rate is plotted. Right, solid line: 1D ¯amelet library, diamond: top left 2D ¯ame, square: middle left 2D

¯ame, triangle: bottom left 2D ¯ame.

(Fig. 39, top). But once a hole exists in the ¯ame, the

scalar dissipation rate x

q

Ed

at the quenching location is

smaller than x

q

and varies depending on the structure of

the edge ¯ame. The ¯uxes of heat along the stoichio-

metric surface are responsible for this departure from

one-dimensional ¯amelet behavior. When the diffusion

¯ame and its extremity have reaction zones of the same

thickness, for x .x

q

the response of the reaction rate

follows the laminar ¯amelet one (Fig. 39, middle). Latter

after quenching, x

q

Ed

decreases and a partially premixed

front develops at the edge of the burning zone, then the

reaction rate reaches values above the laminar ¯amelet

behavior (Fig. 39, bottom). These values are representa-

tive of the existence of burning in a premixed regime at

the extremity of the diffusion ¯ame. Therefore, the

quenching scalar dissipation rate x

q

is the relevant quan-

tity to describe the quenching of a burning ¯amelet, but

diffusion ¯ame quenching leads to edge ¯ame combus-

tion (partially premixed ¯amelets).

² When partially premixed ¯amelets are expected in the

combustion system, for instance at the base of a lifted

turbulent jet-¯ame, it may be interesting in the modeling

to use premixed ¯amelets instead of diffusion ¯amelets.

Bradley and coworkers have proposed ¯amelet closures

along these lines where chemical sources are tabulated

using one-dimensional premixed ¯ames [183]. The

chemical sources are parametrized in terms of the

mixture fraction Z, that determines the equivalence

ratio of the ¯amelet whose progress of reaction is given

by a progress variable c. The mean burning rates are then

obtained by averaging the source terms with a presumed

form for the joint-pdf of Z and c.

Within the family of closures for non-premixed turbulent

¯ames, models based on Eq. (237) represent great progress

compared to the in®nitely fast chemistry hypothesis. When

the chemistry is fast enough with mixing elements featuring

sheet-like properties, they provide interesting results

(Fig. 40) and further re®nements of ¯amelet models are

under development [184]. It is also worthwhile to note

that with SLFM, the turbulent micromixing of chemical

species does not need any particular treatment, since diffu-

sion is directly included and coupled with chemistry in Eq.

(237).

9.4.3. Mixing modeling in SLFM

In steady laminar ¯amelet models (SLFM), the fuel/air

turbulent mixing is captured via the joint pdf

~

P(Z

p

; x

p

; x; t):

Most of the ¯amelet models explicitly suppose that the

mixture fraction and its dissipation rate are two uncorrelated

quantities:

~

P(Z

p

; x

p

; x; t) <

~

P(Z

p

; x; t)

~

P(x

p

; x; t)

A b-function is assumed to presume

~

P(Z

p

; x; t) (see Section

6.4.2) and a log-normal distribution is used for

~

P(x

p

; x; t)

[63]:

~

P(x

p

; x; t) =

1

2p

_

x

p

s(x; t)

exp 2

1

2s

2

(x; t)

(ln x

p

2g(x; t))

2

_ _

The two parameters s and g are provided by the ®rst and

second moments of the scalar dissipation rate:

~ x = exp g 1

s

2

2

_ _

and

¯

x

//2

= ~ x

2

(exp(s

2

) 21

In Ref. [63], s

2

= 2 is proposed.

Another alternative is to estimate a mean scalar dissipa-

tion rate under stoichiometric condition [89]. Starting from

the solution of a steady strained planar counter-¯ow ¯ame:

Z(h) =

1

2

erfc(h=

2

_

); where h(j) =

a

D

_

_j

0

r(j

p

)

r

0

dj

p

a is the strain rate, j the coordinate normal to the stoichio-

metric plane and the assumption r

2

D < cst = r

2

0

D

0

was

used. The corresponding scalar dissipation rate is given by:

x(Z) =

a

2p

exp(22[erfc

21

(2Z)]

2

)

= x

0

exp(22[erfc

21

(2Z)]

2

) = x

0

F(Z) (238)

erfc

21

denotes the reciprocal of the complementary error

function and x

0

is the maximum value of the scalar dissipa-

tion rate. Mean and conditional scalar dissipation rate are

then related by:

~ x

st

= ~ xF(

~

Z;

¯

Z

//2

) (239)

where F is given by:

F(

~

Z;

¯

Z

//2

) =

F(Z

st

)

_1

0

F(Z

p

)

~

P(Z

p

)dZ

p

(240)

and to retain in the ¯amelet library the pro®les

Y

SLFM

i

(Z

p

; ~ x

st

) for averaging with

~

P(Z

p

) (Eq. (236)).

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 251

Fig. 40. Top: Favre-averages mean values of temperature and OH

mole fraction along the centerline of a jet ¯ame. Bottom: radial

pro®les of mean temperature and OH mole fraction. Comparison

between experiments and ¯amelet modeling [177].

Here again, the stumbling block is the estimation of the

mean scalar dissipation rate ~ x: The linear relaxation model

may be written (Eq. (229)):

~ x = C

x

¯

Z

//2

(k=1)

where C

x

is a constant depending on the topological and

spectral properties of the mixture fraction ®eld. More

recently, it was proposed to choose [33]:

~ x = C

x

(DZ

F

)

2

(k=1)

C

x

< 2; DZ

F

is the thickness of the reaction zone of these

laminar ¯ame measured in mixture fraction space. The reac-

tion zone is centered at the location Z = Z

st

: As the stoichio-

metric surface is close to the air side for hydrocarbon ¯ame

(Z

st

p1); one may write:

DZ

F

< 2Z

st

which provides an interesting approximation of this reaction

thickness. This last expression was successful to model

turbulent ¯ame lift-off [185].

9.4.4. Conclusion

The laminar ¯amelet model presumes the conditional

mean of species

_

rY

i

uZ = Z

p

_

averaging over the response

of laminar prototype ¯ames:

_

rY

i

uZ = Z

p

_

= r

_

x

p

Y

SLFM

i

(Z

p

; x

p

)

~

P(x

p

)dx

p

In our generic classi®cation of turbulent combustion model-

ing (Section 6) and Eq. (230), SLFM may be viewed as a

presumed pdf technique involving conditional mean values

determined from laminar ¯ame solutions. Another alterna-

tive is to build models where quantities are estimated intro-

ducing a direct treatment of micromixing and small scale

diffusion (see Sections 9.6±9.8).

9.5. Flame surface density modeling, coherent ¯ame model

Flame surface density concepts were ®rstly introduced for

the coherent ¯ame model (CFM) [84]. The balance equation

for the ¯ame surface density S was based on phenomeno-

logical considerations starting from balance equation for a

material surface where combustion effects have been in-

tuitively added [84,186]. Recent works [88,187,188] have

provided an exact balance equation, identifying the ¯ame

surface to the stoichiometric iso-surface Z = Z

st

: The ¯ame

surface density is then de®ned as:

S = u7Zud(Z 2Z

st

) =

_

u7ZuuZ = Z

st

_

P(Z

st

) (241)

where

P(Z

st

) denotes the probability of ®nding Z = Z

st

:

Starting from this de®nition and the balance equation for

the mixture fraction Z (Eq. (41)), an exact but unclosed,

balance equation for the ¯ame surface density S may be

derived (Section 6.3.2).

In ¯ame surface density models, under a ¯amelet assump-

tion, it is assumed that the interface between fuel and

oxidizer in non-premixed burners behaves, for any values

of Z, as S(Z

p

) < u7Zu = S (Section 6.3.2). The mean burn-

ing rate

_ v

i

is then expressed as the product of S by

_

V

i

the

local reaction rate per unit of ¯ame area. S accounts for

¯ame turbulence interaction and

_

V

i

for the chemistry.

According to Eq. (121)

_ v

i

=

_ v

i

u7Zu

_ _

s

u7Zu =

_

V

i

S (242)

The local burning rate

_

V

i

is estimated from a one-dimen-

sional laminar ¯ame as:

_

V

i

=

_ v

i

u7Zu

_ _

s

<

_1

0

_ v

i

u7Zu

dZ =

_11

21

_ v

i

dj (243)

where j is the coordinate along the normal to the ¯ame

front.

In diffusion ¯amelets, where _ v

i

depends on the mixture

fraction Z and the scalar dissipation x, expressions (96), (91)

and the relation:

_

_ v

i

uZ

p

_

=

_

x

_ v

i

(Z

p

; x

p

)

P

c

(x

p

uZ

p

)dx

p

(244)

lead to:

_

V

i

=

1

u7Zu

_1

0

_

x

_ v

i

(Z

p

; x

p

)

P

c

(x

p

uZ

p

)

P(Z

p

)dx

p

dZ

p

(245)

Then, using Eq. (243):

_

V

i

<

_11

21

_

x

_ v

i

(x

p

; j)

P

c

(x

p

uZ

p

)dx

p

dj

<

_

x

_11

21

_ v

i

(x

p

; j)dj

_ _

........,,........

_

V

i

(x

p

)

P(x

p

)dx

p

(246)

The reaction rate per unit of ¯ame area should, therefore, be

estimated from the integrated reaction rate of a ¯amelet

submitted to a scalar dissipation rate x with the distribution

P(x

p

): In practice,

_

V

i

is estimated as

_

V

i

=

_

V

i

( ~ x); where ~ x

is the averaged scalar dissipation rate. The coherent ¯ame

model, then, corresponds to a simpli®ed version of the

steady laminar ¯amelet model. The integrated burning rate

_

V

i

may also be expressed as a function of the species mol-

ecular ¯uxes, see Eq. (53) in Section 3.2.

The introduction of effects of unsteadiness was realized in

Ref. [186], following the analysis proposed in Ref. [172] in

the context of SLFM, a ¯ame time response is then intro-

duced when calculating

_

V

i:

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 252

9.6. MIL model

MIL (`ModeÁle Intermittent Lagrangien' or Lagrangian

Intermittent Model) was proposed by Borghi in Ref. [68].

As in the steady laminar ¯amelet model (SLFM) and the

conditional moment closure (CMC), MIL incorporates the

¯ame structure in Z space. Arguing that mixing occurs with-

out reaction for large values of the scalar dissipation rate,

corresponding to fast mixing, the ¯ame structure is

constructed dynamically from two possibilities of diffusion

combustion: Mixing without reaction (before ignition) and

in®nitely fast chemistry (after ignition) (Fig. 41). The tran-

sition between mixing without combustion and a `mixed is

burnt' regime is controlled by the position of an ignition

time delay, t

ig

, within the distribution of fuel/air turbulent

mixing times t

k

, t

t

, 11, where t

k

. (n=1)

1=2

is the

Kolmogorov time.

The ignition delay t

ig

is associated to a scalar dissipation

rate x

ig

(t

ig

< x

21

ig

): The main objective of MIL is to account

for unsteadiness in the coupling between small scale diffu-

sion and chemistry, by means of the spectral distribution of

micromixing times. For a given value of t

ig

< x

21

ig

; mixing

times smaller than t

ig

(t

t

,t

ig

or x .x

ig

) prevent ignition

and contribute to pure mixing. Whereas for x ,x

ig

, corre-

sponding to t

t

.t

ig

, mixing is slower than chemistry and

combustion develops.

The inputs of MIL are:

1. A table of ignition delay times t

ig

depending on mixture

fraction, fuel mass fraction and mean temperature, t

ig

=

t

ig

(

~

T;

~

Y

F

; Z): This look-up table is easily derived from

partially stirred reactor (PaSR) calculations, where

~

T

and

~

Y

F

are the mean conditions in the PaSR. Detailed

chemistry can be used to construct such a t

ig

table

[189].

2. The pdf of turbulent mixing times

P(t

t

; x; t) (equiva-

lent to the pdf of the inverse of x).

3. The pdf of the mixture fraction

~

P(Z

p

; x; t):

4. The mean concentration of fuel

~

Y

F

:

The output is the mean burning rate

~

_ v

F

to calculate

new

~

Y

F

values, accounting for auto-ignition and transi-

ent burning in turbulent ¯ames. Assuming that combus-

tion occurs only when ignition is faster than turbulent

fuel/air mixing, the probability to ®nd ignition at a

particular location is:

a(Z; x; t) =

_11

t

ig

(Z;x;t)

P(t

t

; x; t)dt

t

where

P(t

t

; x; t) is the pdf of the micromixing times

having a mean value equal to the integral length scale

time (k/1). Since t

ig

depends on position in mixture

fraction space and mean temperature, a is also a func-

tion of Z and of the local ¯ow conditions through posi-

tion and time. Once

P(t

t

; x; t) and

~

P(Z

p

; x; t) are known,

the mean burning rate of fuel is estimated as:

~

_ v

F

(x; t) =

_1

0

_11

t

ig

(Z

p

;x;t)

_ v

MIL

F

(Z

p

; t

p

t

)

~

P(Z

p

; x; t)

P(t

p

t

; x; t)dt

p

t

dZ

p

(247)

A partially stirred reactor (PaSR), combining mixing

and chemical effects, is introduced to calculate

_ v

MIL

F

(Z

p

; t

p

t

) [68], leading to:

dY

p

F

dZ

=

(

~

Y

p

F

2Y

F

) 1t

p

t

_ v

MIL

F

(

~

Z 2Z

p

)

(248)

This technique is close to the one chosen in Ref. [190]

for modeling NO

x

production in turbulent jet ¯ames. Eq.

(248) describes a local ¯ame element, where molecular

diffusion, D7

2

Y

F

, is modeled as a linear relaxation term

(

~

Y

F

2Y

F

)=t

t

: Solving Eq. (248) with boundary condi-

tions similar to the ones of ¯amelet libraries provides

_ v

MIL

F

(Z

p

; t

p

t

):

A generic shape for

P(t

t

; x; t) was proposed [189,191]:

P(t

t

; x; t) = (q 21)

t

(q21)

t

q

k

exp 2

(q 21)

q

t

t

t

k

_ _

q

_ _

q(x; t) is determined so that

~

Y

MIL

F

calculated along the MIL

trajectory in mixture fraction space (Fig. 41), weighted by

the b-pdf, reproduces the mean

~

Y

F

:

Fig. 42 shows an example of numerical simulation of a

lean premixed prevaporized (LPP) combustion chamber

using MIL. The LPP system is divided into three parts:

®rst, liquid fuel is injected and vaporized (prevaporized

tube), then is mixed with air in the premixing tube before

entering the main combustion chamber. Combustion is

expected to reduce pollutant emission in a lean premixed

regime. The premixing tube operates for a large range of

inlet temperatures (from 300 to 800 K) and a large range of

pressures (from 1 to 40 bar). For the highest temperature and

pressure conditions, some undesirable effects, like auto-

ignition or ¯ashback, may be observed in the premixed

tube. These ®ndings are reproduced with turbulent combus-

tion closures like MIL describing strongly coupled micro-

mixing and ignition phenomena [191].

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 253

Fig. 41. Sketch of MIL in mixture fraction space.

9.7. Conditional moment closure

In®nitely fast chemistry (IFCM), steady laminar ¯amelets

(SLFM), and MIL suggest that non-premixed turbulent

¯ames may be conveniently studied using conditional aver-

aging in the mixture fraction space. Instead of getting Q

i

=

_

rY

i

uZ = Z

p

_

=

_

ruZ = Z

p

_

from laminar ¯amelets, the CMC

approach proposes to solve a balance equation for Q

i

[102,103,105], which is written, neglecting two terms

(Eq. (114)):

_

ruZ = Z

p

_

2Q

i

2t

= 2

_

ru

i

uZ = Z

p

_

2Q

i

2x

i

.......,,.......

(I)

1

_

rxuZ = Z

p

_

2

2

Q

i

2Z

p2

........,,........

(II)

1

_

_ v

i

uZ = Z

p

_

.....,,.....

(III)

(249)

In addition to turbulent transport included in (I), micro-

mixing (II) and the chemical source (III) are unclosed. In

the simplest version of CMC, the ¯uctuations of chemical

source in mixture fraction space are neglected leading to:

_

_ v

i

uZ = Z

p

_

= _ v

i

__

Y

1

uZ = Z

p

_

; ¼;

_

Y

N

uZ = Z

p

__

(250)

As for any combustion models, the scalar dissipation rate is

unknown (II). With CMC, the conditional average

_

rxuZ = Z

p

_

needs closure. This conditional mean is an

important ingredient of the pdf transport equation and

micromixing models are available (Section 9.8.1). Simi-

larly, when the pdf is known via its presumed b-shape

(Section 6.4.2), the conditional value of x may be approxi-

mated from

~

P(Z

p

) to obtain

_

rxuZ = Z

p

_

[192].

CMC has been the subject of many works and re®nements

during the past years [105]. The following points have been

addressed:

² The approximation of the conditional chemical source

using Eq. (250) is much less restrictive than to write

~

_ v

i

<

_ v

i

(

~

Y

1

; ¼;

~

Y

N

) but implies a weak level of ¯uctuations for

a given value of Z. When calculating the chemical source,

the conditional pdf

P

c

(Y

i

uZ

p

) is approximated by a peak

located at

_

Y

i

uZ = Z

p

_

: Accordingly, this method is

expected to provide interesting results for conditions

not too far from in®nitely fast chemistry. The ¯uctuations

of species concentrations and temperature for a given

level of mixture fraction may be included in the condi-

tional burning rate by using second order CMC. This

leads to a variety of modeling re®nements where more

transport equations for conditional quantities are studied,

modeled and solved [105].

² The estimation of

_

rxuZ = Z

p

_

from a b-pdf is feasible

for simple shapes of

~

P(Z

p

); but as shown from the pdf

balance equation (Eq. (110)), the detail of

_

rxuZ = Z

p

_

is

linked to the detail of the shape of the pdf, that may not be

correctly captured with a presumed pdf.

² Within a homogeneous ®eld, the CMC equation reduces

to:

_

ruZ = Z

p

_

2Q

i

2t

=

_

rxuZ = Z

p

_

2

2

Q

i

2Z

p2

1

_

_ v

i

uZ = Z

p

_

(251)

This last formulation may be compared with the ¯amelet

equation (Eq. (237)) replacing Q

i

by Y

i

[107]. These two

turbulent combustion models have similarities in term of

their general formalism, but their underlying physical

assumptions strongly differ. CMC incorporates diffusion

as turbulent micromixing and decouples it from chemis-

try. On the other hand, the ¯amelet assumption deals with

a coupled representation of diffusion and reaction as in a

laminar ¯ame for a given value of x, and therefore, a

given thickness of the laminar diffusive zone.

Once the number M of conditional values required to

capture the conditional means of the N chemical species is

determined, CMC modeling requires the solution of N £ M

balance equations. CMC was also extended to reacting

particles [193].

9.8. Pdf modeling

In premixed turbulent combustion modeling, BML and

related closures assume a bimodal probability density func-

tion (pdf) of the progress variable. In non-premixed ¯ames,

using in®nitely fast chemistry (IFCM), steady laminar

¯amelet (SLFM), or conditional moment closure (CMC),

the pdf of the mixture fraction was assigned a b-shape.

The objective of PDF modeling is to relax all hypotheses

concerning the shape of pdfs. Once a methodology has been

developed to calculate pdfs, it is possible to construct turbu-

lent combustion closures in which all the values taken by

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 254

Fig. 42. Snapshot of temperature in a TURBOMECA-ONERA LPP engine computed with MIL [191]. 1: Vaporized tube, 2: premixed tube,

3: ¯ame tube. Combustion starts in the premixed tube.

species and temperature in mixture fraction space may be

reached (Figs. 5 and 43). As in ¯amelet modeling and CMC,

a detailed description of chemistry may be incorporated in

the modeling. Differential diffusion is easily introduced in

SLFM when tabulating counter-¯ow ¯ames, these prefer-

ential diffusion effects may be more dif®cult to introduce

in CMC, in the context of pdfs, some works have been made

in this direction [194,195].

Two levels of complexity may be chosen:

² One may solve the balance equation for the joint pdf of

the thermochemical variables, species, temperature, etc.

(Eq. (111) in Section 6):

r

2

2t

~

P(Z

p

; x; t)

= 2 r

N

i=1

2

2Y

p

i

2u´7Y

i

uY(x; t) = Y

p

_ _

1 _ v

p

i

_ _

~

P(Z

p

; x; t)

_ _

2 r

N

i=1

N

j=1

2

2Y

p

i

2

2Y

p

j

D

2Y

i

2x

k

2Y

j

2x

k

_

_

_

_

Y(x; t) = Y

p

_ _

~

P(Z

p

; x; t)

_ _

(252)

² Velocity ®eld statistics may also be included into the pdf,

then the balance equation for the joint pdf of velocity and

species is resolved (Eq. (113)). Recent developments

have also included time scale information into the pdf

[196,197].

There are two major issues with pdf modeling:

² Models must be proposed for capturing micromixing and

pressure ¯uctuation and viscous effects (the latter only

when the velocity ®eld is incorporated into the pdf).

² A new numerical treatment is required and leads to the

decomposition of the pdf into a large number of com-

putational particles required to succeed in the computa-

tional implementation of pdf methods [93,198,199].

The chemical composition and velocities of these par-

ticles evolve according to simple differential equations,

resulting from a discrete Monte Carlo simulation of the

continuous pdf equation.

The main advantage of the pdf method lies in the

possibility of treating complex chemical sources directly

(Section 6).

9.8.1. Turbulent micromixing

From the discussion of fundamentals properties of ¯ames

and of turbulent combustion models, the scalar dissipation

rate appears as a key quantity in both premixed and non-

premixed turbulent systems. Using pdfs and CMC, one

needs to express the conditional mean value of the scalar

dissipation rate (or the conditional mean value of diffusion).

This term represents turbulent micromixing [95,96].

Using turbulence modeling, such as the k±1 closure, a

mean mechanical time t

t

< (k=1) and a microscale time

t

k

< (n=1)

1=2

may be estimated. A global picture of mixing

should therefore distinguish between macro and micro-

mixing [200]:

The macromixing is organized from a cascade mechanism

between large and small scales. The study of this cascade in

a fully developed turbulence shows that the mixing

frequency f

Z

is a function of the characteristic length l

Z

[201]:

f

Z

(l

Z

) ,

1

n

_ _

1=2

h

k

l

z

_ _

2=3

The speed at which Z is mixed increases when the charac-

teristic length l

Z

decreases, suggesting that the cascade

mechanism is limited by the mixing frequency f

Z

(l

Z

) of the

largest scales. When the Schmidt number of Z is greater than

unity (n .D

i

), the impact of the molecular dissipation is

less than the effect of viscous dissipation, then the smallest

scales of Z are smaller than the Kolmogorov scales. For

those small scales (l

Z

,h

k

) the mixing frequency is of the

order of a constant f

Z

(l

Z

) = (1=n)

1=2

: This cascade process

from large to small scale does not change the energy in the

system and therefore does not affect

¯

Z

//2

:

The molecular diffusion acting at smaller scales makes

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 255

Fig. 43. Two-dimensional pdf of mixture fraction and reaction

progress variable for a round jet non-premixed ¯ame [104].

the gradients smoother and then induces the micromixing,

phenomenon required to put the reactants in contact with the

reaction zone. When n = D; this process starts at the

Kolmogorov scale, when n ± D; the Batchelor scale

becomes the relevant limiting length:

l

B

, (nD

2

=1)

1=4

= Sc

21=2

h

k

All these approximations are given for non-reactive ¯ows to

exactly quantify the effect of combustion is an arduous task.

This dif®culty is also encountered in ¯amelet or ¯ame

surface modeling, but appears in a different form.

In pdf modeling, the micromixing closure must estimate

the term

_

D7

2

Y

i

uY = Y

p

_

describing the molecular diffu-

sion of each chemical species, or mimic its impact on the

pdf. Various techniques have been proposed [95].

9.8.2. Linear relaxation model, IEM/LMSE

In b-pdf modeling (Section 9.2.2), the scalar dissipation

rate of the mixture fraction was expressed as ~ x = (

¯

Z

//2

=t

t

):

The extension of this linear relaxation to conditional diffu-

sion provides the linear mean square estimation (LMSE) or

interaction by exchange with the mean (IEM) model. First

proposed in the pdf context by Dopazo and O'Brien [202],

IEM was also developed and used in chemical engineering

applications by Villermaux [203].

Fig. 44 shows the conditional mean of the diffusive

budget of the mixture fraction Z measured in a DNS of a

mixing layer. A quasi-linear response is observed close to

the mean, with a slope that can be approximated using the

inverse of t

t

< (k=1): Since the gradient of Z must relax

to zero for Z = 0 and Z = 1; this quasi-linear response

is combined with the conditions

_

D7

2

ZuZ = 0

_

=

_

D7

2

ZuZ = 1

_

= 0: For a `Gaussian-type pdf' (no peaks

at Z = 0 and Z = 1), the behavior at the edges of the sample

space is not of prime importance to capture the major prop-

erties of micromixing, and a ®rst approximation is given by:

_

D7

2

ZuZ = Z

p

_

<

Z 2Z

p

t

t

_ _

<

Z 2Z

p

k=1

_ _

(253)

This simple model cannot capture the time evolution of the

pdf in homogeneous mixing problems and has many short-

comings, for instance it relaxes the ¯uctuations without

modifying the shape of the pdf [94±96,204].

9.8.3. GIEM model

An improvement of the IEM model was recently

discussed [205], a b-pdf is used to reproduce the correct

behavior of conditional diffusion. The mixture fraction is

chosen as a shadow ®eld to estimate micromixing, this

closure is also called a `Beta-Mapping-Closure'. In a homo-

geneous ®eld, the time evolution of the mixture fraction pdf

is:

2

~

P(Z

p

; x; t)

2t

= 2

2

2Z

p

__

D7

2

ZuZ = Z

p

_

~

P(Z

p

; x; t)

_

since

_

D7

2

ZuZ = 0

_

= 0; one may write:

_

D7

2

ZuZ = Z

p

_

= 2

1

~

P(Z

p

; x; t)

2

2t

_Z

p

0

~

P(Z

1

; x; t)dZ

1

_ _

(254)

When

~

P(Z

p

; x; t) is given by a b-function parameterized

with

~

Z and

¯

Z

//2

; this equation determines (D7

2

ZuZ = Z

p

).

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 256

Fig. 44. Left: Scatter plot of D7

2

Y. Right: (D7

2

YuY(x; t) = Y

p

): These two terms were measured in DNS of an inert mixing layer [234].

The LMSE/IEM model leads to:

_

D7

2

ZuZ = Z

p

_

<

1

t

t

(

~

Z 2Z

p

) (255)

Making a direct analogy GIEM proposes:

_

D7

2

ZuZ = Z

p

_

< C(Z

p

; t)(B

Z

(t) 2Z

p

) (256)

where C(Z

p

, t) is a pseudo-mixing frequency depending both

on the position in mixture fraction space and time. It allows

for reproducing the correct behavior at the edges Z

p

= 0 and

Z

p

= 1 and its time dependence gives the correct behavior

during the relaxation of the pdf. B

Z

(t) denotes the point

where the diffusive ¯ux is zero, relaxing the constraint

B

Z

(t) =

~

Z of IEM, C(Z

p

, t) and B

Z

(t) are determined from

Eqs. (254) and (256).

The same treatment is applied to any reactive species Y

i

,

using the same pseudo-mixing frequency (an assumption

made in most micromixing models):

_

D7

2

Y

i

uY

i

= Y

p

i

; Z = Z

p

_

= C(Z

p

; t)(B

Y

i

(t) 2Y

p

i

)

The parameter B

Y

i

(t) is then estimated from the fact that

micromixing by itself should not change the mean:

2Y

i

2t

_ _

= (2D7

2

Y

i

) < 0 - B

Y

i

(t) =

C(Z; t)Y

i

C(Z; t)

This attractive model conserves the simplicity of linear

relaxation modeling, but overcomes some of its drawbacks.

As

¯

Z

//2

is an input of the closure, the mean scalar dissipation

rate ~ x should be previously modeled. This approach has

similarities with the estimation of the conditional values

of x in CMC using b-pdf [192] and is not a full closure

of micromixing per se.

9.8.4. Stochastic micromixing closures

Micromixing modeling may also be developed by choos-

ing an ad hoc stochastic process mimicking the relaxation of

pdfs due to small scale diffusion. These closures are usually

presented in the context of non-continuous interactions of

volumes of ¯uid. When only diffusion is acting, the generic

form of these models is written [94]:

2

~

P(Z

p

; t)

2t

=

2b

t

Z

_

__

~

P(Z

1

; t)

~

P(Z

11

; t)

×T(Z

1

; Z

11

- Z

p

)dZ

1

dZ

11

2

~

P(Z

p

)

_

b is a parameter of the model and T(Z

1

, Z

11

- Z

p

) is the

probability that a volume of concentration Z

1

interacts with

a volume at Z

11

to evolve into Z

p

. The dif®cult point lies in

the choice of the transition probability T. Multiple choices

are possible and have been made (see Ref. [95] for a

review), the simplest one consists of writing:

T(Z

1

; Z

11

- Z

p

) = d Z

p

2

Z

1

1Z

11

2

_ _

(257)

meaning that the two volumes interact to generate a volume

with Z

p

= (Z

1

1 Z

11

)=2: Eq. (257) was ®rst proposed by

Curl [206] to simulate the production of droplets of a given

diameter by coalescence of two droplets of various sizes and

re-dispersion of the two droplets.

These closures are clearly linked to a purely stochastic

view of micromixing and are well suited to Monte Carlo

simulations [93]. Their applications to ¯ames can be disputed

easilysince mixing described via stochastic processes does not

always account for the presence of reaction zones. However,

much work has been done to improve these models which

have then been quite successful in the calculation of jet

¯ames [97]. Fig. 45 shows pdfs of the temperature calcu-

lated from Eq. (110) [207] using the coalescence-dispersion

model [206] and the k±1 closure for the velocity ®eld [208].

For the same calculations, Fig. 46 illustrates the impact of

micromixing modeling on temperature and CO concentra-

tions predictions, the reduced chemical scheme of Ref.

[209] has been utilized with the LMSE (IEM) and Curl's

coalescence re-dispersion formulation.

Micromixing modeling is obviously one of the greatest

challenges of turbulent combustion modeling, and strong

efforts have been made in this direction. Many types of mixing

models were proposed, a full reviewis beyond the scope of this

paper. Detailed discussion of mixing models in the context

of pdfs may be found in the following list of references

[93,95±97,104,194,196,205,206,208,210±236].

9.8.5. Interlinks PDF/¯ame surface modeling

The unclosed molecular diffusion term in the pdf balance

Eq. (109) may be recast as:

1

r

(7´(rD7Z))uZ = Z

p

_ _

= 2

1

ru7Zu

2

2n

(rDu7Zu)u7ZuuZ = Z

p

_ _

2

_

D7´nu7ZuuZ = Z

p

_

= 2

1

ru7Zu

2

2n

(rDu7Zu)

_ _

s

1kD7´nl

s

_ _

_

u7ZuuZ = Z

p

_

= 2[kv

n

l

s

1kD7´nl

s

]

_

u7ZuuZ = Z

p

_

(258)

The molecular diffusion is decomposed into a normal contri-

bution, that may be expressed in terms of a normal displace-

ment speed w

n

and a curvature term D7´n; according to

Eq. (38). The diffusion coef®cient D is assumed constant

and kD7´nl

s

= Dk7´nl

s

: The ¯ame surface averaged curva-

ture may be modeled from geometrical considerations

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 257

[88,121] as in Section 7.4.4 (Eq. (208)):

k7´nl

s

< 2

Z 2Z

1

L

y

(259)

where L

y

is a wrinkling length scale of the ¯ame surface

and Z

1

is the

Z level where mean curvatures are equal

to zero. Mean curvatures are supposed positive (i.e.

convex toward

Z = 0 region) and negative for

Z = 1.

The comparison of Eqs. (258) and (259) shows that

the geometrical assumption used to model the mean

curvature corresponds to an LMSE±IEM model of the

molecular diffusion, when neglecting the normal displace-

ment speed kw

n

l

s

(Section 9.8.2):

2

1

r

(7´(rD7Z))uZ = Z

p

_ _

=

Z 2Z

p

t

t

(260)

It was shown in (Section 9.8.2) that for a non-reacting

scalar, the IEM model provides an interesting estimation

of the micromixing term around

Z = Z

p

when the turbulent

mixing time is correctly calculated [234] but becomes de®-

cient when Z

p

tends towards 0 or 1, where u7Zu is expected to

relax to zero. This shortcoming is attenuated in the ¯ame

surface density context where the surface averaged curva-

ture k7´nl

s

is multiplied by the ¯ame surface density (Eq.

(90)), ensuring zero values at either side of the ¯ame brush.

The turbulent mixing time t

t

and the ¯ame wrinkling length

scale, L

y

, are then related as L

y

< D

_

u7ZuuZ = Z

p

_

t

t

; where

_

u7ZuuZ = Z

p

_

characterizes laminar ¯ame elements.

Expressing the mixing time from linear relaxation hypo-

thesis, t

t

<

¯

Z

//2

= ~ x and using Eq. (116) under a ¯amelet

assumption yields L

y

<

¯

Z

//2

=u7Zu: Wrinkling length scales

depend on scalar ¯uctuations given by Eq. (226), eviden-

cing, once again, the need for modeling scalar dissipation

rates.

9.8.6. Joint velocity/concentrations pdf modeling

When the joint pdf of velocity and species enters the

problem, one needs to model additional viscous and

pressure terms related to velocity ¯uctuations (see Eq.

(113)). In turbulent non-reactive ¯ows, experiments have

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 258

Fig. 45. Comparison between measurements (left) and numerical predictions (right) of unweighted temperature pdf at different locations in the

Masri et al. [207] `L' jet ¯ame (location x=D = 20). Pdf calculations by Jones and Kakhi [208] using the Curl mixing model [206].

2

In non-premixed turbulent ¯ames, the mean normal displace-

ment speed kw

n

l

s

is expected to be negligible against curvature

effects [88] because such ¯ames are unable to propagate. This is

not the case in premixed combustion, where the IEM±LMSE model

requires a correction to account for kw

n

l

s

that is of the order of the

laminar ¯ame speed [237].

shown that the velocity ¯uctuations have pdfs that may be

approximated with Gaussian functions. Closures for the

velocity pdf may then be obtained using a Langevin type

model, a simple stochastic model reproducing this velocity

distribution, as done in Ref. [238].

Basic properties of stochastic methods [239] show that an

equation with a diffusive term for a pdf

P:

2

P(Z

p

)

2t

= 2

2

2Z

p

(A(Z

p

; t)

P(Z

p

)) 1

1

2

2

2

2

2

Z

p

(B(Z

p

; t)

P(Z

p

))

is well reproduced by a stochastic process, Z(t), similar to

the Langevin process:

dZ(t) = A(Z(t); t)dt 1

B(Z(t); t)

_

dW(t)

where dW(t) is a white noise (Wiener process). Using this

equivalence, the Langevin model [238] for the pressure and

viscous unclosed terms of Eq. (113) may be written:

__

1

r

2p

/

2x

i

uu = u

p

; Z = Z

p

_

2

_

n7

2

u

/

i

uu = u

p

; Z = Z

p

_

_

P(u

p

; Z

p

; x; t)

= G

ij

(u

p

j

2 u

j

)

P(u

p

; Z

p

; x; t) 1C

0

1

2

2u

p

i

[

P(u

p

; Z

p

; x; t)]

(261)

The physical properties of this closureare includedinthe tensor

G

ij

. The simplest model considers a linear relaxation for the

velocity ®eld, then G

ij

= (1=2 13C

0

=4)(1=k)d

ij

; where C

0

is a

constant. More sophisticated methods have been proposed

[240], where a parallel is drawn between Reynolds stress

closures (modeled equations for u

/

i

u

/

j

) and the tensor G

ij

[241].

As already indicated, a numerical solution for the pdf

implies the development of accurate Monte Carlo solutions

[93]. The pdf ®eld is decomposed into a set of stochastic

particles whose evolution simulates the turbulent ¯ame.

When a Lagrangian context is adopted [242], each particle

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 259

Fig. 46. Radial pro®les of CO mass fraction and temperature at different locations in Masri et al. [207] `B' jet ¯ame, pdf calculations [208]. Two

mixing models have been used: ®lled circles: measurements; solid line: coalescence±dispersion micromixing Curl model [206]; dashed line:

LMSE micromixing model (IEM). Reduced chemical scheme of Jones and Lindstedt [209].

n is characterized by its position x

p

n

(t); its velocity u

p

n

(t) and

concentration(s) Y

p

n

(t): The Langevin model given in Eq.

(261) corresponds to the system for the discrete particles:

dx

in

dt

= u

in

du

in

dt

= 2

1

r

2 p

2x

i

_ _

1n

2

2

u

i

2x

j

2x

j

_ _

1G

ij

(u

j

n

2 u

j

) 1(C

0

1)

1=2

dW

i

dt

_ _

dY

p

i

dt

=

(

~

Y

i

2Y

p

i

)

(k=1)

1 _ v

i

(Y

p

i

)

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

(262)

The last dif®culty is the determination of the mean pressure

®eld (2 p=2x

i

): Methods have been proposed to calculate it

from Monte Carlo simulations [101]. In most of Monte

Carlo simulations, the Lagrangian simulation is coupled

with Eulerian calculations of means values providing

the mean pressure ®eld and the dissipation rate of the

velocity ¯uctuations. However, full pdf methods includ-

ing a model for the turbulent frequency following a

¯uid particle have also been developed with success [197].

Fully detailed reviews on pdf methods may be found in

Refs. [93±95].

10. Conclusion

Modeling turbulent combustion is a challenging task. In

turbulent ¯ames, various dif®culties (strong heat release,

complex chemistry, large range of time and length scales,

etc.) are added to the complexity of constant-density turbu-

lent ¯ows.

A review of the most classical Reynolds (or Favre) aver-

aged Navier±Stokes (RANS) models has been proposed.

Three main ingredients must be modeled:

² Reynolds stresses: ¯ u

i

u

j

2 ~ u

i

~ u

j

:

² Turbulent transport of species mass fractions:

¯

u

i

Y

k

2 ~ u

i

~

Y

k

:

² Mean reaction rate of species:

_ v

k

:

Most works have focus on re®ned descriptions of the

mean reaction rate

_ v

k

: To provide this reaction rate, tools

have been proposed, which are based on a pure statistical or

more oriented geometrical view of turbulent ¯ames. All

these various models are related via a scalar dissipation

rate, a key quantity representing micromixing of the reac-

tants occurring at the small scales where chemical reactions

develop. The links and the similarities between combustion

models have been demonstrated here. The precise knowl-

edge of these relations may be useful to understand the exact

implications of the underlying physical hypothesis. Major

models are then directly derived from the proposed expres-

sions, which illuminate unexpected links between well-

known closure schemes.

Few studies have focussed on species turbulent transport,

generally modeled using a gradient expression, whereas

counter-gradient turbulent transport is known to appear in

some situations. Reynolds stresses are generally described

from well established turbulent models (mostly k, 1), simply

re-written in terms of Favre averaged quantities. So far,

combustion effects on the ¯ow (¯ame induced turbulence

generation, higher viscous dissipation, etc.) are not inten-

tionally included in the calculations.

There are many open questions and challenges left. One

of them is the understanding of the coupling between spray

and combustion including detailed chemistry. DNS with a

two-way coupling between a dilute spray and a carrier phase

were recently performed to progress in this direction [243]

and closures exist [243±245], but more work is needed to

reach a level where liquid atomization, vaporization

together with ¯ames are properly described.

RANS combustion models will remain useful for the next

few years. However, large eddy simulations (LES) stands as

very promising technique for turbulent combustion:

² Combustion ¯ow ®elds generally exhibit large scale

motions [26].

² LES appear as a promising tool to capture combustion

instabilities.

² Flames are mainly driven by mixing. LES is a good

candidate to capture unsteady turbulent mixing.

² LES may directly provide part of the description of turbu-

lence/combustion interactions because zones of fresh and

burnt gases, having different turbulence characteristics, are

instantaneously identi®ed at the level of the resolved grid.

LES is at a very early stage for combustion applications and

only few works have been done in this direction, mainly

devoted to feasibility tests (two-dimensional simulations,

constant density ¯ows, etc.). Nonetheless, the results of these

preliminary tests suggest that LES will rapidly become a

complementary way to carefully simulate and understand

turbulent combustion systems. So far LES modeling is essen-

tially based on the same modeling strategies as the ones

described here for RANS turbulent combustion modeling.

References

[1] LinÄaÂn A, Williams FA. Fundamental aspects of combustion.

Oxford: Oxford University Press, 1993.

[2] Clavin P. Premixed combustion and gasdynamics. Annu Rev

Fluid Mech 1994;26:321±52.

[3] Buckmaster J, Ludford G. Theory of laminar ¯ames.

Cambridge: Cambridge University Press, 1993.

[4] Clavin P. Dynamics of combustion fronts in premixed gases:

from ¯ames to detonations. Proc Comb Inst 2000;28:569±85.

[5] Roberts WL, Driscoll JF. A laminar vortex interacting with a

premixed ¯ame: measured formation of pockets of reactants.

Combust Flame 1991;87:245±56.

[6] Mueller CJ, Driscoll JF, Reuss DL, Drake MC, Rosalik ME.

Vorticity generation and attenuation as vortices convect

through a premixed ¯ame. Combust Flame 1998;112:342±58.

[7] Smooke MD, Mitchel RE, Keyes DE. Numerical solution of

two-dimensional axisymmetric laminar diffusion ¯ames.

Combust Sci Technol 1989;67:85±122.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 260

[8] Givi P. Model free simulations of turbulent reactive ¯ows.

Prog Energy Combust Sci 1989;15:1±107.

[9] Poinsot T. Using direct numerical simulations to understand

premixed turbulent combustion. Twenty-Sixth Symposium

(International) on Combustion. Pittsburgh: The Combustion

Institute, 1996 p. 219±32.

[10] Baritaud T, Poinsot T, Baum M. Direct numerical simulation

for turbulent reacting ¯ows. Paris: Editions Technip, 1996.

[11] Poinsot T, TrouveÂ A, Candel S. Applications of direct

numerical simulations of premixed turbulent combustion.

Prog Energy Combust Sci 1996;21:531±76.

[12] Vervisch L, Poinsot T. Direct numerical simulation of non-

premixed turbulent ¯ame. Annu Rev Fluid Mech

1998;30:655±92.

[13] BarreÁre M, Prud'homme R. Equations fondamentales de

l'aeÂrothermochimie. Paris: Masson, 1973.

[14] Williams FE. Combustion theory. 2nd ed. Reading, MA:

Addison-Wesley, 1985.

[15] Kuo KK. Principles of combustion. New York: Wiley, 1986.

[16] Poinsot T, Veynante D. Theoretical and numerical combus-

tion. Philadelphia, PA: Edwards, 2001.

[17] TrouveÂ A, Poinsot T. The evolution equation for the ¯ame

surface density. J Fluid Mech 1994;278:1±31.

[18] Favre A. Statistical equations of turbulent gases. In: SIAM,

editor. Problems of hydrodynamics and continuum

mechanics. Philadelphia: SIAM, 1969. p. 231±66.

[19] Bray KNC, Libby PA, Masuya G, Moss JB. Turbulence

production in premixed turbulent ¯ames. Combust Sci Tech-

nol 1981;25:127±40.

[20] Shepherd IG, Moss JB, Bray KNC. Turbulent transport in

con®ned premixed ¯ame. 19th Symposium(International) on

Combustion. Pittsburgh: The Combustion Institute, 1982.

p. 423±31.

[21] Piomelli U, Chasnov JR. Large eddy simulations: theory and

applications. In: HallbaÈck H, Henningson DS, Johansson

AV, Alfredsson PH, editors. Turbulence and transition

modelling, Dordrecht: Kluwer Academic Publishers, 1996.

p. 269±336.

[22] Ferziger J. Large eddy simulation: an introduction and

perspective. In: MeÂtais O, Ferziger J, editors. New tools in

turbulence modelling. Berlin: Les Editions de PhysiqueÐ

Springer, 1997. p. 29±47.

[23] Lesieur M. Recent approaches in large-eddy simulations of

turbulence. In: MeÂtais O, Ferziger J, editors. New tools in

turbulence modelling. Berlin: Les Editions de PhysiqueÐ

Springer, 1997. p. 1±28.

[24] Lesieur M, MeÂtais O. New trends in large-eddy simulations

of turbulence. Annu Rev Fluid Mech 1996;28:45±82.

[25] Veynante D, Poinsot T. Reynolds averaged and large

eddy simulation modeling for turbulent combustion. In:

MeÂtais O, Ferziger J, editors. New Tools in Turbulence

Modelling. Berlin: Les Editions de PhysiqueÐSpringer, 1997.

p. 105±40.

[26] Coats CM. Coherent structures in combustion. Prog Energy

Combust Sci 1996;22:427±509.

[27] Menon S, Jou WH. Large eddy simulations of combustion

instability in an axisymmetric ramjet combustor. Combust

Sci Technol 1991;75(1±3):53±72.

[28] Ghosal S, Moin P. The basic equations for the large eddy

simulation of turbulent ¯ows in complex geometry. J Comp

Phys 1995;118:24±37.

[29] Peters N. The turbulent burning velocity for large-scale and

small-scale turbulence. J Fluid Mech 1999;384:107±32.

[30] LinÄaÂn A. The asymptotic structure of counter¯ow diffusion

¯ames for large activation energies. Acta Astronaut

1974;1007(1).

[31] LinÄaÂn A, Orlandi P, Verzicco R, Higuera FJ. Effects of non-

unity Lewis numbers in diffusion ¯ames. Studying turbulence

using numerical databasesÐV. In: Moin P, Reynolds WC,

1994. Stanford, CA: CTR, Stanford University. p. 5±18.

[32] Barlow RS, Fiechtner GJ, Carter CD, Chen J-Y. Experiments

on the scalar structure of turbulent CO/H

2

/N

2

jet ¯ames.

Combust Flame 2000;120(4):549±69.

[33] Bray KNC, Peters N. Laminar ¯amelets in turbulent ¯ames.

In: Libby PA, Williams FA, editors. Turbulent reacting

¯ows. London: Academic Press, 1994. p. 63±113.

[34] Cuenot B, Poinsot TJ. Asymptotic and numerical study of

diffusion ¯ames with variable Lewis number and ®nite rate

chemistry. Combust Flame 1996;104(1):111±37.

[35] Burke SP, Schumann TEW. Diffusion ¯ames. Ind Engng

Chem 1928;20:998±1004.

[36] Kim JS, Williams FA. Extinction of diffusion ¯ames with

nonunity Lewis numbers. J Engng Math 1997;31:101±18.

[37] Clavin P, LinÄaÂn A. Theory of gaseous combustion. An intro-

ductive course. In: Verlarde MG, editor. Non equilibrium

cooperative phenomena in physics and related ®elds. New

York: Plenum Press, 1983.

[38] Favier V, Vervisch L. Edge ¯ames and partially premixed

combustion in diffusion ¯ame quenching. Combust Flame

2001;125(1/2):788±803.

[39] LinÄaÂn A. Ignition and ¯ame spread in laminar mixing layer.

In: Buckmaster J, Jackson TL, Kumar A, editors. Combus-

tion in high speed ¯ows. Dordrecht: Kluwer Academic

Publishers, 1994. p. 461.

[40] Phillips H. Flame in a buoyant methane layer. Tenth Sym-

posium (International) on Combustion. Pittsburgh: The

Combustion Institute, 1965.

[41] Kioni PN, Rogg B, Bray KNC, LinÄaÂn A. Flame spread in

laminar mixing layers: the triple ¯ame. Combust Flame

1993;95:276.

[42] Plessing T, Terhoeven P, Peters N, Mansour MS. An exper-

imental and numerical study on a laminar triple ¯ame.

Combust Flame 1998;115(3):335±53.

[43] Kioni PN, Bray KNC, Greenhalgh DA, Rogg B. Experimen-

tal and numerical studies of a triple ¯ame. Combust Flame

1998;116:192±206.

[44] Dold JW. Flame propagation in a nonuniform mixture:

analysis of a slowly varying triple ¯ame. Combust Flame

1989;76:71±88.

[45] Hartley LJ, Dold JW. Flame propagation in a nonuniform

mixture: analysis of a propagating triple-¯ame. Combust

Sci Technol 1991;80:23±46.

[46] Buckmaster J. Edge-¯ames and their stability. Combust Sci

Technol 1996;115:41±68.

[47] Ghosal S, Vervisch L. Theoretical and numerical investiga-

tion of a symmetrical triple ¯ame using a parabolic ¯ame tip

approximation. J Fluid Mech 2000;415:227±60.

[48] Bourlioux A, Cuenot B, Poinsot T. Asymptotic and nu-

merical study of the stabilization of diffusion ¯ames by hot

gas. Combust Flame 2000;120(1/2).

[49] Ruetsch GR, Vervisch L, LinÄaÂn A. Effects of heat release on

triple ¯ame. Phys Fluids 1995;6(7):1447±54.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 261

[50] Domingo P, Vervisch L. Triple ¯ames and partially premixed

combustion in autoignition of nonpremixed mixtures. Twenty-

Sixth Symposium (International) on Combustion. Pittsburgh:

The Combustion Institute, 1996. p. 233±40.

[51] Plessing T, Terhoven P, Peters N, Mansour MS. An experi-

mental and numerical study on a laminar triple ¯ame.

Combust Flame 1998;115(3):335±53.

[52] Echekki T, Chen JH. Structure and propagation of methanol±

air triple ¯ames. Combust Flame 1998;114(1/2):231±45.

[53] Favier V, Vervisch L. Effects of unsteadiness in edge-¯ames

and liftoff in non-premixed turbulent combustion. Twenty-

Seventh Symposium (International) on Combustion. Pitts-

burgh: The Combustion Institute, 1998. p. 1239±45.

[54] Ghosal S, Vervisch L. Stability diagram for lift-off and blow-

out of a round jet laminar diffusion ¯ame. Combust Flame

2001;124(4)646±55.

[55] Vervisch L. Using numerics to help understanding of nonpre-

mixed turbulent ¯ames. Proc Combust Inst 2000;28:11±24.

[56] Borghi R. ReÂactions chimiques en milieu turbulent. PhD

Thesis. UniversiteÂ Pierre et Marie Curie, Paris 6, 1978.

[57] Villasenor R, Pitz RW, Chen JY. Interaction between chemi-

cal reaction and turbulence in supersonic non-premixed

H

2

±air combustion. In: 29th Aerospace Science Meeting,

Reno (Nevada), AIAA paper 91-0375, 1991.

[58] Nieuwstadt FTM, Meeder JP. Large-eddy simulation of air

pollution dispersion: a review. In: Metais O, Ferziger J,

editors. New tools in turbulence modelling. Berlin: Les

Editions de PhysiqueÐSpringer, 1997. p. 264±80.

[59] Meeder JP, Nieuwstadt FTM. Subgrid-scale segregation of

chemically reactive species in a neutral boundary layer. In:

Chollet JP, Voke PR, Kleiser L, editors. Direct and large

eddy simulation II. Dordrecht: Kluwer Academic Publishers,

1997. p. 301±10.

[60] Bray KNC. Turbulent ¯ows with premixed reactants in

turbulent reacting ¯ows. In: Libby PA, Williams FA, editors.

Topics in applied physics, vol. 44. New York: Springer,

1980. p. 115.

[61] Borghi R. On the structure and morphology of turbulent

premixed ¯ames. Rec Adv Aerosp Sci 1985:117±38.

[62] Borghi R, Destriau M. Combustion and ¯ames, chemical and

physical principles. Paris: Editions Technip, 1998.

[63] Peters N. Laminar ¯amelet concepts in turbulent combustion.

Twenty-First Symposium (International) on Combustion.

Pittsburgh: The Combustion Institute, 1986. p. 1231±50.

[64] Tennekes H, Lumley JL. A ®rst course in turbulence.

Cambridge: MIT Press, 1972.

[65] Poinsot T, Veynante D, Candel S. Quenching processes and

premixed turbulent combustion diagrams. J Fluid Mech

1991;228:561±606.

[66] Roberts WL, Driscoll JF, Drake MC, Goss LP. Images of the

quenching of a ¯ame by a vortex: to quantify regimes of

turbulent combustion. Combust Flame 1993;94:58±69.

[67] Bilger RW. The structure of turbulent non premixed ¯ames.

Twenty-Second Symposium (International) on Combustion.

Pittsburgh: The Combustion Institute, 1988.

[68] Borghi R. Turbulent combustion modelling. Prog Energy

Combust Sci 1988;14:245±92.

[69] Y.Y. Lee, Nonpremixed reacting ¯ows near extinction. PhD

Thesis. Cornell University, 1994.

[70] Cuenot B, Poinsot TJ. Effects of curvature and unsteadiness

in diffusion ¯ames, implications for turbulent diffusion

combustion. Twenty-Fifth Symposium (International) on

Combustion. Pittsburgh: The Combustion Institute, 1994.

[71] Cook A, Riley JJ. Direct numerical simulation of a turbulent

reactive plume on a parallel computer. J Comp Phys

1996;129(2):263±83.

[72] Libby PA, Williams FA. Turbulent combustion: fundamental

aspects and a review. In: Libby PA, Williams FA, editors.

Turbulent reacting ¯ows. London: Academic Press, 1994.

p. 2±61.

[73] Cuenot B, Egolfopoulos FN, Poinsot T. An unsteady laminar

¯amelet model for nonpremixed combustion. Combust

Theor Model 2000;4(1):77.

[74] Mahalingam S, Chen JH, Vervisch L. Finite-rate chemistry

and transient effects in direct numerical simulations of turbu-

lent nonpremixed ¯ames. Combust Flame 1995;102(3):285±

97.

[75] Bray KNC. The challenge of turbulent combustion. Twenty-

Sixth Symposium (International) on Combustion. Pittsburgh:

The Combustion Institute, 1996.

[76] Williams FA. Turbulent combustion. In: Buckmaster JD,

editor. The mathematics of combustion. Philadelphia:

SIAM, 1985. p. 116±51.

[77] Kerstein AR, Ashurst W, Williams FA. Field equation for

interface propagation in an unsteady homogeneous ¯ow

®eld. Phys Rev A 1988;37(7):2728±31.

[78] Poinsot T, Echekki T, Mungal MG. A study of the laminar

¯ame tip and implications for premixed turbulent combus-

tion. Combust Sci Technol 1991;81(1±3):45.

[79] Smiljanovski V, Moser V, Klein R. A capturing±tracking

hybrid scheme for de¯agration discontinuities. Combust

Theor Model 1997;1(2):183±215.

[80] Bourlioux A, Moser V, Klein R. Large eddy simulations of

turbulent premixed ¯ames using a capturing/tracking hybrid

approach. Sixth International Conference on Numerical

Combustion, New Orleans, Louisiana 1996.

[81] Kim W-W, Menon S, Mongia HC. Large-eddy simulation of

a gas turbine combustor ¯ow. Combust Sci Technol

1999;143:25±62.

[82] Piana J, Veynante D, Candel S, Poinsot T. Direct numerical

simulation analysis of the G-equation in premixed combus-

tion. In: Chollet JP, Voke PR, Kleiser L, editors. Direct and

large eddy simulation II. Dordrecht: Kluwer Academic

Publishers, 1997. p. 321±30.

[83] Fichot F, Lacas F, Veynante D, Candel S. One-dimensional

propagation of a premixed turbulent ¯ame with a balance

equation for the ¯ame surface density. Combust Sci Technol

1993:90.

[84] Marble FE, Broadwell JE. The coherent ¯ame model of non-

premixed turbulent combustion. Project Squid TRW-9-PU,

Project Squid Headquarters, Chaffee Hall, Purdue Univer-

sity, 1977.

[85] Candel S, Poinsot T. Flame stretch and the balance equation

for the ¯ame area. Combust Sci Technol 1990;70:1±15.

[86] Pope SB. The evolution of surfaces in turbulence. Int J

Engng Sci 1988;26(5):445±69.

[87] Vervisch L, Bidaux E, Bray KNC, Kollmann W. Surface

density function in premixed turbulent combustion model-

ing, similarities between probability density function and

¯ame surface approach. Phys Fluids A 1995;7(10):2496±

503.

[88] Van Kalmthout E, Veynante D. Direct numerical simulations

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 262

analysis of ¯ame surface density models for non premixed

turbulent combustion. Phys Fluids A 1998;10(9):2347±68.

[89] Peters N. Turbulent combustion. Cambridge: Cambridge

University Press, 2000.

[90] Weller HG, Marooney CJ, Gosman AD. A new spectral

method for calculation of the time-varying area of a laminar

¯ame in homogeneous turbulence. 23rd Symposium (Inter-

national) on Combustion. Pittsburgh: The Combustion

Institute, 1990. p. 629±36.

[91] Weller HG. The development of a new ¯ame area combus-

tion model using conditional averaging. Thermo-Fluids

Section Report TF/9307. Imperial College of Science Tech-

nology and Medicine, 1993.

[92] O'Brien EE. The probability density function (pdf) approach

to reacting turbulent ¯ows. In: Williams FA, Libby PA,

editors. Turbulent reacting ¯ows. London: Academic Press,

1980. p. 185.

[93] Pope SB. Pdf method for turbulent reacting ¯ows. Prog

Energy Combust Sci 1985;11:119±95.

[94] Kollmann W. The pdf approach to turbulent ¯ow. Theor

Comp Fluid Dynam 1990;1:285±349.

[95] Dopazo C. Recent developments in pdf methods. In: Libby

PA, Williams FA, editors. Turbulent reacting ¯ows. London:

Academic Press, 1994. p. 375±474.

[96] Dopazo C, Valino L, Fuego F. Statistical description of the

turbulent mixing of scalar ®elds. Int J Mod Phys B

1997;11(25).

[97] Chen JY, Kollmann W. Pdf modeling of chemical nonequi-

librium effects in turbulent non-premixed hydrocarbon

¯ames. Twenty-Second Symposium (International) on

Combustion. Pittsburgh: The Combustion Institute, 1988.

[98] Abramowitz M, Stegun IA. Handbook of mathematical func-

tions. NY, USA: Dover, 1970.

[99] Bradley D, Lawes M, Scott MJ. Combust Flame

1994;99:581±90.

[100] Lundgren TS. Distribution function in the statistical theory of

turbulence. Phys Fluids 1967;10(5):969±75.

[101] Pope SB. Computations of turbulent combustion: progress

and challenges. Twenty-Third Symposium (International)

on Combustion. Pittsburgh: The Combustion Institute, 1990.

[102] Klimenko AY. Fluid Dynam 1990;25:327.

[103] Bilger RW. Conditional moment closure for turbulent react-

ing ¯ow. Phys Fluids 1993;5(2):327±34.

[104] Chen J-Y, Kollmann W. Comparison of prediction and

measurement in non-premixed turbulent ¯ames. In: Libby

PA, Williams FA, editors. Turbulent reacting ¯ows. London:

Academic Press, 1994. p. 213±308.

[105] Klimenko AY, Bilger RW. Prog Energy Combust Sci

1999;25(6):595±687.

[106] Smith NSA. Conditional moment closure of mixing and reac-

tion in turbulent nonpremixed combustion. In: Moin P,

Reynolds WC, editors. Annual research briefs. Stanford,

CA: CTR, Stanford University, 1996. p. 85±99.

[107] Swaminathan N, Bilger RW. Assessment of combustion

submodels for turbulent nonpremixed hydrocarbon ¯ames.

Combust Flame 1999;116(4):519±45.

[108] Borghi R. Turbulent premixed combustion: further discus-

sions on the scales of ¯uctuations. Combust Flame

1990;80:304±12.

[109] Abdel-Gayed RG, Bradley D, Hamid MN, Lawes M. Lewis

number effects on turbulent burning velocity. 20th Sympo-

sium (International) on Combustion. Pittsburgh: The

Combustion Institute, 1984. p. 505±12.

[110] Abdel-Gayed RG, Bradley D. Combustion regimes and the

straining of turbulent premixed ¯ames. Combust Flame

1989;76:213±8.

[111] Yakhot V, Orszag CG, Thangam S, Gatski TB, Speziale

CG. Development of turbulence models for shear ¯ows by

a double expansion technique. Phys Fluids 1992;4(7):

1510.

[112] Gouldin FC. Combustion intensity and burning rate integral

of premixed ¯ames. Twenty-Sixth Symposium (Interna-

tional) on Combustion. Pittsburgh: The Combustion Insti-

tute, 1996. p. 381±8.

[113] Im HG, Lund TS, Ferziger JH. Large eddy simulation of

turbulent front propagation with dynamic subgrid models.

Phys Fluids A 1997;9(12):3826±33.

[114] Spalding DB. Mixing and chemical reaction in steady

con®ned turbulent ¯ames. 13th Symposium (International)

on Combustion. Pittsburgh: The Combustion Institute,

1971. p. 649±57.

[115] Said R, Borghi R. A simulation with a cellular automaton for

turbulent combustion modelling, Twenty-Second Sym-

posium (International) on Combustion. Pittsburgh: The

Combustion Institute, 1988 p. 569±77.

[116] Moss JB, Bray KNC. A uni®ed statistical model of the

premixed turbulent ¯ame. Acta Astronaut 1977;4:291±319.

[117] Mantel T, Borghi R. A new model of premixed wrinkled

¯ame propagation based on a scalar dissipation equation.

Combust Flame 1994;96(4):443.

[118] Lahjaily H, Champion M, Karmed D, Bruel P. Introduction

of dilution in the BML model: application to a stagnating

turbulent ¯ame. Combust Sci Technol 1998;135:153±73.

[119] Poinsot T, Veynante D, TrouveÂ A, Ruetsch G. Turbulent

¯ame propagation in partially premixed ¯ames. Proceedings

of the 1996 Summer Program. Center for Turbulence

Research, 1996. p. 111±36.

[120] Veynante D, Duclos JM, Piana J. Experimental analysis of

¯amelet models for premixed turbulent combustion, Twenty-

Fifth Symposium (International) on Combustion. Pittsburgh:

The Combustion Institute, 1994.

[121] Veynante D, Piana J, Duclos JM, Martel C. Experimental

analysis of ¯ame surface density model for premixed turbu-

lent combustion. Twenty-Sixth Symposium (International)

on Combustion. Pittsburgh: The Combustion Institute,

1996. p. 413±20.

[122] Duclos JM. Etude theÂorique et expeÂrimentale d'une ¯amme

turbulente stabiliseÂe par un barreau cylindrique. PhD Thesis.

Ecole Centrale Paris, 1997.

[123] Martel C. Etude expeÂrimentale de la combustion turbulente

preÂmeÂlangeÂe. Analyse de modeÁles. PhD Thesis. Ecole

Centrale Paris, 1998.

[124] Bray KNC, Champion M, Libby PA. The interaction

between turbulence and chemistry in premixed turbulent

¯ames. In: Borghi R, Murphy SN, editors. Turbulent reacting

¯ows, Lecture Notes in Engineering, vol. 40. Berlin:

Springer, 1989. p. 541±63.

[125] Gouldin FC. Review of closure models for the ¯amelet

regime of premixed turbulent combustion. Joint Meeting of

the British and German Sections of the Combustion Institute,

Queens' College, Cambridge, UK, 1993. p. 52±5.

[126] Gouldin FC, Miles PC. Chemical closure and burning rates in

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 263

premixed turbulent ¯ames. Combust Flame 1995;100(1/

2):201.

[127] Meneveau C, Poinsot T. Stretching and quenching of ¯ame-

lets in premixed turbulent combustion. Combust Flame

1991;86:311±32.

[128] Bailly P. Contribution l'eÂtude de l'interaction turbulenceÐ

combustion dans les ¯ammes turbulentes de preÂmeÂlange aÁ

l'aide de modeÁles du second ordre. PhD Thesis. UniversiteÂ de

Poitiers, 1996.

[129] Bailly P, GarreÂton D, Simonin O, Bruel P, Champion M,

Deshaies B, Duplantier S, Sanquer S. Experimental and

numerical study of a premixed ¯ame stabilized by a rectan-

gular section cylinder, Twenty-Sixth Symposium (Interna-

tional) on Combustion. Pittsburgh: The Combustion

Institute, 1996. p. 923±30.

[130] Duclos JM, Veynante D, Poinsot T. A comparison of ¯amelet

models for premixed turbulent combustion. Combust Flame

1993;95(1/2):101±18.

[131] Bray KNC, Champion M, Libby PA. Premixed ¯ames in

stagnating turbulence part iv: a new theory for the Reynolds

stresses and Reynolds ¯uxes applied to impinging ¯ows.

Combust Flame 2000;120(1/2):1±18.

[132] Domingo P, Bray KNC. Laminar ¯amelet expressions for

pressure ¯uctuation terms in second moment models of

premixed turbulent combustion. Combust Flame

2000;121(4):555±74.

[133] Gouldin FC, Bray KNC, Chen JY. Chemical closure model

for fractal ¯amelets. Combust Flame 1989;77:241±59.

[134] Cant RS, Pope SB, Bray KNC. Modelling of ¯amelet surface

to volume ratio in turbulent premixed combustion. 23rd

Symposium (International) on Combustion, Orleans. Pitts-

burgh: The Combustion Institute, 1990. p. 809±15.

[135] Prasad ROS, Gore JP. An evaluation of ¯ame surface density

models for turbulent premixed jet ¯ames. Combust Flame

1999;116(1±2):1±14.

[136] Cheng WK, Diringer JA. Numerical modelling of SI engine

combustion with a ¯ame sheet model. SAE paper 910268.

International Congress and Exposition, Detroit, 1991.

[137] Choi CR, Huh KY. Development of a coherent ¯amelet

model for a spark ignited turbulent premixed ¯ame in a

closed vessel. Combust Flame 1998;114(3/4):336±48.

[138] Checkel MD, Thomas A. Turbulent combustion of premixed

¯ames in closed vessels. Combust Flame 1994;96(4):351±

70.

[139] Prasad ROS, Paul RN, Sivathanu YR, Gore JP. An evaluation

of combined ¯ame surface density and mixture fraction

models for nonisenthalpic premixed turbulent ¯ames.

Combust Flame 1999;117(3):514±28.

[140] Veynante D, TrouveÂ A, Bray KNC, Mantel T. Gradient and

counter-gradient scalar transport in turbulent premixed

¯ames. J Fluid Mech 1997;332:263±93.

[141] Dopazo C. On conditioned averages for intermittent turbu-

lent ¯ows. J Fluid Mech 1977;81:433±8.

[142] TrouveÂ A. The production of premixed ¯ame surface area in

turbulent shear ¯ow. Combust Flame 1994;99:687±96.

[143] Libby PA, Bray KNC. Countergradient diffusion in premixed

turbulent ¯ames. AIAA J 1981;19:205±13.

[144] Masuya G, Libby P. Non gradient theory for oblique turbulent

¯ame with premixed reactants. AIAA J 1981;19:205±13.

[145] TrouveÂ A, Veynante D, Bray KNC, Mantel T. The coupling

between ¯ame surface dynamics and species mass conserva-

tion in premixed turbulent combustion. In: Center for Turbu-

lence Research, editor. Proceedings of the Summer Program,

Stanford, Stanford: Center for Turbulence Research, 1994.

p. 95±124.

[146] Rutland C, Cant RS. Turbulent transport in premixed ¯ames.

In: Moin P, Reynolds WC, editors. Studying turbulence

using numerical databasesÐV. Stanford, CA: CTR, Stan-

ford University, 1994. p. 75±94.

[147] Kalt PAM. Experimenal investigation of turbulent scalar ¯ux

in premixed ¯ames. PhD Thesis. University of Sydney, 1999.

[148] Boughanem H. Evaluation des termes de transport et de

dissipation de surface de ¯amme par simulation numeÂrique

directe de la combustion turbulente. PhD Thesis. University

of Rouen, France, 1998.

[149] Boughanem H, TrouveÂ A. The occurence of ¯ame instabil-

ities in turbulent premixed combustion. 27th Symposium

International on Combustion 1998.

[150] Paul RN, Bray KNC. Study of premixed turbulent combus-

tion including Landau±Darrieus instability effects. Twenty-

Sixth Symposium (International) on Combustion. Pittsburgh:

The Combustion Institute, 1996. p. 259±66.

[151] Bray KNC, Moss JB, Libby PA. Turbulent transport in

premixed turbulent ¯ames. In: Zierep J, Oertel H, editors.

Convective transport and instability phenomena. Germany:

University of Karlsuhe, 1982.

[152] Veynante D, Poinsot T. Effects of pressure gradients on

turbulent premixed ¯ames. J Fluid Mech 1997;353:83±114.

[153] Libby PA. Theoretical analysis of the effect of gravity on

premixed turbulent ¯ames. Combust Sci Technol

1989;68:15±33.

[154] Chen JY, Lumley JL, Gouldin FC. Modeling of wrinkled

laminar ¯ames with intermittency and conditional statistics.

21st Symposium (International) on Combustion. Pittsburgh:

The Combustion Institute, 1986. p. 1483±91.

[155] Boger M, Veynante D. Large eddy simulations of a turbulent

premixed V-shape ¯ame. Eighth European Turbulence

Conference, Barcelona 2000.

[156] Koochesfahani MM, Dimotakis PE. Mixing and chemical

reactions in turbulent liquid mixing layer. J Fluid Mech

1986;170:83±112.

[157] McMurty PA, Riley JJ, Metcalfe RW. Effects of heat release

on the large-scale structure in turbulent mixing layer. J Fluid

Mech 1989;199:297±332.

[158] Hermanson JC, Dimotakis PE. Effects of heat release in

turbulent, reacting shear layer. J Fluid Mech 1989;199:

333±75.

[159] Jones WP. Turbulence modelling and numerical solution

methods for variable density and combusting ¯ows. In:

Libby PA, Williams FA, editors. Turbulent reacting ¯ows,

London: Academic Press, 1994. p. 309±68.

[160] Magnussen BF, Hjertager BH. On the mathematical model-

ing of turbulent combustion with special emphasis on soot

formation and combustion. Sixteenth Symposium (Interna-

tional) on Combustion. Pittsburgh: The Combustion Insti-

tute, 1976. p. 719±29.

[161] Ferziger JH, PericÂ M. Computational methods for ¯uid

dynamics. Berlin: Springer, 1996.

[162] Coupland J, Priddin CH. Modelling of ¯ow and combustion

in a production gas turbine combustor. In: Durst F, Launder

BE, Lumley JL, Schmidt FW, Whitelaw JH, editors. Turbu-

lent shear ¯ows 5th. Berlin: Springer, 1987. p. 310±23.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 264

[163] BarlowRS, Dibble RW, Chen JY, Lucht RP. Combust Flame

1990(82).

[164] Montgomery CJ, Kosaly G, Riley JJ. Direct numerical simu-

lation of turbulent H

2

±O

2

combustion using reduced chem-

istry. In: AIAA Paper 93-0248. 31st Aerospace Sciences

Meeting and Exhibit, Reno, NV, 1993. p. 1±10.

[165] Peters N. Prog Energy Combust Sci page 1984:319.

[166] Sun CJ, Sung CJ, Wang H, Law CK. On the structure of

nonsooting counter¯ow ethylene and acetylene diffusion

¯ames. Combust Flame 1996;107(4):321±35.

[167] Sung CJ, Liu JB, Law CK. Structural response of counter-

¯ow diffusion ¯ames to strain rate variations. Combust

Flame 1995;102:481±92.

[168] Rogg B. The Cambridge universal laminar ¯amelet computer

code. In: Peters N, Rogg B, editors. Reduced kinetic mechan-

isms for applications in combustion systems, Appendix C.

Berlin: Springer, 1993.

[169] Nandula SP, Brown TM, Pitz RW. Measurements of scalar

dissipation in the reaction zones of turbulent nonpremixed

H

2

±air ¯ames. Combust Flame 1994;99:775±83.

[170] Bish ES, Dahm WJA. A strained dissipation and reaction

layer formulation for turbulent diffusion ¯ames. Western

States Section Fall Meeting. Pittsburgh: The Combustion

Institute, 1993.

[171] Ferreira JC. Flamelet modelling of stabilization in turbulent

non-premixed combustion. PhD Thesis. Swiss Federal Insti-

tute of Technology, ETH, Zurich, 1996.

[172] Haworth DC, Drake MC, Pope SB, Blint RJ. The importance

of time-dependent ¯ame structures in stretched laminar

¯amelet models for turbulent jet diffusion ¯ames. Twenty-

Second Symposium (International) on Combustion. Pitts-

burgh: The Combustion Institute, 1988. p. 589±97.

[173] Pitsch H, Wan YP, Peters N. SAE Paper, 952357, 1995.

[174] Zhang Y, Rogg B, Bray KNC. 2-d Simulation of turbulent

autoignition with transient laminar ¯amelet source term

closure. Combust Sci Technol 1995;105:211±27.

[175] Darabiha N. Transient behaviour of laminar counter¯ow

hydrogen±air diffusion ¯ames with complex chemistry.

Combust Sci Technol 1992;86:163±81.

[176] Mauss F, Keller D, Peters N. Alagrangian simulation of ¯ame-

let extinction and re-ignition in turbulent jet diffusion ¯ames.

Twenty-Third Symposium (International) on Combustion.

Pittsburgh: The Combustion Institute, 1990. p. 693±8.

[177] Pitsh H, Chen M, Peters N. Unsteady ¯amelet modeling of

turbulent hydrogen±air diffusion ¯ames. Twenty-Seventh

Symposium (International) on Combustion. Pittsburgh: The

Combustion Institute, 1998. p. 1057±64.

[178] Mell WE, Nilsen VN, KosaÂly G, Riley JJ. Direct numerical

simulation investigation of the conditional moment closure

model for nonpremixed turbulent reacting ¯ows. Combust

Sci Technol 1991;91:179±86.

[179] Delhaye B. Etude des ¯ammes de diffusion turbulentes.

Simulations directes et modeÂlisation. PhD Thesis. Ecole

Centrale Paris, 1994.

[180] Delhaye B, Veynante D, Candel S. Simulation and modeling

of reactive shear layers. Theoret Comput Fluid Dynam

1994;6:67±87.

[181] Chang CHH, Dahm WJA, Trygvason G. Lagrangian model

simulations of molecular mixing, including ®nite rate chemi-

cal reactions, in a temporally developing shear layer. Phys

Fluid 1991;3(5):1300±11.

[182] Lee YY, Pope SB. Nonpremixed turbulent reacting ¯ow near

extinction. Combust Flame 1995;101:501±28.

[183] Bradley D, Gaskell PH, Gu XJ. The mathematical modeling

of liftoff and blowoff of turbulent non-premixed methane jet

¯ames at high strain rate. Proc Combust Inst 1998;27:1199±

206.

[184] Cuenot B, Egolfopoulos FN, Poinsot T. An unsteady laminar

¯amelet model for non-premixed combustion. Combust

Theor Mod 2000;4:77±97.

[185] MuÈller CM, Breitbach H, Peters N. Partially premixed turbu-

lent ¯ame propagation in jets ¯ames. Twenty-Fifth Sympo-

sium (International) on Combustion. Pittsburgh: The

Combustion Institute, 1994. p. 1099±106.

[186] Fichot F, Delhaye B, Veynante D, Candel S. Strain rate

modelling for a ¯ame surface density equation with applica-

tion to non-premixed turbulent combustion, 25th Symposium

(International) on Combustion. Pittsburgh: The Combustion

Institute, 1994.

[187] van Kalmthout E, Veynante D, Candel S. Direct numerical

simulation analysis of ¯ame surface density equation in non-

premixed turbulent combustion, Twenty-Sixth Symposium

(International) on Combustion. Pittsburgh: The Combustion

Institute, 1996.

[188] Van Kalmthout E, Veynante D. Analysis of ¯ame surface

density concepts in non-premixed turbulent combustion

using direct numerical simulation. Eleventh Symposium on

Turbulent Shear Flows, Grenoble, France 1997.

[189] Obounou M, Gonzalez M, Borghi R. A lagrangian model for

predicting turbulent diffusion ¯ames with chemical kinetic

effects. Twenty-Fifth Symposium (International) on Combus-

tion. Pittsburgh: The Combustion Institute, 1994. p. 1107±13.

[190] Broadwell JE, Lutz A. Aturbulent jet chemical reaction model:

NO

x

production in jet ¯ames. Combust Flame 1998:114.

[191] Ravet F, Vervisch L. Modeling non-premixed turbulent

combustion in aeronautical engines using pdf-generator.

36th Aerospace Sciences Meeting and Exhibit AIAA paper

98-1027, Reno, NV 1998.

[192] Girimaji SS. Phys Fluids 1992;4:2529.

[193] Smith NSA, Ruetsch GR, Oefelein J, Ferziger JH. Simulation

and modeling of reactive particles in turbulent nonpremixed

combustion. In: Moin P, Reynolds WC, editors. Studying

turbulence using numerical databasesÐVII. Stanford, CA:

CTR, Stanford University, 1998. p. 39±60.

[194] Kerstein AR. A linear-eddy model for turbulent scalar trans-

port and mixing. Combust Sci Technol 1988;60:391±421.

[195] Fox RO. The lagrangian spectral relaxation model for differ-

ential diffusion in homogeneous turbulence. Phys Fluids

1999;11(6):1550±71.

[196] Fox RO. The spectral relaxation model of the scalar dissipa-

tion rate in homogeneous turbulence. Phys Fluids

1995;7(5):1082±94.

[197] Van Slooten Jayesh PR, Pope SB. Advances in pdf modeling

for inhomogeneous turbulent ¯ows. Phys Fluids 1998;10(1):

246±65.

[198] Spielman LA, Levenspiel O. A Monte Carlo treatment for

reacting and coalescing dispersed phase systems. Chem

Engng Sci 1965;20:247±54.

[199] Valino L. A®eld Monte Carlo formulation for calculating the

probability density function of a single scalar in turbulent

¯ow. Flow Turb Combust 1998;60(2):157±72.

[200] Fox RO. Computational methods for turbulent reacting ¯ows

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 265

in the chemical process industry. Revue de L'Institut Fran-

cËais du PeÂtrole 1996;51(2).

[201] Corrsin S. The isotropic turbulent mixer. Part II. Arbitrary

Schmidt number. AIChE J 1964;10:870±7.

[202] Dopazo C. Non-isothermal turbulent reactive ¯ows: stochas-

tic approaches. PhD Thesis. University of New York, Stony

Brook, 1973.

[203] Villermaux J. Micromixing phenomena in stirred reactors.

Encycl Fluid Mech 1986:707.

[204] Eswaran V, Pope SB. Direct numerical simulation of the

turbulent mixing of a passive scalar. Phys Fluids

1988;31(3):506±20.

[205] Tsai K, Fox RO. The bmc/giem model for micromixing in

non-premixed turbulent reacting ¯ows. Ind Engng Chem Res

1998:6.

[206] Curl RI. Dispersed phase mixing. Theory and effects in

simple reactors. AIChE J 1963;175(9).

[207] Masri AR, Bilger RW, Dibble RW. Turbulent non-premixed

¯ames of methane near extinction. Combust Flame

1988;73:261±8.

[208] Jones WP, Kakhi M. Pdf modeling of ®nite rate chemistry

effects in turbulent nonpremixed jet ¯ames. Combust Flame

1998;115(1/2):210±29.

[209] Jones WP, Lindstedt RP. Global reaction schemes for hydro-

carbon combustion. Combust Flame 1988;73:233±49.

[210] Dopazo C. Relaxation of initial probability density functions

in the turbulent convection of scalar ®elds. Phys Fluids

1979;22(20).

[211] Janicka J, Kolbe W, Kollmann W. Closure of the transport

equation for the probability density function of turbulent

scalar ®elds. J Non-Equilib Thermodyn 1979;4(47).

[212] Pope SB. An improved turbulent mixing model. Combust Sci

Technol 1982;28:131±5.

[213] Pope SB. Consistent modeling of scalars in turbulent ¯ows.

Phys Fluids 1982;26:404±8.

[214] Leonard A, Hill J. Direct numerical simulation of turbulent

¯ows with chemical reaction. J Scient Comput

1988;3(1):25±43.

[215] Kraichnan RH. Closures for probability distributions. Bull

Am Phys Soc 1989;34:2298.

[216] Dutta A, Tarbell JM. Closure models for turbulent reacting

¯ows. AIChE J 1989;35(12).

[217] Gao F. An analytical solution for the scalar probability

density function in homogeneous turbulence. Phys Fluids

1991;3(4):511±3.

[218] Valino L, Dopazo C. A binomial sampling model for scalar

turbulent mixing. Phys Fluids A 1990;2(7):1204±12.

[219] Norris AT, Pope SB. Turbulent mixing model based on

ordered pairing. Combust Flame 1991;83:27±42.

[220] Valino L, Dopazo C. A binomial langevin model for turbu-

lent mixing. Phys Fluids A 1991;3(12):3034±7.

[221] Valino L, Ros J, Dopazo C. Monte Carlo implementation and

analytic solution of an inert-scalar turbulent-mixing test

problem using mapping closure. Phys Fluids A

1991;3(9):2191±8.

[222] Pope SB. Mapping closure for turbulent mixing and reaction.

Theoret Comput Fluid Dynam 1991;2:255±70.

[223] Gao F. Mapping closure and non-gaussianity of the scalar

probability density functions in isotropic turbulence. Phys

Fluids A 1991;3(10):2438±44.

[224] Gao F, O'Brien EE. Joint probability density function of a

scalar and its gradient in isotropic turbulence. Phys Fluids

1991;3(6).

[225] Gao F, O'Brien EE. A mapping closure for multispecies

®ckian diffusion. Phys Fluids 1991;3(3).

[226] Kerstein AR. Linear-eddy modelling of turbulent transport.

Part 6. Microstructure of diffusive scalar mixing ®elds. J

Fluid Mech 1991;231:361±94.

[227] Chen J-Y, Kollmann W. Pdf modeling and analysis of ther-

mal NO formation in turbulent nonpremixed hydrogen±air

jet ¯ames. Combust Flame 1992;88:397±412.

[228] Jiang T-L, Givi P, Gao F. Binary and trinary scalar mixing by

Fickian diffusionÐsome mapping closure results. Phys

Fluids A 1992;4(5):1028±35.

[229] Kerstein AR. Linear-eddy modelling of turbulent transport.

Part 7. Finite rate chemistry and multi-stream mixing. J Fluid

Mech 1992;240:289±313.

[230] Fox RO. Improved Fokker±Planck model for the joint scalar,

scalar gradient pdf. Phys Fluids 1994;1(6):334±48.

[231] Masri AR, Subramaniam S, Pope SB. A mixing model to

improve the pdf simulation of turbulent diffusion ¯ames.

Twenty-Sixth Symposium (International) on Combustion.

Pittsburgh: The Combustion Institute, 1996.

[232] Fox RO. On velocity-conditioned scalar mixing in homo-

geneous turbulence. Phys Fluids 1996;8(10):2678±91.

[233] Fox RO. On the relationship between lagrangian micromix-

ing models and computational ¯uid dynamics. Chem Engng

Process 1998;37:521±35.

[234] ReÂveillon J, Vervisch L. Subgrid mixing modeling: a

dynamic approach. AIAA J 1998;36(3):336±41.

[235] Subramaniam S, Pope SB. A mixing model for turbulent

reactive ¯ows based on Euclidean minimum spanning

trees. Combust Flame 1998;115(4):487±514.

[236] Subramaniam S, Pope SB. Comparison of mixing model

performance for nonpremixed turbulent reacting ¯ows.

Combust Flame 1998;117(4):732±54.

[237] Anand MS, Pope SB. Calculations of premixed turbulent

¯ames by pdf methods. Combust Flame 1987:67.

[238] Haworth DC, Pope SB. A generalized Langevin model for

turbulent ¯ows. Phys Fluids 1986;29:208±28.

[239] Gardiner CW. Handbook of stochastic methods. Berlin:

Springer, 1997.

[240] Pope SB. On the relationship between stochastic Lagrangian

models of turbulence and second-moment closures. Phys

Fluids 1994;6:973±85.

[241] Wouters HA, Peeters TWJ, Roekaerts D. On the existence of

a stochastic Lagrangian model representation for second-

moment closures. Phys Fluids 1996;8(7):1702±4.

[242] Pope SB. Lagrangian pdf methods for turbulent ¯ows. Annu

Rev Fluid Mech 1994;26:23.

[243] ReÂveillon J, Vervisch L. Spray vaporization in non-premixed

turbulent ¯ames: a single droplet model. Combust Flame

2001;121(1/2):75±90.

[244] Hollmann C, Gutheil E. Modeling of turbulent spray diffu-

sion ¯ames including detailed chemistry. Twenty-Sixth

Symposium (International) on Combustion. Pittsburgh: The

Combustion Institute, 1996. p. 1731±8.

[245] Calimez X. Simulation aÁ petite eÂchelle par une meÂthode VOF

d'eÂcoulement diphasiques reÂactifs. PhD Thesis. Ecole

Centrale Paris, 1998.

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 266

194

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266

6. Tools for turbulent combustion modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2. Scalar dissipation rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3. Geometrical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1. G-®eld equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.2. Flame surface density description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.3. Flame wrinkling description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4. Statistical approaches: probability density function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.2. Presumed probability density functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.3. Pdf balance equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.4. Joint velocity/concentrations pdf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.5. Conditional moment closure (CMC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5. Similarities and links between the tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7. Reynolds-averaged models for turbulent premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1. Turbulent ¯ame speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2. Eddy-break-up model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3. Bray±Moss±Libby model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.2. BML model analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.3. Recovering mean reaction rate from tools relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.4. Reynolds and Favre averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.5. Conditional averagingÐcounter-gradient turbulent transport . . . . . . . . . . . . . . . . . . . . . 7.3.6. Extensions to partially premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4. Models based on the ¯ame surface area estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.2. Algebraic expressions for the ¯ame surface density S . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.3. Flame surface density balance equation closures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.4. Analysis of the ¯ame surface density balance equation . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.4.1. Turbulent transport. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.4.2. Strain rate induced by the mean ¯ow ®eld, AT. . . . . . . . . . . . . . . . . . . . . . . . 7.4.4.3. Strain rate aT due to turbulent motions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.4.4. Propagation and curvature terms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.5. Flame stabilization modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.6. A related approach: G-equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8. Turbulent transport in premixed combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2. Direct numerical simulation analysis of turbulent transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3. Physical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.1. Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4. External pressure gradient effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5. Counter-gradient transportÐexperimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6. To include counter-gradient turbulent transport in modeling . . . . . . . . . . . . . . . . . . . . . 8.7. Towards a conditional turbulence modeling? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9. Reynolds averaged models for non-premixed turbulent combustion . . . . . . . . . . . . . . . . . . . . . . 9.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2. Fuel/air mixing modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ~ 9.2.2. Balance equation and simple relaxation model for x . . . . . . . . . . . . . . . . . . . 9.3. Models assuming in®nitely fast chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.1. Eddy dissipation model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.2. Presumed pdf: in®nitely fast chemistry model . . . . . . . . . . . . . . . . . . . . . . . . 9.4. Flamelet modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

212 212 214 214 214 216 218 218 218 219 219 221 221 221 223 223 223 224 224 225 226 227 227 228 229 229 229 232 234 234 234 235 236 237 237 237 237 238 238 238 239 241 242 242 243 244 245 245 246 246 246 247 247 247 248 248

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266

195

9.4.2. Flame structure in composition space, ................... 9.4.3. Mixing modeling in SLFM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5. Flame surface density modeling, coherent ¯ame model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.6. MIL model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.7. Conditional moment closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8. Pdf modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.1. Turbulent micromixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.2. Linear relaxation model, IEM/LMSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.3. GIEM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.4. Stochastic micromixing closures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.5. Interlinks PDF/¯ame surface modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.6. Joint velocity/concentrations pdf modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Y iSLFM Z p ; xp

249 251 252 252 253 254 254 255 256 256 257 257 258 260 260

1. Introduction The number of combustion systems used in power generation and transportation industries is growing rapidly. This induces pollution and environmental problems to become critical factors in our societies. The accurate control of turbulent ¯ames therefore appears as a real challenge. Computing is now truly on par with experiment and theory as a research tool to produce multi-scale information that is not available by using any other technique. Computational ¯uid dynamics (CFD) is ef®ciently used to improve the design of aerodynamical systems, and today no real progress in design can be made without using CFD. With the same objectives, much work has been devoted to turbulent combustion modeling, following a variety of approaches and distinct modeling strategies. This paper is intended to provide a generic review of these numerical models. A wide range of coupled problems are involved in turbulent ¯ames: ² The ¯uid mechanical properties of the combustion system must be well known to carefully describe the mixing between reactants and, more generally, all transfer phenomena occurring in turbulent ¯ames (heat transfer, molecular diffusion, convection, turbulent transport, etc.). ² Detailed chemical reaction schemes are necessary to estimate the consumption rate of the fuel, the formation of combustion products and pollutant species. A precise knowledge of the chemistry is absolutely required to predict ignition, stabilization or extinction of reaction zones together with pollution. ² Two (liquid fuel) and three (solid fuel) phase systems may be encountered. Liquid fuel injection is a common procedure and the three-dimensional spatial distribution of gaseous reactants depends on complex interactions

between the breakdown of the liquid sheets, the vaporization of the liquid, turbulent mixing, and droplet combustion. ² Radiative heat transfer is generated within the ¯ame by some species and carbon particles resulting from soot formation and transported by the ¯ow motion. In furnaces, walls also interact with combustion through radiative transfer. Turbulent combustion modeling is therefore a very broad subject. All the aspects of the problem are not addressed in the present review. We will only focus on the closure schemes developed and used to understand and calculate turbulent transport and mean burning rates in turbulent ¯ames. The detail of chemistry, its reduction, tabulation, etc. are not considered. However, the links existing between the models are identi®ed, showing similarities which are sometimes much stronger than is usually thought. Numerical modeling of ¯ames is developed from the following steps (Fig. 1): ² Under assumptions such as the high activation energy limit, asymptotic analysis [1±4] allows the analytical determination of ¯ame properties in well-de®ned model problems (ignition, propagation of ¯ame front, instabilities and acoustics, etc.). This approach, limited to simpli®ed situations, leads to analytical results exhibiting helpful scale factors (dimensionless numbers) and major ¯ame behaviors. Asymptotic analysis is particularly well suited to perform quantitative comparison between various phenomena. ² Simpli®ed experiments are useful to understand the basic properties of combustion (laminar ¯ames, ¯ame/vortex interactions, etc.) [5,6]. These experiments are accompanied by numerical simulations of laminar ¯ames incorporating complex chemistry and multi-species transports along with radiative heat losses [7].

3): 2ruj ui 2tij 2rui 2p 1 2 1 1 Fi 2xi 2t 2xj 2xj 2 1 . the ultimate step is the simulation of a real combustion device.196 D. a ®rst part is devoted to a brief description of laminar ¯ames (Section 3). ² For given chemistry and transport model. Because of the large number of degrees of freedom involved in turbulent combustion. Once the models have been implemented in numerical codes. understand ¯ames and improve models. These instantaneous local balance equations are. The next three sections are devoted to combustion modeling in the context of Reynolds Averaged Navier±Stokes (RANS) equations.2. implementation of these models into CFD. and to isolate some speci®c phenomena (heat release. Lewis number¼) [8±12]. 2. Following a short presentation of the balance equations for reactive ¯ows (Section 2). using the classical lettering [13±15]: ² Mass: 2ruj 2r 0 1 2t 2xj ² Momentum (i 1. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. Then. a full DNS of a practical system cannot be performed and averaging techniques leading to unclosed equations are necessary. 1 is quite formal. validation procedures are required. species and energy transport equations. we review the available closures for the mean reaction rate (Section 7) and turbulent transport (Section 8). the physical analysis leading to turbulent combustion diagrams is developed (Section 5). For premixed turbulent combustion. in direct numerical simulation (DNS) all the scales of the turbulence (time and length) are calculated without resorting to closures for turbulent ¯uxes and mean burning rate. Veynante.1. Turbulent ¯ames are analyzed in simple con®gurations to extract data impossible to measure in experiments. Con®gurations as close as possible to actual industrial systems are chosen for these tests. Instantaneous balance equations The basic set of balance equations comprises the classical Navier±Stokes. Balance equations 2. L. Models for turbulent ¯ames are then developed: closure techniques are proposed for unknown terms found in exact averaged balance equations. the modeling of the mean burning rate in non-premixed turbulent ¯ames is addressed. modeling tools available to derive turbulent combustion models are described and the relations between a priori quite different formalisms are established (Section 6). experimental measurements and comparison between these experimental data and the numerical results. After a presentation of the unsuccessful Taylor's expansion for closing the mean burning rate (Section 4). The decomposition discussed in Fig. Combustion modeling steps. Turbulent combustion modeling is actually a continuous ring between theoretical studies to analyze combustion. Then. 1. The numerical modeling is validated against measurements obtained from experiments. In a subsequent section (Section 9).

Reynolds and Favre averaging Unfortunately. In practical situations.e. brings unclosed correlations such as u H Q H that are unknown and must be modeled. One direct consequence of these instabilities is an increase of the premixed ¯ame area and of the global reaction rate [14. Averaging the mass balance equation leads to: 2r 2 1 rui 1 r H u Hi 0 2t 2xi 12 where the velocity/density ¯uctuations correlation r H u Hi appears. the full numerical solution of the instantaneous balance equations is limited to very simpli®ed cases (DNS [9. and Reynolds averaging induces some additional dif®culties. The numerical procedure is called Reynolds Averaged Navier±Stokes (RANS) modeling.17]. ¯uctuations of density are observed because of the thermal heat release. ui tij and ui Fj are negligible (low Mach number assumption) [16].12]). ~ a Favre (mass weighted) [18] average Q is introduced and ~ any quantity is then decomposed into Q Q 1 Q HH : rQ ~ Q . d ij is the Kronecker symbol. Veynante. Species molecular diffusivities are generally described using the Fick law. i. where the number of time and length scales present in the ¯ow is not too great. ² Species (N species with k 1. Soret effect (species diffusion under temperature gradients) and molecular transport due to pressure gradients are usually neglected. the enthalpy diffusive ¯ux (Eq. the viscous tensor is given by the Newton law: 2 3 2uj 2ui 2 2uk 2 ml dij tij ml 1 5 3 2xj 2xi 2xk where the molecular viscosity ml depending on the ¯uid properties is introduced. ¼.D. These equations are closed by expressions for the species molecular ¯uxes and the viscous forces. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 197 where t ij denotes the viscous force tensor and Fi a body force. the Lewis number Lek of the species k. Nevertheless. ² Total enthalpy ht h 1 ui ui =2 : 2ruj ht 2rht 2p 2 1 1 Jh 1 ui tij 1 uj Fj 2t 2xj j 2t 2xj 4 where ui tij and ui Fj denote respectively the power due to viscous and body forces.11.2. More complex expressions may be used to describe multispecies molecular diffusion. assuming a major species: Jk 2 j ml 2Yk Sck 2xj 6 Sck is the Schmidt number of the species k. 2. To avoid the explicit modeling of such correlations. de®ned as: ml 7 Sck rDk Dk is the molecular diffusivity of the species k relative to the major species. In turbulent ¯ames. Each quantity Q is split into a mean Q and a deviation from the mean denoted by Q H : Q Q 1 QH with QH 0 11 where Jk is the molecular diffusive ¯ux of the species k j and v k the mass reaction rate of this species per unit _ volume. widely used in non-reacting ¯uid mechanics. radiative heat transfer and Dufour effect (enthalpy diffusion under mass fraction gradients) are neglected. an additional step is introduced by averaging the balance equations to describe only the mean ¯ow ®eld (local ¯uctuations and turbulent structures are integrated in mean quantities and these structures have no longer to be described in the simulation). L. r À Á ~ HH f r Q2Q 0 Q r 13 . The Prandtl number is written as a function of the thermal diffusivity l and the constant pressure speci®c heat Cp : mi Cp Pr 9 l Then. This assumption is generally made to simplify turbulent ¯ame modeling. Enthalpy diffusion is described according to the Fourier law: 4 5 N Pr ml 2h 2Yk h Jj 2 8 1 2 1 hk Sck Pr 2xj 2xj k1 The Prandtl number Pr compares the diffusive transport of momentum (viscous forces) and temperature. thermo-diffusive instabilities occur in premixed systems when the Lewis number is lower than unity (for example for hydrogen). N): 2ruj Yk 2Jk 2rYk j 1 2 1 vk _ 2t 2xj 2xj 3 Then. (8)) is simpli®ed and mass fraction and enthalpy balance equations are formally identical if 2P=2t. In the previous expressions. all ¯uids are assumed to be Newtonian. comparing thermal and mass diffusivities is introduced: 3 2 Sck l Lek 10 Pr rCp Dk Under the assumption of unity Lewis number. To overcome this dif®culty. the previous instantaneous balance equations may be ensemble averaged to derive transport equations for the mean quantity Q: This classical Reynolds averaging technique. especially in premixed ¯ames when species mass fractions and temperature are assumed to be equivalent variables.

and Sckt a turbulent Schmidt number for the species k. HH Y g ² Species ugHH and temperature u HH T turbulent ¯uxes. and the knowledge of steady statistical means is indeed not always suf®cient to describe turbulent combustion. On the contrary. etc. 17). Veynante. however. while the effects of the smaller one are modeled. However. are usually small j j compared to turbulent transport.3. averages. Most experimental techniques determine ~ Reynolds averaged data Q and differences between Q and Q may be signi®cant (Section 7. ¼. etc. see Sections 7. Jh . most combustion works are based on turbulence modeling developed for nonreacting ¯ows. assuming a suf®ciently large turbulence level (large Reynolds numbers limit). energy. Q and ~ Reynolds.5 and 8). with experimental results are not obvious. L. LES is widely studied in the context of non-reactive ¯ows [21± 24]. Accordingly. These equations. k 1. such as: g ² Reynolds stresses u HH u HH : The turbulence model provides i j . Nonetheless. resulting from the rolling up of shear layers. 2.e. that may differ from the instantaneous ones. the complex coupling between micromixing and chemical reactions occurring at unresolved scales needs models. A relation between Q and Q requires the knowledge of density ¯uctuations correlations r H Q H remaining hidden in Favre averaging (see Section 7. The closure may be done directly or by deriving balance equations for these Reynolds stresses. allow only for the determination of mean quantities. such as k±1.3. ² Species chemical reaction rates v k : Turbulent combus_ tion modeling generally focuses on the closure of these mean burning rates. smaller scales feature more universal properties. ² Laminar diffusive ¯uxes Jk . Nevertheless. simply rewritten in terms of Favre averaging. This result is true for any balance equations (momentum. providing ~ Favre averaged quantities Q. theoretical and experimental works have demonstrated that this assumption may be wrong in some premixed turbulent ¯ames and counter-gradient turbulent transport may be observed [19. 2. Filtering and Large Eddy Simulation The objective of Large Eddy Simulation (LES) is to explicitly compute the largest structures of the ¯ow (typically the structures larger than the computational mesh size).198 D.20] (i. in an opposite direction compared to the one predicted by Eq.3.4. As in RANS. 19 The objective of turbulent combustion modeling is to propose closures for the unknown quantities appearing in the averaged balance equations. An alternative is to use large eddy simulation (LES). j k j These ¯uxes are usually closed using a gradient transport hypothesis: HH Y k r ugHH 2 j ~ mt 2Y k Sckt 2xj 20 ² Comparisons between numerical simulations. mass fractions. Favre averaging is only a mathematical formalism: ~ ² There is no simple relation between Favre. are observed in turbulent ¯ames.). estimated from the turbulence model. Strong unsteady mixing effects. turbulence models may be more ef®cient when they have to describe only the smallest structures. Favre averaged balance equations Averaging instantaneous balance equations yields: ² Mass: 2r uj ~ 2r 1 0 2t 2xj ² Momentum (i 1.2. N): 2r ugHH HH Y k 2Jk 2r uj Y k ~ ~ 2r Y k ~ j j 1 2 2 1 vk _ 2t 2xj 2xj 2xj ~ ² Total enthalpy ht : 2r uj ht ~ ~ 2r ht ~ 1 2t 2xj 2r u HH h HH gt 2p 2 h j 2 1 Jj 1 ui tij 1 uj Fj 1 2xj 2t 2xj 18 16 where m t is the turbulent viscosity.4): r Q r Q 1 r HQ H ~ 15 an approximation for this term. LES possesses some attractive properties: ² Large structures in turbulent ¯ows generally depend on the geometry of the system. Q.3. (20). Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 The Favre averaged continuity equation: 2r 2r ui ~ 1 0 2t 2xi 14 is then formally identical to the Reynolds averaged continuity equation for constant density ¯ows.3): 2r u HH u HH gj 2r uj ui ~ ~ 2t ij 2r ui ~ 2p i 1 2 2 1 1 F i 17 2xi 2t 2xj 2xj 2xj ² Chemical species (for N species.4 and Fig. closed with appropriate models. and heat release effects on the Reynolds stresses are generally not explicitly included. its application to combustion modeling is still at an early stage [25].

2. large heat transfers and. the relevant quantities Q are ®ltered in the spectral space (components greater than a given cut-off frequency are suppressed) or in the physical space (weighted averaging in a given volume). Standard ®lters are: ² A cut-off ®lter in the spectral space: ( F k 1 if k # p=D 0 otherwise 22 To summarize: f Q HH ± 0 29 ~ ~ ~ f Q H 0. according 1 Q H : But. ² With LES. dx3 1 21 21 21 25 In combusting ¯ows. the destruction of the device. for example. The derivation of balance equations for the ®ltered quan ~ tities Q or Q requires the exchange of ®ltering and differentiation operators. i 1. with different turbulence characteristics (Section 8. the uncertainties due to this operator exchange are neglected and assumed to be incorporated in subgrid scale modeling. In LES. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 199 ² Turbulent mixing controls most of the global ¯ame properties. instead of being averaged.D. Filtering the instantaneous balance equations leads to equations formally similar to the Reynolds averaged balance equations given in Section 2. where k is the spatial wave number. x2 . L. x3 1=D3 0 if uxi u # D=2. Q Q. Favre ®ltering. Veynante. Q H may be nonzero: À Á Q H x Q xp 2 Q xp F x 2 xp dxp Q xp F x 2 xp dxp 2 Q xp F x 2 xp dxp 27 Q x 2 Q x where Q x ! Q x1 F xp 2 x1 dx1 F x 2 xp dxp Q x1 F xp 2 x1 F x 2 xp dx1 dxp ± Q x 28 where F is the LES ®lter. x2 . are true when a cut-off ®lter in the spectral space is chosen (Eq. 3 otherwise 23 where x1 . ² mass: 24 2r uj ~ 2r 1 0 2t 2xj 30 ~ ~ ~ Q ± Q. the hydrodynamic ¯ow ®eld and acoustic waves. when the ®lter size varies (®lter size corresponding to the mesh size. In LES. This ®lter corresponds to an averaging of the quantity Q over a box of size D . x3 are the spatial coordinates of the location x. ² Most reacting ¯ows exhibit large scale coherent structures [26]. This exchange is theoretically valid only under restrictive assumptions and is wrong. x2 . unsteady large scale mixing (between fresh and burnt gases in premixed ¯ames or between fuel and oxidizer in non-premixed burners) is simulated. which are also especially observed when combustion instabilities occur. This point has been carefully investigated [28]. dx2 . ² A Gaussian ®lter in the physical space: F x F x1 . x2 . is introduced as: r Q x rQ xp F x 2 xp dxp 26 ~ Instantaneous balance equations (Section 2) may be ®ltered to derived balance equations for the ®ltered quantities Q or ~ Q: This derivation should be carefully conducted: Any quantity Q may be decomposed into a ®ltered component Q and a `¯uctuating' component Q H . x3 dx1 . all the frequency components greater than a cut-off wave number kc p=D vanish.LES may be a powerful tool to predict the occurrence of such instabilities [27] and consequently improve passive or active control systems. Q HH 0 The relations used in RANS Q Q. (22)). depending on the spatial location). In most simulations. They need to be avoided because they induce noise variations of the main properties of the system. Q H ± 0. This ®lter preserves the length scales greater than the cut-off length scale 2D: ² A box ®lter in the physical space: ( F x F x1 . in disagreement with classical to: Q Q Reynolds averaging (ensemble average). These instabilities result from the coupling between heat release. a mass-weighted. even in some extreme cases. This may help to describe some properties of the ¯ame/turbulence interaction. Then. large structures are explicitly computed and instantaneous fresh and burnt gases zones. x3 6 pD2 3=2 ! 6 exp 2 2 x2 1 x2 1 x2 1 2 3 D . The ®ltered operation is de®ned by: Q x Q xp F x 2 xp dxp 21 All these ®lters are normalized: 1 1 1 1 1 1 F x1 .7) are clearly identi®ed.3: Q ± Q.

The local imbalance between diffusion of heat and chemical consumption leads to the propagation of the front.200 D. are about 0. part of the information contained in the unresolved scales is lost (and should be modeled). from the Zeldovich/Frank-Kamenetskii (ZFK) theory [14. This progress variable may be de®ned as a reduced temperature or a reduced mass fraction: c T 2 Tu Tb 2 Tu or c u YF 2 YF b u YF 2 YF 35 . Compared to RANS. SL and the kinematic viscosity of the fresh gases. The propagation speed SL of a laminar ¯ame depends on various parameters (fuel and oxidizer compositions. for example. Laminar premixed ¯ames Fig. ¼. 2. Ref. 3. n : Ref dl SL <4 n 34 ~ where Q and Q denote LES ®ltered quantities instead of ensemble means. The unknown quantities are: À Á g ~ ~ ² Unresolved Reynolds stresses ui uj 2 ui uj . the laminar ¯ame speed. This relation. d l. requiring a subgrid scale turbulence model. Compared to direct numerical simulations (DNS). Major properties of premixed. Structure of a laminar plane premixed ¯ame. The scope of this review is limited to RANS. is then almost constant. 2. À Á g ~ ~ ² Unresolved species ¯uxes uj Yk 2 uj Y k and enthalpy À Á g ~ ~ ¯uxes uj ht 2 uj ht : ² Filtered laminar diffusion ¯uxes Jk . non-premixed and partially premixed ¯ames 3.1±1 m/s. such as c 0 in the fresh gases and c 1 in the fully burnt ones.) and is about 0. etc. 3): 2r uj ui ~ ~ 2r ui ~ 1 2t 2xj 2 i 2t ij 2 h 2p r ui uj 2 ui uj 2 g ~ ~ 1 1 Fi 2xj 2xi 2xj 31 ² Chemical species (N species. A strong temperature gradient is observed (typical ratios between burnt and fresh gases temperatures are about 5±7). Either using RANS or LES. N): i 2r Y j ~ 2r uj Y k ~ ~ 2 h g ~ ~ 1 2 r uj Yk 2 uj Y k 1 v k 32 _ 2xj 2t 2xj ² Total enthalpy ht h 1 ui ui =2 i 2p 2r uj ht ~ ~ 2r ht ~ 2 h g ~ ~ 1 2 r uj ht 2 uj ht 1 2t 2xj 2xj 2t 1 2 h Jj 1 ui tij 1 uj Fj 2xj 33 The structure of a laminar premixed ¯ame is displayed in Fig. the basic tools and formalism of turbulent combustion modeling are somehow the same for both techniques.15] is often implicitly used in theoretical derivation of models for premixed turbulent combustion. There is an interesting relation between the thermal ¯ame thickness. LES provides valuable information on the large resolved motions. coupled to subgrid scale models may be numerically solved to simulate the unsteady where the thermal thickness d l corresponds to a temperature jump of 98% of the temperature difference between fresh and fully burnt products. L. fresh gases are preheated and then start to burn. fresh gases temperature. d l. Another characteristic of a premixed ¯ame is its ability to propagate towards the fresh gases. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 behavior of the ®ltered ®elds. Fresh gases (fuel and oxidizer mixed at the molecular level) and burnt gases (combustion products) are separated by a thin reaction zone (typical thermal ¯ame thicknesses. ² momentum (for i 1. For a simple one-step irreversible chemical scheme: reactants 3 products the ¯ame is described using a progress variable c. The ¯ame Reynolds number.1. k 1. Jh : j j ² Filtered chemical reaction rate v k : _ These ®ltered balance equations. combustion occurs at the unresolved scales of the computations. Because of the temperature gradient and the corresponding thermal ¯uxes. Then. Veynante.1±1 mm). 2. Most of the RANS combustion models can be modi®ed and adapted to LES modeling. derived.

introducing the displacement speed v of the iso-c surface: À Á ! _ 2c 1 7´ rD7c 1 v 1 u´7c 37 j7cj wu7cu 2t r u7cu | {z } displacement speed Eq. Tu and Tb are respectively the local. w2 where T. Generic structure of a laminar diffusion ¯ame. YF. The structure of a steady diffusion ¯ame therefore depends on ratios between characteristic times representative of molecular diffusion and chemistry [30]. the unburnt gas u b and the burnt gas temperatures. whereas wc incorporates wrinkling surface effects and may be expressed using Markstein lengths [29]. wr is due to the reaction rate _ v : In a ®rst approximation. 4. YF is non-zero for a rich combustion (fuel in excess). In a counter-¯owing fuel and oxidizer ¯ame (Fig. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 201 progress variable: À Á À Á 2rc 1 7´ ruc 7´ rD7c 1 v _ 2t 36 The previous Eq. Sketch of a counter-¯owing fuel and oxidizer diffusion ¯ame. (36) may be recast in a propagative form.2.D. YF and YF are respectively the local. Increasing the jet velocity. In opposition with premixed ¯ames: Fig. w n 1 wr may be modeled with the laminar ¯ame speed. unburnt gas and burnt gas fuel mass fracb tions. L. A steady planar diffusion ¯ame with determined thickness may be observed in the vicinity of the stagnation point. fuel and oxidizer are on both sides of a reaction zone where the heat is released. The thicknesses of the mixing zone and of the reaction zone vary with these characteristic times. 3 and 4). [2] and references therein. 3. the displacement speed may be split into three contributions: w À Á 1 1 v nn : 7 rD7c 2 D7´n 1 _ rj7cj rj7cj Á 1 2 À 1 rDu7cu 2 D7´n 1 v _ ru7cu 2n ru7cu |{z} | } {z | {z } wn wc wr 38 Fig. The burning rate is controlled by the molecular diffusion of the reactants toward the reaction zone (Figs. (36) then describes the displacement of an iso-c surface with the displacement speed w measured relative to the ¯ow. but are mainly mixing controlled. Introducing the vector n normal to the iso-c surface and pointing towards fresh gases n 27c=u7cu. the amount of heat transported away from the reaction zone is exactly balanced by the heat released by combustion. For a unity Lewis number (same molecular and thermal diffusivities). without heat losses (adiabatic combustion) and compressibility effects. the two de®nitions (35) are equivalent and mass and low Mach number energy balance equations reduce to a single balance equation for the where 2=2n n´7 denotes a normal derivative. 4). The propagation of reactive fronts has been the subject of various developments and more discussion may be found in Ref. 3. wn corresponds to molecular diffusion normal to the iso-c surface. ² Diffusion ¯ames do not bene®t from a self-induced propagation mechanism. Laminar diffusion ¯ames In laminar diffusion ¯ames. wc is related to the curvature 7´n of this surface and corresponds to tangential diffusion. . quenching occurs when the heat ¯uxes leaving the reaction zone are greater than the chemical heat production. SL. Veynante.

ZL Z: In experiments or in simulations involving complex chemistry. n i are the stoichiometric molar coef®cients of the reaction. d YO and d YP are the variations of fuel. oxidizer and product mass fractions. Z evolves through the diffusive layer from zero (oxidizer) to unity (fuel): Y Y f F 2 O 11 YF. the mixture fraction is de®ned from mass fractions of atomic elements [32]. The mixture fraction and these additional conserved scalars are linearly related and one may write: Y O x.o YO. A local orthogonal coordinate system attached to the surface of stoichiometric mixture is introduced and the derivatives in the stoichiometric plane are denoted '. Wi denotes the species molar weight and v is the reaction rate. YO. [33]). The internal structure of diffusion ¯ames is usually discussed using the extent of mixing between fuel and oxidizer.o YO. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 ² The thickness of a diffusion ¯ame is not constant. t TF. T ) (conserved scalars) may be derived by combining the variables (YF. Let us consider the irreversible single step chemical reaction between fuel and oxidizer: F 1 sO 3 1 1 sP where s is the mass stoichiometric coef®cient.o 2 T x. T ). t YO. f is the chemical equivalence ratio: f sYF.o 40 The mixture fraction follows the balance equation: 2rZ 1 7´ ruZ 7´ rD7Z 2t 41 where Lei is the Lewis number of the species i.e. The relations between Z and ZL are given in Table 1.202 D. but depends on the local ¯ow properties. t | {z } Mixing 1 Cp Z x. but an additional mixture fraction: F ZL YF Y 2 O 11 YF. with the mass stoichiometric coef®cient s nO WO =nF WF : The mixture fraction Z is then de®ned by normalizing w using values in the fuel and oxidizer streams.o 1 TO. DF DO D). Similarly. t YF. For unity Lewis number and using Eq.o 2 TO.o is the fuel mass fraction in the fuel feeding stream.o and TF. (41).o F11 44 should be introduced. the diffusion ¯ame is fully determined when the mixture fraction Z and any one of T. T ) and (YO. When LeO LeF . In term of mass fraction.o 1 2 Z x.23). YO. t | {z } Mixing nP dYP nF dYF 1 no dYo where d YF.o Z f11 39 1 nO WO Cp Z x. satisfying [31]: r 2 3 DZ 1 l 7´ 7Z L Dt L Cp 45 where: L LeO 1 1 f= 1 1 F with F Le O =LeF f 46 YF.o YO.o is the oxidizer mass fraction in the oxidizer stream (for instance. [14] or Ref. Mass fractions and temperature balance equations may be reorganized into a new frame where Z is one of the coordinates (see for instance Ref. Using these algebraic relations.o 2 TO. It is ®rst assumed that fuel and oxidizer molecular diffusivities are equal (i. t Q | {z } Combustion 43 where TO.o Z x. in air. Veynante. this chemical reaction may be written: Other Schwab±Zeldovitch variables w (YF. T ) and w (YO.o 2 T x. YO YF 2 YO =s is introduced. YF. L.o 1 TO. a Schwab±Zeldovitch variable) w YF . the species transport equation . or YO is known. The balance equations _ for mass fractions and temperature are necessary to identify the properties of the ¯ame: 2rYF 1 7´ ruY F 7´ rD F 7Y F 2 nF WF v _ 2t 2rYO 1 7´ ruY O 7´ rD O 7Y O 2 nO WO v _ 2t 2 3 2rT l Q _ 7T 1 n F WF v 1 7´ ruT 7´ CP 2t Cp The molecular diffusion is expressed using the Fick law. a conserved scalar (quantity that is not in¯uenced by the chemical reaction.o < 0.o are the temperatures of the fuel and oxidizer streams respectively. the chemical rate of fuel and oxidizer are respectively v F _ nF WF v and v O nO WO v : Q is the amount of heat released _ _ _ by the combustion of an unit mass of fuel. Combining the transport equation for YF and YO. t nF WF Q |{z} Combustion 42 YF x. The conserved scalar approach may still be useful when fuel and oxidizer molecular diffusivities differ. t TF.

the time derivative vanishes and for unity Lewis numbers.of 1 YF. x.o Z f 1 1 1 TO. (47).D. 5. x is the scalar dissipation rate of the mixture fraction Z: 2 3 2Z 2Z Du7Zu 2 xD 48 2xj 2xj measuring the inverse of a diffusive time tx x21 : As this time decreases. summarized in Table 1. The counter¯ow diffusion ¯ame (Fig. When iso-Z surface curvatures are not too strong. When the chemistry is in®nitely fast. and more generally when large velocity gradients are found. species mass fractions and temperature.o Z 1 1 f 2 1=f T TF. ZLst Z ZL 1 1 F= 1 1 f YF 0 YO YO. or solving Eq. ZLst Z f Z L 1 1 F 2 1=F 1 1= 1 1 f YO 0 YF YF. A coordinate j is de®ned across the one-dimensional ¯amelet such as: dZ 2n Z ´7Z u7Zu dj 51 where nZ denotes the normal vector to the iso-Z surfaces. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 203 Table 1 Piecewise relations for in®nitely fast chemistry including non-unity Lewis number. A diffusive time tx < x21 st Du7Zu 2 21 st and a chemical time t c are combined to build a ZZ È Damkohler number Dap tx =tc < tc xst 21 : The response .o Tf 11f Q Cp r 2Yi 22 Y rx 2i 1 v i _ 2t 2Z 49 Neglecting unsteady effects. Zst 1= 1 1 f and ZLst 1= 1 1 F (see Eq. Then. the temperature depends on mixing through Z. the in®nitely fast chemistry hypothesis cannot be invoked everywhere. 5). pointing towards Z 0: The reaction rate may be integrated In many combustion systems. Veynante.34.o 1 TO.o Fuel side Z . the balance equation for the mass fractions reduces to: rDu7Zu 2 2 Yi dZ 2Z 2 52 Assuming that r . but not on the rate of mixing x [35]. Diffusion combustion is limited by two regimes corresponding to pure mixing of the reactants and in®nitely fast chemistry (Fig. The subscript o denotes a quantity measured in pure fuel or oxidizer. the diffusion ¯ame is completely determined as a function of the mixture fraction Z and the scalar dissipation rate x (or 7Z): Y i Yi Z.t) and full solutions for various laminar ¯ames may be derived from asymptotic developments [30.34]. Eq. piecewise relationships exist between Z. Zst and ZL .o 1 2 Z 1 1 f T Tf 2 TO.o 2 Tf Z f 1 1 2 1=f 1 Tf may be written: r 2Yi 22 Y 1 ru' ´7 ' Yi rx 2i 1 7 ' ´ rD7 ' Yi 2t 2Z _ 2 rD7 ' lnu7Zu´7 ' Yi 1 v i 47 in the direction j through the ¯ame. the gradients measured along the stoichiometric surface are smaller than the gradients in the direction Z perpendicular to the stoichiometric surface. Z st and ZL . x An expression for x (Z. using Eq. ZL. D and u7Zu do not vary across the ¯ame (this is typical of a ¯amelet assumption used in turbulent combustion modeling where the ¯ame is assumed very _ thin). Under these hypothesis. The characterization of diffusion ¯ames from the in®nitely fast chemistry situation to the quenching limit is therefore of fundamental interest for turbulent combustion. T T Z. 4) is a generic con®guration well suited to reproduce and to understand the structure and the extinction of laminar diffusion ¯ames. (46)).36]. (43) provides the maximum ¯ame temperature Tf obtained when YF YO 0 and Z Zst 1= 1 1 f TF. mass and heat transfers through the stoichiometric surface are enhanced. (50) leading to Fig. the ¯ame structure is fully described by: rx 22 Yi 1 vi 0 _ 2Z 2 and rx 22 T 1 vT 0 _ 2Z 2 50 showing that the chemical reaction rate is directly related to the function T(Z. x ). Tf is the ¯ame temperature Oxidizer side Z . V i becomes: ! ! 2Y i Z1 2Y Z1 _ V i < 2rDu7Zu 2 rD i 53 2Z Z0 2j Z0 This last relation illustrates how the integrated reaction rate of a species i is directly related to the molecular diffusion ¯ux of that species through the ¯ame. For example in ignition problems or in the vicinity of stabilization zones. L. These extinction phenomena have been theoretically described using asymptotic developments [30. (50): 1 1 1 v x. Z _i _ Vi v i jdj _ dZ u7Zu 21 0 2 1 0 In Eq.

T ) (Fig. oxidizer and temperature lies between the in®nitely fast chemistry limit and the pure mixing case. T(Z) is replaced by the mean temperature calculated for a given value of Z. when the ¯ow is turbulent. extinction develops. L. The value of È the Damkohler Dap at the extinction point may be estimated q by quantifying the leakage of fuel (or oxidizer) through the stoichiometric surface [37]. Starting from a situation where the chemistry is fast. as observed in a . Fig.e. whereas the characteristic thickness of the reaction zone lr depends on both ld and the Damkohler number [12]. 6) [14].e. for a given state in the mixing between fuel and oxidizer. some partial premixing of the reactants may exist before the reaction zone develops. These cases delineate the domain where ¯ames may develop in planes (Z. Two limit cases are thus important for non-premixed turbulent combustion modeling: pure mixing without combustion (Da p 3 0) and in®nitely fast chemistry (Da p 3 1). decreasing Da p (increasing x ) makes the burning rate and transport through the stoichiometric surface greater.3. YO) and (Z. 5). Then. Inner structure of non-premixed ¯ames. Partially premixed ¯ames In non-premixed combustion. until chemistry cannot keep up with the large heat ¯uxes. YF).204 D. Many turbulent combustion models are based on this description of diffusion ¯ame. Veynante. by traveling along the normal to the stoichiometric surface. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. Then. The distribution in mixture fraction space of fuel. 5). The thickness of the diffusive zone ld is estimated from the scalar dissipation rate x at È the stoichiometric surface. 5. (Z. of the burning rate to variations of Da p leads to the so-called `S' curve (Fig. for a given location within a diffusion ¯ame. T(Z) can be constructed and characterizes the combustion regime (i. Moreover. fast or slow chemistry. i. 3. the pure diffusive/reactive layer.

Dap and Dap are the critical values of Dap tc xst 21 at quenching and ignition. oxidizer. while a lean partially premixed ¯ame develops on the air side (Fig. further downstream [39].D. many ¯ames in burners are stabilized by the recirculation of burnt gases. 6. The overall structure. The resulting premixed kernel tends to propagate towards fresh gases and contributes to the stabilization of the trailing diffusion ¯ame. more recent experiments have con®rmed the existence of triple ¯ames in laminar ¯ows [41±43]. The mixtures feeding the reaction zone are then not always pure fuel and pure oxidizer. Theoretical studies [44±48] and numerical simulations [49±53] have been devoted to triple ¯ames. This de¯ection also induces a decrease of the mixture fraction gradient in the trailing diffusion ¯ame. These two premixed ¯ames are curved because their respective propagation velocities decrease when moving away from the stoichiometric condition. Triple ¯ame velocity response to variations of scalar dissipation rate may be derived by approximating the ¯ame tip by a parabolic pro®le and using results from expansions in parabolic-cylinder coordinates. combustion starts in a region where fuel and oxidizer have been mixed in stoichiometric proportion. respectively. leading to stabilization mechanisms controlled by the mixing between fuel. In a mixing layer con®guration. This analysis was used by Ghosal and Vervisch to include small but ®nite heat release and gas expansion. and burnt gases. increasing with the scalar dissipation rate imposed in front of the ¯ame. Generic response of the heat released by a one-dimensional strained diffusion ¯ame versus Damkholer number. where the reactants can be mixed before auto-ignition occurs. a diffusion ¯ame may be stabilized at the splitter plate by the combination of heat losses with viscous ¯ow effects. ² Laminar or turbulent ¯ame stabilization. t cis a given chemical q i time and xst Du7Zu2 st is the scalar dissipation rate under stoichiometric conditions. is usually called `triple ¯ame'. L. or. The propagation speed of triple ¯ames is controlled by two parameters: the curvature of the partially premixed front. may not be the unique relevant model problem. and the amount of heat release. In this latter case. Furthermore. ² After quenching of the reaction zone. The triple ¯ame is an interesting model problem to approach partially premixed combustion. ZZ laminar diffusion ¯ame. Veynante. There are situations where partial premixing is clearly important: ² Auto-ignition in a non-homogeneous distribution of fuel and oxidizer. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 205 È Fig. Since this pioneer work. the reactants may mix leading to possibility of re-ignition and combustion in a partially premixed regime [38]. composed of two premixed ¯ames and of a diffusion ¯ame. 7). making the triple ¯ame speed greater than the propagation speed of a planar stoichiometric ¯ame. when combustion does not start at the very ®rst interface between fuel and oxidizer in the vicinity of burner exit. the triple ¯ame velocity UTF may be written [47]: s b l p 54 xst UTF < SL 1 1 a 2 Zst 1 1 a 4nF 2 2 rCp where a Tburnt 2 Tfresh =Tburnt is de®ned from the . In a laminar shear layer where the mixing between cold fuel and oxidizer develops. the stoichiometric premixed kernel evolves to a rich partially premixed ¯ame in the direction of the fuel stream. The triple ¯ame velocity decreases when increasing the scalar dissipation rate at the ¯ame tip. so that fuel and oxidizer may mix without burning. Such triple ¯ames have been ®rstly experimentally observed by Phillips [40]. The dashed line denotes in®nitely fast chemistry. The effect of heat release is to de¯ect the ¯ow upstream of the triple ¯ame.

Consider a simple irreversible reaction between fuel (F) and oxidizer (O): F 1 sO 3 1 1 sP where the fuel mass reaction rate v F is expressed from the _ Qn T ~ Tn HHn 3 56 where Pn and Qn are given by: Pn n k1 21n2k n 2 1! n 2 k! k 2 1!2 k TA ~ T k . values of x st. illustrates the dif®culties arising from the non-linear character of chemical sources. b a TA =Tburnt is the Zeldovitch number [14]. This simple formalism. and TA is the activation temperature. temperatures on both sides of a stoichiometric premixed ¯ame for the same mixture. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. L. Schematic of a freely propagation triple ¯ame. (54) may be combined with Landau±Squire solution for non-reacting laminar round jet to construct a stability diagram for lift-off and blowout of jet laminar diffusion ¯ames [54]. the averaged reaction rate v F cannot be easily expressed as a function _ ~ ~ of the mean mass fractions Y F and YO the mean density r ~ and the mean temperature T: The ®rst simple idea is to expand the mean reaction rate v F as a Taylor series: _ T exp 2 A ~ T 2 ~ Tb Tb 1 1 3 2 11 T HHn TA Pn n . based on series expansion. As the reaction rate is highly non-linear. n F the stoichiometric coef®cient of the fuel and x st is measured far upstream in the mixing layer where the triple ¯ame propagates. A variety of studies suggest that ®nite rate chemistry and quenching in non-premixed combustion are somehow linked to partially premixed combustion [55]. but non-zero. The value of the scalar dissipation rate at the triple point is of the order of xst = 1 1 a2 [47]. The triple ¯ame velocity given by Eq. 7. A direct analysis: Taylor's expansion A direct approach to describe turbulent combustion is ®rst discussed in this section. 57 Qn b b 1 1¼ b 1 n 2 1 n! .206 D. Veynante. Arrhenius law as: T _ v F 2Ar2 T b YF YO exp 2 A T 55 where A is the pre-exponential constant. exp 2 11 ~ ~ T T n1 11 n1 4. TA is the activation temperature. These relations are valid for small values of a and moderate.

(58) leads to various dif®culties. (58). Because of non-linearities.). First. etc. Turbulent premixed combustion regimes as identi®ed by Borghi and Destriau [62]: (a) ¯amelet (thin wrinkled ¯ame).D. . new quantities HH T HHn such as Yg have to be closed using algebraic expressions k or transport equations. v F becomes [56]: _ 3 2 ~ ~ ~ F 2Ar 2 T b Y F Y O exp 2 TA v _ ~ T 3 2 HH HH HH YgHH YgHH YgHH FY O FT OT 1 1 P1 1 Q1 Â 11 ~ ~ ~ ~ ~ ~ Y FY O Y FT Y OT 5 3 2 HH HH T T YgHH 2 YgHH 2 1 P2 1 Q2 1 P1 Q1 F 2 1 O 2 1 ¼ ~ ~ ~ ~ Y FT Y OT 4 58 Eq. Expression (58) is quite complicated. Fig. 8. Veynante. but is only valid for a simple irreversible reaction and cannot be easily extended to realistic chemical schemes (at least 9 species and 19 reactions for hydrogen combustion. reaction rate closures in turbulent combustion are not based on Eq. For these reasons. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 207 The mean reaction rate. and (c) thickened ¯ame regime. (b) thickened wrinkled ¯ame regime. several hundred species and several thousand reactions for hydrocarbon combustion. L. large errors exist when only few terms of the series expansion are retained.

60± 62]. is then introduced: 2 3 HH Y YgHH Yg F YO F O 59 aFO 2 2 12 ~ ~ ~ ~ Y FY O Y FY O to characterize the mixing between the reactants F and O. If they are perfectly separated Yg 0 and aFO 21: On F YO HH Y the other hand.1. Nevertheless. This time may also be viewed as a diffusive time scale. as long as quenching does not occur. These diagrams could be a support to the selection and development of the relevant combustion model for a given situation. most practical situations correspond to high or medium È values of the Damkohler numbers. the ¯ame front is thin and its inner structure is not affected by turbulence motions which only wrinkle the ¯ame surface. 5. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Models are rather derived from physical analysis as discuss below. The turbulent ¯ow is characterized by a Reynolds number comparing turbulent transport to viscous forces: Re u H lt v 61 mixed by turbulent structures before reaction. a low Damkohler number Da p 1 corresponds to a slow chemical reaction. For turbulent premixed ¯ames. 1 Estimating the turbulent time from turbulent integral scale characteristics È tt lt =u H . This ¯amelet regime or thin wrinkled ¯ame regime (Fig. t c. velocity and time scales describing turbulent ¯ow ®eld and chemical reactions. [59] in a large eddy simulation context). a perfect mixing YgHH 0 leads to aFO F O 0: This segregation factor may be either postulated or provided by a balance equation (see Ref. may be estimated as the ratio of the thickness d l and the propagation speed SL of the laminar ¯ame.2. Then. 5. This transition is described in term of the Karlovitz number Ka: Ka where u H is the velocity rms (related to the square root of the turbulent kinetic energy k). the Kolmogorov scales). The internal structure of the ¯ame is not strongly affected by turbulence and may be described as a laminar ¯ame element called a `¯amelet'. Turbulent combustion involves various lengths. The turbulent structures wrinkle and strain the ¯ame surface.23. Introduction The objective is to analyze premixed turbulent combustion regimes by comparing turbulence and chemical characteristic length and time scales.62] is retained here. È The Damkohler number compares the turbulent (t t) and the chemical (t c) time scales: t Da t 62 tc È In the limit of high Damkohler numbers Da q 1.2. It is worth noting that various chemical time scales may be encountered: fuel oxidation generally corresponds to short chemical time scales Da q 1 whereas pollutant production or destruction such as CO oxidation or NO formation are slower.2. using Eq. the chemical time is short compared to the turbulent one. this approach is used in some simulations of supersonic reacting ¯ows [57] or to describe reaction in atmospheric boundary layer where the temperature T may be roughly assumed to be constant [58]. On the È other hand. a FO. only the ®rst two terms in the series expansion are kept. Scales and diagrams for turbulent combustion 5. L. have a turbulent time t k larger than t c (turbulent motions are too slow to affect the ¯ame structure). In these situations. In this perfectly stirred reactor limit. A formalism combining recent analysis [29. lt is the turbulence integral length scale and n the kinematic viscosity of the ¯ow. The physical analysis is mainly based on comparison between these scales. the Damkohler number becomes: Da tt l S t L tc dl u H 63 where a velocity ratio (u H /SL) and a length scale ratio (lt/d l) are evidenced. the chemical time scale.e. A segregation factor. Combustion regimes È For large values of the Damkohler number Da q 1. the mean reaction rate may be expressed from Arrhenius laws using mean mass fractions and temperature. In turbulent ¯ames. 8a) occurs when the smallest turbulence scales (i. This analysis leads to combustion diagrams where various regimes are presented as function of various dimensionless numbers [14. corresponding to the ®rst term of the Taylor's expansion (58). Reactants and products are tc d u l k tk lk SL 64 1 This chemical time t c corresponds to the time required for the ¯ame to propagate over a distance equal to its own thickness.29. (34) tc dl =SL 1=Ref d2 =n: l .2. Veynante.1. corresponding to a thin reaction zone distorted and convected by the ¯ow ®eld. a physical approach is required to derive models for turbulent combustion. the mean reaction rate becomes: T _ v 2A 1 1 aFO r 2 T b Y F Y O exp 2 A ~ ~ ~ 60 ~ T 5. Turbulent premixed combustion diagram 5.208 D. Introduction _ As the mean burning rate v cannot be found from an averaging of Arrhenius laws.

9). Damkohler. turbulent motions become able to affect the ¯ame inner structure but not necessarily the reaction zone.D. Da. lt/d l. 1: wrinkled ¯ame. For example. the following turbulent premixed ¯ame regimes are proposed [29]: ² Ka . results recovered experimentally [66]. Turbulent premixed combustion diagrams were analyzed using direct numerical simulations of ¯ame/vortex interactions [65]. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 209 The size lk and the velocity uk of Kolmogorov structures are given by [64]: 2 3 31=4 v lk . 9. but cannot modify the reaction zone which remains thin and close to a laminar reaction zone (Fig. In this case. 8b). [65]. L. corresponding to a unity ¯ame Reynolds number Ref (Eq. ² Ka . The laminar propagation is predominant and turbulence/combustion interactions remain limited. the ¯ame thickness d l may be based on the thermal thickness or on the diffusive thickness.2. Comments This analysis. 1: Flamelet regime or thin wrinkled ¯ame regime (Fig. have small lifetimes because of viscous dissipation and therefore only limited effects on combustion. . has a thickness d r much lower that the thermal thickness d l of the ¯ame dr < 0:1dl : The Karlovitz number based on this reaction thickness is: 2 2 2 dr dr dl 1 dl 2 Ka 70 Kar < < 100 lk 100 lk dl lk Then. turbulent motions are able to affect and to thicken the ¯ame preheat zone. for Karlovitz numbers larger than unity Ka . H and a set of two parameters (Re. Results show that the ¯amelet regime seems to extend above the Klimov±Williams criterion (see Fig. and the distributed combustion regime where the ¯ame inner structure is strongly modi®ed by turbulence motions. leading to a rough classi®cation of combustion regimes as a function of characteristic numbers. In this situation. As u H may be viewed as the rotation speed of the larger turbulent motions. This reaction zone. The integral length scale lt may be written: 2 H3 3 u lt 66 1 using v dl SL . Following this classi®cation. such analyses are only qualitative and should be used with great care. 100 (Kar . Ka. Ka # 100 Kar . The Karlovitz number also compares the ¯ame and the Kolmogorov length scales according to: 2 dl Ka 69 lk The Karlovitz number is used to de®ne the Klimov± Williams criterion. previously described. similar to the one presented on Fig. DNS has revealed that small turbulent scales. rather than Ka 1: ² Various effects are not taken into account. Ka) are necessary to discuss regimes in the case of premixed reactants. and Karlovitz. uk v11=4 65 1 where 1 is the dissipation rate of the turbulent kinetic energy k. 1: wrinkled ¯ame with pockets (`corrugated ¯ames'). ² 1 . which is not the case in combustion systems. Unsteady and curvature effects play an important role neglected here. Da). has been developed as a support to derive and choose turbulent combustion models. 5. A recent analysis [29] has shown that. Two subdivisions may be proposed depending on the velocity ratio u H /SL: Ð(u H /SL) . most practical applications correspond to ¯amelet or thickened wrinkled ¯ame regimes. 1. Peters [29] shows that the criterion Ka 100 (i.e. lt/d l). Kar 1 is in quite good agreement with the transition proposed in Ref. yields 2 H 33=2 u lt 21=2 67 Ka SL dl È Reynolds. which are supposed in classical theories to have the strongest effects on ¯ames. 1. 8c). These various regimes are generally displayed on a logarithmic diagram (u H /SL. at least when the length scale ratio. 1 : Thickened wrinkled ¯ame regime or thin reaction zone. Accordingly. This criterion was ®rst interpreted as the transition between the ¯amelet regime Ka . ² Some quantities used are not clearly de®ned. (34)). Nevertheless. (Re. is suf®ciently large. ² All regime limits are based on order of magnitude estimations and not on precise derivations. In this situation. For example. the limits between the various regimes may noticeably change. 8a). A diagram such as the one displayed on Fig. turbulent structures are unable to wrinkle the ¯ame surface up to ¯ame front interactions. the ¯amelet regime limit could correspond to a Karlovitz number Ka 0:1 or Ka 10. Veynante. Ka) or (Da. larger structures become able to induce ¯ame front interactions leading to pockets. 1): Thickened ¯ame regime or wellstirred reactor. 9 cannot be readily used to determine the combustion regime of a practical system from (u H /SL) and (d l/lt) ratios: ² The analysis is based on the assumption of a homogeneous and isotropic turbulence unaffected by heat release.3. Re. corresponding to Ka 1. preheat and reaction zones are strongly affected by turbulent motions and no laminar ¯ame structure may be identi®ed (Fig. numbers are related as: Re Da2 Ka2 68 Ð(u /SL) . where heat is released. delineating between two combustion regimes.

A laminar diffusion ¯ame is fully determined from a È Damkohler number Dap tc xst 21 . Introduction Two numbers. Turbulent premixed combustion diagram [29. where various characteristic scales have been retained depending on the authors [33. the ¯ame is thinner than any turbulent scale. For example. 5. On the other hand. the CO2 formation from carbon monoxide (CO) and OH radical in the burnt gases is much slower with a chemical time of the same order as turbulent times. The Klimov±Williams criterion Ka 1 corresponds to a ¯ame thickness d l equal to the Kolmogorov scale lk. the fresh and the burnt gases temperature (see Section 8). d r. [65] from direct numerical simulations is also displayed. therefore.67±71]. is thinner than any turbulent scale and is not affected by turbulent motions (the criterion is plotted assuming dr < 0:1dl ).2) and no ®xed reference length scale can be easily identi®ed for diffusion ¯ames. to characterize the size of the smaller vortex able to affect the ¯ame front was used [63]. a length ratio and a velocity ratio. [140] to delineate between gradient (above) and counter-gradient (below) turbulent transport is displayed assuming a heat release factor t Tb =Tu 2 1 6 where Tu and Tb are. Below the line delineating the Peters criterion (Ka 100 or Kar 1). Additional lengths have also been introduced. For instance the Gibson scale lG. 9. whereas a Damkohler number is chosen for the reaction zone [72].62]. L. respectively. ² A velocity ratio (turbulence intensity to premixed laminar ¯ame speed) and a length ratio (integral scale to premixed laminar ¯ame thickness) may be constructed [68] to delineate between regimes. Combustion regimes are identi®ed using the length scale lt =dl and the velocity u H =SL ratios. The criterion proposed in Ref. the reaction zone thickness. the thickness of the ¯ame depends on the aero- dynamics controlling the thickness of the local mixing layers developing between fuel and oxidizer (Section 3.3. Non-premixed turbulent combustion diagram 5. This dif®culty is well illustrated in the literature. Below this line. using for instance thicknesses of pro®les in mixture fraction space [67]. In actual turbulent combustion. This length was de®ned as the size of the vortex having the same velocity as the laminar ¯ame speed SL. where the value of the chemical time t c depends on the fuel chemistry [30] . The ¯amelet regime limit devised by Poinsot et al. In addition. ² All these analyses are implicitly based on a single step irreversible reaction. ² Additional length scales have been introduced in the literature. These reactions may correspond to a large range of chemical time scales.1.3. These classi®cations of non-premixed turbulent ¯ames may be organized in three major groups: ² The turbulent ¯ow regime is characterized by a Reynolds È number. exhibit no intrinsic characteristic speed. a large number of chemical species and reactions are involved (several hundred species and several thousand reactions for propane burning in air). Veynante.210 D. ² The mixture fraction ®eld is retained to describe the f È turbulent mixing using Z HH2 and a Damkohler number (ratio of Kolmogorov to chemical time) characterizes the ¯ame [33]. have been used to identify premixed turbulent combustion regimes. The problem is more dif®cult in non-premixed turbulent combustion because diffusion ¯ames do not propagate and. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. the propane oxidation may assumed to be fast compared to turbulent time scale.

² For ®nite rate chemistry. DaLFA). where a is the order of a global one-step reaction. 1 3 v i _ 71 Summarizing these effects in a generic diagram is an arduous task. t is a chemical time and r the characteristic size of the vortices. In this diagram presented on Fig. however. when Dap # Daext . Veynante. In this number. a characteristic mixing length.2). 11): 2 3=2 3 ~ 21 < lt < k 72 lz < u7Zu 1 Turbulent small scale mixing mainly depends both on velocity ¯uctuations. transporting the iso-Z surfaces (stirring). DaLFA and Daext. ld D=xst 1=2 : According to asymptotic developments [30]. the reaction zone thickness is of the order of lr < ld Dap 21= a11 . a departure from laminar ¯amelet is expected. u H is the level of velocity ¯uctuations. A diagram for laminar ¯ames submitted to curvature associated to a time varying strain rate was obtained by Cuenot and Poinsot from DNS results of ¯ame/vortex interaction [70]. This diagram delineates the steady laminar ¯amelet assumption (LFA) validity regions. the ¯ame thickness is d i < ld. the burning rate does not immediately follow variations of x st. strong unsteadiness effects are observed. On the other hand. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 211 Fig. Da p . xst Du7Zu st . a main dif®culty arises when effects of unsteadiness need to be quanti®ed. Also. leading to a distribution of x st for given rates of strain. This È analysis identi®es two limiting Damkohler numbers. Because diffusion ¯ames do not feature a ®xed reference length. 10. di ld is the ¯ame thickness (ld < u7Zu 21 ). Daext . and diffusion between these iso-surfaces that compose the mixing layer of thickness ld. ¯ame extinction occurs. the reaction zones develop within a mean mixing zone whose thickness lz is of the order of the turbulent integral length scale lt (Fig. st and. modifying the burning rate (Eq. with D 1=2 ld < 73 ~ x st where x st denotes the conditional mean value of the scalar ~ dissipation rate x for Z Zst : When transport of species and heat by velocity ¯uctuations is faster than transfer in the diffusion ¯ame. whereas r and u H denote respectively the characteristic size and velocity of the vortex pair.D.e. the inner . unsteadiness in diffusion ¯ames develops at two levels [73]: ² The mixture fraction ®eld Z does not immediately respond to velocity ¯uctuations. In a non-premixed turbulent ¯ame. when the velocity ®eld ¯uctuates. (Section 3. Because a strong correlation exists between x st and velocity gradients taken along the stoichiometric line [74]. leading to a second level of unsteadiness. 10. u H 3 Unsteadiness in mixing 3 xs 3 Burning rate unsteadiness for Dap . In a steady laminar ¯ame the local rate of strain is directly related to x st (and to a ¯ame thickness). the ¯ame front may be viewed as a steady laminar ¯ame element and its inner structure is not affected by vortices. L. In the È intermediate Damkohler number range (i. this effect is not the dominant one when ®nite rate chemistry occurs. the scalar dissipation rate 2 under stoichiometric condition Z Zst . when the Kolmogorov scale lk is of the order of the ¯ame thickness. When Da p is larger than DaLFA. tx x21 . measures at the same time a mechanical time. the quenching limits and the zone where unsteady and curvature effects are important during ¯ame vortex interaction. (71)). Non-premixed ¯ame/vortex interaction regimes by Cuenot and Poinsot [70].

when Dap # Daext ). When the Damkohler number Da tt =tc .212 D. lr the reaction zone thickness. 12. The basic ingredients to describe turbulent ¯ames remain the quantities introduced for laminar ¯ame analysis: the progress variable c for premixed combustion (c 0 in fresh gases and c 1 in burnt gases. structure of the reaction zone may be modi®ed by the turbulence. or. described in .1. The ¯ame position would correspond to values of the progress variable c lying between 0 and 1.e. In a practical combustion devices. the reaction rate is limited by turbulent mixing. This condition may be simply expressed as Dap $ DaLFA On the other hand. 13: The burning rate may be quanti®ed in terms of turbulent È mixing. the ¯ame is expected to have a laminar ¯ame structure. or even quenching. one may write: a ~ ld < a1 lk and x st < 2 74 tk where a1 $ 1 and a2 # 1 (the maximum local strain rate would correspond to ld pk ). It is then concluded that the non-linear character of the problem requires the introduction of new tools. see Section 3. a 1 and a 2 would evolve in space and time according to ¯ow ¯uctuations. (58)). As the classi®cation of premixed turbulent ¯ames. In a given burner. 6. it is likely that one may observe at different locations ¯amelet behavior and strong unsteadiness. Tools for turbulent combustion modeling 6.2). the Damkohler number comparing turbulent ¯ame scale and chemical ¯ame scale is recast as: p t t t t a2 Da t t k < t < a2 Re Dap 75 ~ tc tk tc tk x st tc È Constant Damkohler numbers Da p correspond to lines of slope 1/2 in a log±log (Da. 12. to Z taking on values in the vicinity of Z Zst : Three main types of approaches are summarized in Fig. Veynante. The simplest and most direct approach is to develop the chemical rate in Taylor series as a function of species mass fractions and temperature (Eq. for large chemical times (i. l È Then using tt =tk Re. Sketch of a non-premixed turbulent ¯ame. Z is the mixture fraction. velocity and scalar energetic spectra. Re) diagram. 11. These new tools must be designed to describe turbulent ¯ames and have to provide an estimation of mean production or consumption rates of chemical species. Schematic of non-premixed turbulent combustion regimes È as function of the Damkohler number Da tt =tc (constructed from the turbulent integral time scale t t and chemical time t c) and Re the turbulent Reynolds number. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. L. practical interest that should be approximated by turbulent combustion models. comparing turbulent (t t) and chemical (t c) characteristic times. for example) or on quantities that may be easily modeled or obtained from closed balance equations. extinction occurs. This analysis is limited by its low accuracy and by the rapidly growing complexity of the chemistry (Section 4). a generic description of the main concepts used to model turbulent combustion is proposed. is large (a common assumption in combustion modeling). When the chemistry is suf®ciently fast (large Da values). and. Laminar ¯ames are encountered for low Reynolds numbers Re . the mixture fraction Z for non-premixed ¯ames (Z is a passive scalar. Introduction The mean heat release rate is one of the main quantities of Fig. these considerations are limited by the numerous hypothesis necessary to derive the regimes.2). ld the diffusive thickness. In this section. Relations between the various approaches are also emphasized. They also need to be based on known quantities (mean ¯ow characteristics. but the discussion of the closure strategy is postponed to subsequent sections. lt the turbulence integral length scale. As diffusion ¯ame scales strongly depend on the local ¯ow motions. see Section 3. with Z 0 in pure oxidizer and Z 1 in pure fuel. 1: Results are summarized in Fig.

The density of ¯ame surface area per unit volume is also useful to estimate the burning rate. accordingly. Following this view. In the geometrical analysis. Three types of analysis of premixed or non-premixed turbulent ¯ames. plays a dominant role in combustion modeling. terms of scalar dissipation rates [75]. L. A ¯ame normal analysis is derived by focusing the attention on the structure of the reacting ¯ow along the normal to the ¯ame surface. even for ®nite rate chemistry. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 213 Fig. The ¯ame is then envisioned as an interface between fuel and oxidizer (non-premixed) or between fresh and burnt gases (premixed). . The small scale dissipation rate of species controls the mixing of the reactants and. 13. the ¯ame is described as a geometrical surface. this approach is usually linked to a ¯amelet assumption (the ¯ame is thin compared to all ¯ow scales).D. scalar ®elds (c or Z) are studied in terms of dynamics and physical properties of iso- value surfaces de®ned as ¯ame surfaces (iso-c p or iso-Zst). This leads to ¯amelet modeling when this structure is compared to one-dimensional laminar ¯ames. Veynante.

the f variance c HH2 is de®ned as: ~ ~ f r c HH2 r c 2 c2 r c HH2 2 c2 rc2 2 r c2 f ~ 76 gradient). r x is the dissipation rate of the ¯uctuations of the ~ scalar ®eld. The determination of these pdfs leads to pdf modeling. these ¯uctuations characterize non-homogeneities and intermittencies. viewed as an interface between fresh and burnt gases in premixed combustion or as an interface between fuel and oxidizer in non-premixed situations. Geometrical description The ¯ame front is here described as a geometrical entity. One may then expect a very strong coupling between mixing phenomena and g chemical reaction. Since the burning rate depends on the contact between the reactants. 6. when mean gradients are neglected: ~ r x < rD7c HH ´7c HH 78 Then. This preliminary analysis shows that dissipation rate of scalars is a very key concept of turbulent combustion and. Mean values and correlations are then extracted via the knowledge of one-point probability density functions (pdf). A one-point statistical analysis restricted to a particular value of the scalar ®eld is related to the study of conditional statistics. the mean burning rate is proportional to the scalar dissipation rate of Z or c. 6. This analysis is generally linked to the assumption of a suf®ciently thin ¯ame. the exact transport equation f for c HH2 reads: 2r c HH2 f 1 7 r uc HH 2 1 7 r ug 2 ~g HH c HH 2t ~ 7´ rD7c HH 2 1 2c HH 7´ rD7c 2v c HH _ 22r u HH c HH ´7c 2 2rD7c HH ´7c HH 1 |{z} g ~ |{z} |{z} Production Dissipation Source 77 In addition to the two diffusive terms 7´ rD7c HH 2 and ~ 2c HH 7´ rD7c. turbulent mixing occurs between fresh and burnt gases. x appears in any tools used to model ¯ames. G-®eld equation The balance equation for the progress variable may Starting from the balance equation for the progress vari~ able (Eq. then the new equation is multiplied by c HH and averaged. Scalar dissipation rate f f In a ®rst step.214 D. fresh fuel and fresh oxidizer have to be mixed at the molecular level for reacting and the ¯ame is mainly controlled by turbulent mixing occurring between the fresh gases. Two formalisms have been proposed: ®eld equation or ¯ame surface density concept.1. In the literature. (36)). Hereafter: ~ ~ ~ ~ r x rD7c´7c rD7c´7c 1 2rD7c HH ´7c 1 rD7c HH ´7c HH leading to. This is observed in the equation for c HH 2 ~ where. In consequence. After straightforward manipulations. L. when assuming very fast chemistry and a combustion limited by mixing. the transport equation for c HH2 or Z HH2 is derived. For instance. In a non-premixed ¯ame. x and the chemical source v c HH are _ involved.2. In the case of the progress variable. Within a premixed system. various expressions have been associated to the terminology scalar dissipation rate (in laminar ¯ame theory. ~ ~ In the simpli®ed case of homogeneous ¯ames (no c or Z . which are non-zero. the scalar dissipation rate enters directly or indirectly the expression for the mean burning rate. in any models. directly or indirectly. c is decomposed into c c 1 c HH . Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 The statistical properties of scalar ®elds may be collected and analyzed for any location within the ¯ow. there f is no chemical source acting on the evolution of Z HH2 : The mixture fraction Z is sensitive to chemistry only via density change. Veynante.3. Conditional statistics which are obviously linked to the geometrical analysis and to ¯ame surfaces when the conditioning value is c p or Zst. The main stumbling block in turbulent combustion modeling and bridges between the various modeling concepts emerge through the scalar dissipation rate. making the coupling between chemistry and mixing different than in the case of premixed combustion. 6. it actually quanti®es a diffusion speed. a factor 2 and be written in term of instantaneous (c) or ¯uctuating (c HH ) values of the concentration species. Section 3.3. It may include the density r . the time evolution of the scalar variances are governed by: Premixed combustion: 2r c HH 2 g _ 22rD7c HH ´7c HH 1 2v c HH dt Non-premixed combustion: 2r Z HH 2 g 22rD7Z HH ´7Z HH dt These equations have important implications: The scalar dissipation rate directly measures the decaying speed of ¯uctuations via turbulent micromixing. two important terms are found: The ¯uctuating part of the scalar dissipation rate 2rD7c HH ´7c HH and a correlation v c HH _ involving the chemical source. but expected small for large Reynolds number ¯ows (especially the second one).2). at the same time.

Each G iso-surface is then related to a ¯ame front. Averaging the progress variable Eq. pointing toward the fresh gases. was introduced to simulate the propagation of fronts [76]. The reactant consumption and the heat release rate are controlled by the consumption speed Sc: 1 ru Sc v dj _ 82 21 where Tu is the temperature of unburnt gases. a model has to be where r u is the fresh gases density and j the spatial coordinate along the normal direction to the ¯ame front. especially in high ¯ame front curvature zones [78]. This velocity is easily decomposed into the ¯ow velocity u and w the relative velocity of the iso-c surface measured with respect to the ¯ow. The `G-equation' may be written: 2G 2G 1 uc ´7G 0 or 1 u´7G wu7Gu 2t 2t 80 w and Sc are related but may be quite different. generally much thicker than the ¯ame front. has to be modeled. (79) leads to: P Q ~ 2c ru R 27´ r u HH c HH 1 rwu7cu S g ~ ~ ~ u7cu 84 1 u´7c r 2t ~ u7cu |{z} ST where the turbulent ¯ame speed. The G-equation becomes: ~ 2G ru ~ ~ ~ 1 u´7G S u7Gu 2t r T The mean turbulent ¯ame brush is then located at the ~ points where G G 0 : The G-equation does not required thin ¯ame elements per se. Usually. This formalism may be extended to turbulent ¯ames. even in the case of a steady laminar ¯ame propagating at the constant laminar ¯ame speed SL. a G-®eld whose level G G 0 represents the ¯ame surface. ST.D. this technique is well suited to two dimensional simulations. The G-equation provides a kinematic description of the ¯ame front and involves its displacement speed w. leading to u c u 1 wn. de®ned relatively to the unburnt gases. The displacement speed w may also be quite different from the laminar ¯ame speed SL. Three approaches have been proposed to overcome this dif®culty: ² Flame front tracking technique: The displacement of the ¯ame front is evaluated from the displacement speed w. the displacement speed w of the G-iso-surface is affected by thermal expansion and should be corrected for density variations. In laminar ¯ames. In non-constant density ¯ows where thermal heat release is included. Only the G-®eld. L. any other effects (heat losses. . A more dif®cult point is the coupling required between the G-equation and the species or energy balance equations. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 215 be written: 2c 1 1 u´7c 7´ rD7c 1 v wu7cu _ 2t r 2c or 1 u c ´7c 0 2t 79 where uc is the absolute propagation velocity of the progress variable ®eld. Q the reaction heat release and H denotes the Heaviside function (the Heaviside function is smoothed on the mesh of the simulation). ² Under the assumption of constant density (thermodiffusive assumption). Based on a purely geometric approach. (37). needs to be resolved. but its extension to 3D cases may not be straightforward. Of course. the G-equation may be used for low cost direct numerical simulations. ² Temperature (or energy) reconstruction: The temperature ®eld is directly estimated from the G-®eld as [81]: T Tu 1 Q H G 2 Gp Cp 83 This kinematic description of premixed combustion possesses some attractive aspects: ² The internal ¯ame front structure does not need to be resolved on the computational mesh. where n 2 7c=u7cu is the local normal vector to the iso-c surface.80]. Accordingly. The overall turbulent ¯ame is only viewed as a propagating surface without solving for the internal ¯ame structure. This formalism is therefore a good candidate for the numerical simulation of large systems where the knowledge of the internal structure of the ¯ame brush is not required [79]. This correction is: r w u wu 81 r where wu is the ¯ame displacement speed relative to the unburnt gases of density r u. The coupling between the consumption speed Sc and the displacement speed w is a very key point in G-®eld modeling. Nevertheless. and one may argue that a single simulation corresponds to the computation of several ¯ames. compressibility) may be included. leading to an estimation of the volume of burnt gases produced along with the thermal heat release [79. This approach does not require a balance equation for the energy but is not applicable when heat losses or compressibility effects (for example in an internal combustion engines) occur. The relative displacement speed w is given by Eq. ² Estimation of the heat release rate from the G-®eld: The G-®eld is used to estimate the heat release rate to be incorporated in the balance energy equation from a formulation similar to Eq. (83) [82]. Veynante. G x is de®ned as the distance of the given location x to the ¯ame front [77].

a laminar ¯ame where curvature effects have been introduced. x. submitted or not to a steady strain. the ¯ame surface density S of the iso-c p surface is estimated from the conditional gradient of the progress variable c [86]: S c p u7cud c 2 c p u7cuuc c p P cp 86 where d c 2 cp is a local measure of the probability (see Section 6. Ç V i is related to the properties of the local ¯ame front and is generally estimated from a prototype laminar ¯ame. Two equivalent forms of the progress variable equation may be used: ² A classical reaction/diffusion formulation: 2c 1 _ 1 u´7c 7´ rD7c 1 v r 2t 87 Ç where V i is the mean local burning rate per unit of ¯ame area integrated along the normal direction to the ¯ame surface. is to decouple the chemical description (V i) from the ¯ame/turbulence interaction (S ). . L.33]. 2. In premixed combustion. supplied for ST. incorporating more or less complexity. More rigorous derivations were obtained from geometrical considerations [17. Derivation of an equation for u7cu from the equation for the instantaneous progress variable c. details are not given here. turbulent transport and propagation may be separated to carefully model the effect of turbulence on w. 14. Averaging the u7cu d c 2 c p balance equation leading to an exact equation for the ¯ame surface density S u7cud c 2 c p : This derivation is valid for any iso-scalar surface (c p can take any values between zero and unity) and S c p . 3.4. S measures the available ¯ame area d A per unit volume d V. u7cuuc c p is the conditional average of u7cu for c c p and P cp is the probability to ®nd c cp at the given location. From this de®nition and the balance equation for the progress variable c. an exact equation for the ¯ame surface density S may be derived according to the following steps [87]: 1. Using a phenomenological analysis. 14. or even a laminar unsteady strained and curved ¯ame. curved and strained by the velocity ®eld [17.3. For instance.3). diffused. The ¯ame surface density S may be estimated either from algebraic relations (see Section 7.4. when doing so. the derivation of its balance equation is quite tedious [87] and similar to the derivation of a balance equation for the probability density function (Section 6.2. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. summarized Ç on Fig. the instantaneous displacement speed of the premixed front [29. t is also called a surface density function. Flame surface density description The ¯ame is identi®ed as a surface and the ¯ame surface density S is introduced. The ¯ame surface is convected.4. Flame surface density modeling. Veynante.85] and from a statistical description [86±88] leading to an exact. The mean burning rate of a species i is then modeled as: _ _ v i V iS 85 balance equation. Derivation of an equation for u7cu d c 2 c p by conditioning the previous u7cu balance equation. one may consider a planar laminar ¯ame. The derivation using statistical tools is now brie¯y summarized.216 D. but unclosed. this balance equation was ®rst proposed by Marble and Broadwell [84] for non-premixed turbulent combustion. 6.3).83]. The main advantage of this formulation.2) or as a solution of a balance equation.

These two formulations of the S balance equation induce the following comments: ² The two balance Eqs. The last two terms in Eq. Veynante. (89) and (90) correspond to unsteady effects and to the ¯ame surface convection. many effects are incorporated in the ¯ame front propagation speed. Integrating Eq. 2/2c p is a derivative in the sample space c p. a ¯ux term in the sample space is found in the reaction/diffusion derivation. Eq.1). the local burning zone is not in®nitely thin and the ¯ame front. In the propagative form (Eq. identi®ed as the location of the maximum reaction. (37)). ² The derivation of the balance Eqs. This is a priori true when the ¯ame is in®nitely thin. (86) across iso-surface levels leads to: 1 1 S c p dcp u7cuuc c p P cp dcp u7cu 95 0 0 . The last two terms in the RHS of the propagative Eq. ² In combustion modeling. and kQls is a function of c p. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 217 ² A propagative form: 2c 1 u´7c wu7cu 2t 88 where the displacement speed w of the ¯ame front relatively to the ¯ow was introduced (Eq. describes whether the ¯ame is actually burning or not. (89) and (90) implicitly forgets some mathematical singularities that may become important in particular situations. w. The propagative formulation is often written in term of p total stretch Ac of the iso-c p surface 1 2S 1 7´ ku 1 wnl s S A c p S 2t z}|{ k | nn : 7u 1 w7´n l s 7´u 2 {z } tangential strain rate Curvature ² The strain rate term is also split into a contribution due to the mean ¯ow and a contribution due to turbulent velocity ¯uctuations: ~ ~ k7´u 2 nn : 7ul s S 7´u 2 knnl s : 7ul s S |{z} AT 92 ~ 1 k7´u HH 2 nn : 7uS | {z } aT 94 The LHS terms in the two balance Eqs. This equation corresponds to the G-equation [77] (Section 6. This may be for instance the case when analyzing data of direct numerical simulations where the ¯ame front has to be resolved on the computational grid. This may not be the case when two close ¯ame fronts interacts. (90)). In Eq. The surface average of Q. the normal vector n is assumed to be well de®ned and having a ®nite derivative. For example. is de®ned as: Qu7cujc c p Qu7cud c 2 c p 91 kQl s u7cud c 2 c p u7cujc c p Here the notation kQls implicitly indicates that the mean is taken for c cp . Two different types of equation for S are associated to these two formulations of the progress variable balance equation. assuming that the Ç local reaction rate per unit ¯ame area. V i. However in real turbulent ¯ames. L.3. ~ ² Using the Favre decomposition u u 1 u HH . The ®rst term in the RHS expresses the action of the tangential strain rate on the ¯ame surface. the convection ¯ux term may be decomposed into mean and turbulent components: ~ kul s S uS 1 ku HH l s S 93 89 Â Ã k7´u 2 nn : 7ul s S 2 7´ kwnl s S 1 kw7´nl s S where n is the unit vector normal to the c c p surface and pointing towards the fresh gases n 27c=u7cu: 7´n corresponds to the ¯ame front curvature. that may differ from the laminar ¯ame speed SL [78]. (89) describe respectively reaction/diffusion effects All these de®nitions may be extended to ¯ame fronts which are not in®nitely thin. (89) and (90) are mathematically strictly equivalent but the problem is not expressed in the same way. a single iso-c p surface is assumed to correspond to the ¯ame front. (90) the imbalance between diffusion and reaction is cast in the form of the propagation velocity w. (89) recalls that transfer phenomena between iso-surfaces are involved in this propagation. may differ from the c p iso-surface. kQls.D. ² A reaction/diffusion formulation: 2S 1 7´ kul s S k7´u 2 nn : 7ul s S 2t B 2 3C 1 2 1 2 7´ rD7c 1 v _ S u7cu 2n r s C 5 4B 2 1 2 7´ rD7c 1 v S _ 2c p r s ² A propagative formulation: 2S 1 7´ kul s S 2t 90 along the ¯ame normal direction and ¯uxes in the sample space c p. On the other hand. 2=2n n´7 is a derivative normal to the ¯ame front. (90) correspond to front convection due to a normal propagation and combined propagation/curvature effects.

One needs to develop closures for the turbulent ¯ux of ¯ame surface. tDY [92±96].218 D. t quanti®es the probability to ®nd. Consider a non-premixed ¯ame where the chemistry is reduced to a single step reaction. This approach has been explored by Weller et al. but the wrinkling factor may be more convenient for initial and/or boundary conditions (J $ 1 everywhere). for example. t where n and np are the unit vectors normal to the instantaneous ¯ame front and to the mean propagating direction respectively. velocity. the ¯ame surface is assumed to behave such as S c p < u7cu S: All balance equations remain formally the same but have to be closed. n p 2 u7cu 7cuc c p Then: u7cuuc c p S c p J c p 7cuc c p 7cuc c p P cp where Y p is the sample space variable corresponding to the random variable Y. u7cu follows a balance equation similar to the S (c p) balance Eq. The pdf satis®es the following simple relations: P Y p . or pollutants like CO.3. (90). and where radiative heat losses are neglected.4. a generalized ¯ame surface wrinkling. In probability density function methods. for any values of c p. require the description of the ¯ame front internal structure. L. [90.3. x. x. a joint probability density p p function P Y1 . x.3.The conditional pdf Pc YF uZ p . etc. Statistical approaches: probability density function 6. The probability density function (pdf) P Y p . x. t: For such ¯ames. x. they are initially based on a geometrical view describing the ¯ame as a thin interface. ¼. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 that is a direct estimation of the inverse of the mean local ¯ame thickness d d 1=u7cu 21 and may be viewed as a `generalized ¯ame surface density'. The basic idea is to introduce the ratio J (c p) of the ¯ame surface to its projection in the direction of propagation np: u7cuuc c p P cp S c p p J c np ´ nu7cud c 2 c p np ´ nu7cud c 2 c p u7cuuc c p 1 np ´knls np ´ nu7cuuc c p 97 closely related. When more than one variable is required to capture the ¯ame structure. temperature. 6.3. ¼. t is introduced As in Section 6. J . Veynante. Flame wrinkling description The previous formalism may be recast in term of ¯ame surface wrinkling. The turbulent ¯ame is then fully described by the joint pdf of mixture fraction and fuel mass fraction. J c p J for any c p value. x. where surface averages kQl s are replaced by `generalized surface averages' kQls which do not depend on c p: 1 1 kQls kQl s S c p dcp 96 u7cu 0 In ¯ame surface density models. tdY p 1 P Y p . tdY p Y x. YN . under a ¯amelet assumption. is introduced: 1 S c p dcp u7cu S 0 J 100 1 u7cu u7cu 7cuc c p P cp dcp 0 For an in®nitely thin ¯ame front. 5). ¼. Flame surface density and ¯ame wrinkling factor are . t is introduced [97]. The mean burning rate (or any mean quantity) is then estimated as: v Y1 x. 6. x.4. for a given location x and a time t. Introduction Predictions of radicals and intermediate species such as OH. Balance equations may also be derived and closed for J (c p) or J : These equations are considerably more complicated than ¯ame surface density balance equations. x. a variable Y (mass fraction. these conditional means are also useful in a pdf context. as well as for the effects of curvature and strain rate. Y p 1 DY=2: This probability is equal to P Y p . tdY p Y H2 x. Even though G-®eld and density of ¯ame surface S need some statistical treatments. one wishes to relax this assumption by focussing on the statistical properties of intermediate states within the ¯ame front. t _ Y1 ¼ YN p p p p p p v Y1 Y1 . it is interesting to focus on the statistical properties of the fuel mass fraction YF for a given value of the mixture fraction Z (Section 3.2. p P YF . YN . These vectors are chosen pointing towards the fresh gases and are given by: 7cuc c p 7c n2 98 . The same modeling issues emerge using the Gequation formalism [89]. Z p .91]. YN P Y1 . for the propagation velocity of the surface. for intermediate states between fresh and burnt gases in premixed ¯ames or between fuel and oxidizer in non-premixed ¯ames.2 and p Fig. Laminar combustion would be parameterized with two variables.2). tdY1 ¼dYN _ 101 99 Conditional statistics have been used to de®ned the ¯ame surface density (Section 6. fuel mass fraction YF and mixture fraction Z (see Section 3.2).) within the range Y p 2 DY=2. t Y Y Y Y p 2 Y2 P Y p . x.1.

D. usually with a Beta function [72]: À Áb21 À Á Z p a21 1 2 Z p ~ P Z p . t d cp 2 c x.3). tP Z p . x. t 0 By de®nition: otherwise In this decomposition. P Z p . Because the probability to ®nd all the possible values is constant and equal to unity: 1 P cp . x. Z HH2 vanishes when reactants are perfectly mixed and reaches its maxiÀ Á ~ ~ mum value. 0 f Z HH2 1 À 0 Á À Á ~ 2~ Z p 2 Z P Z p . t describes the statistical p properties of fuel/air mixing. t dZ p 0 a a 1 1¼ a 1 n 2 1 a 1 b a 1 b 1 1¼ a 1 b 1 n 2 1 where C can either be time or any spatial coordinates. t satis®es: 2 2 2c x. t Â (105) 2C This presumed pdf should reproduce the mean of the f ~ mixture fraction Z and its variance Z HH2 : 1 À Á ~ ~ Z Z p P Z p . This ¯ux is driven by a velocity equal to the conditional mean of the time evolution of the progress variable c. t cp P cp . t is given by: 2c 1 _ 2u´7c 1 7´ rD7c x. x. when fuel and oxidizer are completely segregated (Section 5. whereas Pc YF uZ p . (106) shows that the time evolution of the pdf is controlled by a ¯ux in the sample space c p.3. x. x. Eq. x. then the pdf takes the form of a double peak function. useful to obtain all equations discussed in this section. x. t 2t 2c 2t 106 where conditional averaging Q cd cp 2 c x. x. (102) are key relations. After simple manipulations combining Eqs. The time evolution of a probability density function is easily derived by expressing the pdf as the average of a function d cp 2 c x. x. c . ² To solve a balance equation for the pdf. t d c p 2 c x. t.4. x. (102) and (103) with Eq. In other words. L. t 102 The time evolution of the reactive species c x. t 22 p 1 p2 d cp 2 c x. tdYF 0 p variable c. t Q cuc x. t 2C 2 2 2 22 c x. t=2t is non-zero for the value c cp . t p p P c . t dZ p . Z 1 2 Z . x. Pdf balance equation The pdf balance equation is ®rst written for the progress . the probability of ®nding the occurrence c cp is modi®ed. t is introduced. t 1=Dcp if cp 2 Dcp =2 . Presumed probability density functions The simplest approach to the mixing between fuel and oxidizer in a non-premixed ¯ame is to presume the shape ~ of the pdf P Z p . t 1 À Áb21 1 Z 1 a21 1 2 Z 1 dZ 0 P cp . when the mean of 2c x.2). (104) the transport equation for the pdf may be written: 42 5 3 2 p 2 2c x.4. t de®ned as [100]: À Á d cp 2 c x. x. t 2C 2c 2C 2 2 2c x. t 1 v r 2t and d c p 2 c x. x. Z p . 6. t 2 p uc x. t Pc YF uZ p . t d cp 2 c x. t p d c 2 c x. t Z are linked to the internal structure of the ¯ame front. t Z p YF Pc YF uZ p . cp 1 Dcp =2 d cp 2 c x. t 2c p 2C 104 103 22 d cp 2 c x. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 219 and de®ned for a given value of Z as: p p P YF . ba 21 $0 aZ ~ f Z Z HH2 This technique requires the solving of a balance equation for f the mean and the variance of Z (Section 9. Veynante. with peaks located at Z 0 and Z 1: This approach is also used in premixed ¯ames replacing the mixture fraction Z by the progress variable c [99]. t dZ p and using the relation [98]: 1 À Á ~ Z p n P Z p . Two main directions may be chosen to built numerical models from pdf: ² To presume the pdf shape from available mean quantities. x. 6. x. t The conditional mean is readily de®ned from: 1 p p p YF uZ x. t and p YF uZ x. These relations (104) and (105) and Eq. x. t c P c . tdcp 1 0 The two-parameters a and b are determined as: 2 3 ~ ~ 1 ~ Z 1 2 Z 2 1 $ 0. t 2 2c 2c 2C 32 2 2c x.2. t d cp 2 c x.

tdcp 0 42 D 2r 2Yi 2Yj uY x. t cp P cp . t is usually negligible compared to the transport due to velocity ¯uctuations. x. x. t 2t 2 nt 7P c p . ¼. x. t 2t i 2 h 2 p 2u´7cuc x. x. ¼. YN d Y1 2 Y1 ¼d YN 2 YN 93: ~ p r 2 ~ p p P Y1 . t r cp d c x. the ¯uctuating part becomes: 2nt 7P c p . ¼. x. t cp P cp . L. x. This term insures that the density of probability is transported from point to point in the sample space c p in a conservative manner. x. t cp v cp _ _ The main advantage of pdfs in turbulent combustion lies in this availability to deal with chemistry. t c r named the micromixing term and remaining unclosed (as any term involving spatial derivatives). x. t leading to the pdf balance equation: 2 p P c . the pdf balance equation is derived using the same formalism with p p p p r P Y1 . t cp P cp . ¼YN . x. The ¯ux depends on the conditional mean of the time evolution of the progress variable given by Eq. x. The total dissipation rate. x. D7 2 P cp . t Y p 2xk 2xk 3 Using diffusion or scalar dissipation rate. ¼. t _ 2c i 22 h 2 p2 xuc x. t _ ² In terms of scalar dissipation rate: 2 p P c . x. YN r Y1 . x. t 1 p2 xuc x. YN . tuc x. (107). t _ 2c i 22 h 2 p2 xuc x. t=2t also includes a conditional diffusive term: 2 3 1 p 108 7´ rD7cuc x. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 an increase or a decrease of the density of probability for c cp . t 2 cp r cp P cp . using Eq. t and ¯uctuating components. t cp P cp .220 D. x. (105) and assuming rD < cst : i 2 h D7 2 cuc x. t When more than one species is taken into account. x is recovered as: x 1 xuc x. x. t 1 u´7P c p . averages are also introduced in pdfs. for instance when r r c : ~ r P cp . t N N 2 2 2Yip 2Yjp i1 j1 The ®rst term in the RHS. t (111) . t 2t 2r N 2 h 2u´7Y i uY x. (103): 2 3 2c uc x. x. Nonetheless. x. x. t 110 2c Weighted. t c 1 v uc x. the probability density function balance equation may be organized in two different forms: 5 p ~ p Â P Y1 . and the conditional mean of 2c x. t c p 2 1 3 1 p 7´ rD7c x. x. t Y p 1 v p _i 2Yip i1 i p ~ p Â P Y1 . or Favre. YN . implies a modi®cation of this density for other values of c p. t 2t 422 3 2 1 2u´7c 1 7´ rD7cuc x. x Du7cu 2 . (102) with Eq. t 2cp i 2 h 2D7 2 P cp . t cp P cp . t c p 2 p 2c r 5 3 (109) 1 v cp P cp . t 2c The convective term may be split into mean. this advantage is offset by the fact that reactants are brought to the reaction zone by diffusion. t cp 2t 2u´7cuc x. x. t cp _ r 107 In Eq. t 2c 2 ² In terms of molecular diffusion: 2 p P c . any term de®ned in one-point (as chemical source) is closed. t 2 p v cp P cp . the conditional value of the source term is a function of c de®ned in one-point and is exactly known since: v cuc x. Using an eddy viscosity model. u´7P c p . This micromixing term may be rewritten with the scalar dissipation rate. Veynante. x. x. t c p 1 v cp P cp . justifying the convective term in sample space.

The unclosed terms (RHS) are the pressure ¯uctuations. EQi appears from molecular diffusion along with differential diffusion effects across the iso-c surface. On the other hand. All these phenomena remain to be closed. convection of the pdf in velocity space. the convective term in physical space can be decomposed into mean and ¯uctuating parts. between fuel and oxidizer in nonpremixed combustion. t c p . turbulence models. but also for the conditional value of the chemical source of Yi. ² The pdf of the progress variable P c p . c . x. Conditional moment closure (CMC) Conditional moment closure modeling was ®rst proposed in Refs. One equation has to be solved for each value of c p retained. c cp P up .c c 2ui r 2xi |{z} Pressure fluctuations ~ In premixed ¯ames. x. L.4. The transport equation for this joint composition/velocity pdf is given below for the progress variable c. the probability density function entering expression (113). such as k± 1 .5. The number of these values is determined from the accuracy required to estimate both the mean from Eq. x. whereas for non-premixed combustion. and ®nally the closed chemical source. micromixing (x enters this term) and chemical source. the joint velocity/ concentration pdf is introduced. t 1 up ´7P u p . In premixed ¯ames. we consider the case of a turbulent premixed ¯ame represented with a progress variable c. Without loss of the generic character of the discussion. À Á any conditional quantity rYi uc cp corresponding to the conditional average of Yi along the iso-surface c cp : 6.105]: 2Q 2Q 22 Q i i i 2 rui uc cp ruc cp 1 rxuc cp 2t 2xi 2cp2 1 v i uc cp 1 EQi 1 EYi _ (114) The two last terms of Eq. EYi represents the effects of turbulent ¯uctuations on the deviation from the conditional mean. no longer required for the mean ¯ow. calculated for a given value of the progress variable c p or the mixture fraction Z.4. Closures for the conditional values of the scalar dissipation rate. u H ´7cuc x. 6. the mean value.e. three quantities are useful: ² The scalar dissipation rate of the progress variable ~ r x rD7c´7c. However. x. cp . x. t 2t 1 2p 2 p p P u .103]. a priori. cp .4. c . The three ®rst terms on the right hand side are unclosed. t: ² The ¯ame surface density S or the mean ®eld G: Simple links exist between these quantities: ² The conditional value of the scalar dissipation rate enters i 2 h 2 p n7 2 u Hi uu up . t is presumed. t _ 2cp 42 5 3 2 1 2p H p p P up . Once this joint pdf is known. P cp . cp . [102. c cp P up . the ¯uctuating part of the viscous dissipation and micromixing. x. cp . convection in physical space by the random velocity ®eld (incorporating turbulent transport). x. x. (110). are. . the conditional quantity is the progress variable c. tdcp 113 0 One may solve a balance equation for the conditional quantities Qi(c p) de®ned as: rYi uc cp Qi cp ruc cp This balance equation is [104. here the convective velocity is the mean of the viscous dissipation and the mean pressure gradient. they describe convective transport. CMC may also be viewed as a multi-surface description. t 2ui | {z} Viscous dissipation 2 i 2 h D7 2 cuu up . (113) and the second order derivative in the sample space (i. cp . (114). the mixture fraction is used. As was done in Eq. 6. t p u u .107]. As with the pdf. To describe this turbulent ¯ame. here only conditional moments rYi uc cp are considered. except when the frequency of mixing is also included in the joint pdf [101]. Similarities and links between the tools Major links between the tools used in turbulent combustion modeling are now developed.D. Y i of Yi may be estimated as: 1 ~ rYi rYi uc cp P cp . x. t 2cp | {z } Molecular diffusion 112 The LHS terms are closed and represent respectively accumulation. are required [106. 2 p p P u . and micromixing. Veynante.5. However. where the RHS contains the unclosed terms. Joint velocity/concentrations pdf To avoid the gradient transport assumption for the turbulent ¯ux. (114) are usually neglected. t 1 n7 2 ui 2 r 2xi 2up i 1 2 v cp P up . 2 2/2c p2) found in Eq. x. an equation for 1 is still needed to estimate a characteristic mixing time. or. the idea is to focus on particular states between fresh gases and fully burnt product in premixed ¯ames. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 221 The same terms are identi®ed on the RHS: convection by the conditional velocity and by the chemical source.

x . In premixed combustion. (96). cp dY1 .222 D. Yi and a sampling scalar c: v _ Y1 116 where the generalized surface average kQl s is de®ned by Eq. YN . t.cuccp _ 122 p p decomposing the joint-pdf as P Y1 . Y1 . ¼. ¼. based . S u7cuuc c p P cp (Eq. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Table 2 Exact relations between mean scalar dissipation rate. YN ucp dY1 . The CMC formalism (see Section 6. the ¯ame surface S . x. Since x is proportional to u7cu. x . of iso-surface c c p and ~ generalized ¯ame surface density S u7cu: Connections with the G-equation are readily obtained with Eq. c is the conditional variable (progress variable in premixed ¯ames. the mean c-scalar dissipation rate is also a function of surface densities: 1 ~ r x rDu7cu 2 krDu7cul s S c p dcp krDu7cul s u7cu 0 where the conditional mean Qi cp may be written: rYi uc cp krYi =u7cul s p Qi c kr=u7cul s ruc cp 119 and appears as directly related to c cp surface averaged quantities. (91) with Eq. the ~ pdf P c p and the dissipation rate x are very strongly related. Relations between ¯ame surface densities and scalar dissipation rates were anticipated by Borghi [108]. (86)). temperature. kQl s is a surface average (Eq.5) may be reorganized in term of ¯ame surface density: 1 1 ( rY ) i ~ rYi rYi uc cp P cp . The ¯ame surface density is then: S G 0 u7GuuG G 0 P G0 117 Hence. the ¯ame surface density may be written: 2r 3 x uc cp P cp . YN ucp P cp : Note that such a relation may be used when in premixed ¯ames the progress variable. ¯ame surface density function and generalized ¯ame surface density are then related via the scalar dissipation rate. tdcp S c p dcp 0 0 u7cu s 118 ¼ YN 1 0 p p p p p v i Y1 . cp p p Pc Y1 . (117).YN . ¼. ¼. YN . dYN P cp dcp _ p Y1 YN |{z} À Á ¼ v . x. x. tdxp 115 D x p and characteristic length scales dc D=x of the iso-c p distribution are embedded within S (c ). t D s xp p p P c . P cp . ¼.¼. The mean burning rate is given by: v _ 1 0 v cp P cp dcp _ 120 Using relations (86) and (91) and Eq. is related to the pdf via the conditional value of u7cu. or surface density.4. x. ¼. x. the ¯ame front is identi®ed to a given level G G0 [29]. conditional means. dYN dcp _ p 1 0 ! p p p p p v i Y1 . When the local reaction rate is a function of various quantities (species mass fractions. (86) and Eq. t S c p .). ¼. cp Pc Y1 . Veynante. Using the joint pdf of c and x . L. (96). cp P Y1 . probability density function. YN . previous relations may be recast in terms of the Gequation. ¼. this expression becomes: 1 1 ( v ) _ v v uc cp P cp dcp _ S c p dcp _ 0 0 u7cu s ( ) ( ) ( ) v _ v _ 1 v _ u7cu p S rx ~ x=D s u7cu s krDu7cul s u7cu s 121 Models based on probability density functions. t and x : ² The surface density function S c p . therefore there is a ~ direct connection between P c p . mixture fraction in non-premixed combustion) Scalar dissipation rate r x ~ Probability density function P cp 1 0 Flame surface density S c p 1 0 rx ~ P cp S u7cu ± xuc cp via PDF Eq: rx ~ krDu7cul s rDu7cu 2 uc cp P cp dcp u7cuuc p P cp dcp ± 1 0 krDu7cul s S c p dcp S c p = u7cuuc c p 1 S c p dcp 0 the pdf transport equation (Eq. These expressions are extended to burning rate depending on many species using multi-dimensional pdfs and conditional averaging. The links between the combustion modeling tools are summarized in Table 2. (116). YN . kQls corresponds to generalized surface average and is de®ned in Eq. etc. (110)). (96). (91)). S (c p). x. or ¯ame surface. Combining Eq. using a G-®eld equation.

etc.113]. etc. S (c p) of iso-surface c c p : Connections with the G-equation are readily obtained with Eq. x. [114]. 7. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Table 3 Tools for turbulent combustion modeling Description Geometrical Tools G-®eld with G Gp at the ¯ame ~ ~ ~ ~ 2G=2t 1 u´7G r u =r ST u7Gu Flame surface S Algebraic closure or transport equation Small scales Statistical Scalar dissipation rate. From experimental data [109.). x . t Z p 1 YF Pc YF uZ p . Z p . (91)). t (1) Presumed 1 ¯amelets (1) Strategy to generate the pdf ~ (2) Transport equation (2) Solve 2P=2t ¼ Conditional mean. P cp or surface density. u' is the turbulent velocity (i. Turbulent ¯ame speed Turbulent premixed ¯ames may be described in terms of a global turbulent ¯ame speed ST.D. but at the light of these relations. Eddy-Break-Up model Devised in Ref. Unfortunately. The previous relations may be used to carefully compare the proposed closure schemes and Table 3 summarizes the tools and their related modeling issues. Veynante. kQls corresponds to generalized surface average and is de®ned in Eq. t u7cuuc c p P cp . ¯ow geometry. many closures are essentially equivalent.e the RMS velocity). probability density function. The fundamentals of turbulent combustion modeling clearly rely on pdf. non-unity Lewis numbers. G-®eld and scalar dissipation rate concepts. Experimental data exhibit a large scatter because they depend on various parameters (chemistry characteristics. While this global approach is not particularly well suited to close Favre averaged transport equations.1. it may be of interest in the context of Large Eddy Simulation [79. (96) Tools Averaged reaction rate v 1 v _ _ S c p dcp 0 u7cu s u7cu s r 1 x r ~ x ~ p p krDu7cul s 0 krDu7cul s S c dc 1 0 Scalar dissipation Probability density function Flame surface density v cp P cp dcp _ v 1 v _ _ S c p dcp u7cu 0 u7cu s u7cu s . Various expressions for mean reaction rates are displayed in Table 4. the following expression has been proposed: 2 H 3n ST u 11a 123 SL SL where a and n are two model constants of the order of unity. CMC. (117). Major differences between the various approaches only appear when closing the unknown quantities. and the temperature are independent (compressibility effects. heat losses. kQls is a surface average (Eq. 7. ¯ame surface density. x. L. this model is based on a phenomenological analysis of turbulent combustion assuming high Table 4 ~ Exact expressions for the averaged reaction rate as a function of the mean scalar dissipation. tdYF 0 ~ Fundamental links: S x. x. the turbulent ¯ame speed ST is not a fully well de®ned quantity [112]. Reynolds-averaged models for turbulent premixed combustion 7. turbulence scales.110] or theoretical analysis (Renormalization group theory [111]).). r x rDu7Yu 2 ~ Algebraic closure or transport equation Modeling issues Propagation speed ST _ v i Combustion S Total stretch curvature 1 strain rate ~ Fast chemistry v < x _ ~ Provide the rate of micromixing _ Vi Turbulence 223 Probability density function p ~ Micromixing P YF . presumed PDF Micromixing and conditional source p p p YF uZ x.2. t and r x krDu7cul s u7cu on reactant mass fractions.

d cp and d 1 2 cp are respectively the Dirac delta functions corresponding to fresh gases c 0 and fully burnt ones c 1: Normalization of the probability density function: 1 0 P cp . especially in highly strained regions. (124) and (125) but. as described in Section 3. as: q f c HH2 v 2CEBU r 125 _ tt f Mass fraction ¯uctuations Y HH2 (or progress variable ¯uctuaF f ) must be modeled and may be estimated from a HH2 tions c balance equation (see Eq. tf cp . this model. x. Bray. or. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 È Reynolds Re @ 1 and Damkohler Da @ 1 numbers. The reaction zone is viewed as a collection of fresh and burnt gases pockets. assuming excess of oxidizer. irreversible chemical reaction between two reacting species. walls. etc. fully burnt and burning gases contributions (Fig. t t is estimated from the turbulence kinetic energy k and its dissipation rate 1 according to: chemical features [115]. the mean reaction rate is mainly controlled by the turbulent mixing time t t. c HH2 is easily estimated because c2 c : ~ ~ ~ r c HH2 r c 2 c2 r c2 2 c2 r c 1 2 c e ~ f 126 The square root has been introduced for dimensional reasons in Eqs. 15): P cp . c. constant chemical properties. private communication). x.). the mean reaction rate is expressed as: q f Y HH2 F _ v F 2CEBU r 124 tt where Y HH denotes the fuel mass fraction ¯uctuations and F CEBU is a model constant of the order of unity. fresh gases (R) and combustion products (P) is considered: R3P Classical assumptions are made to simplify the formulation: perfect gases. 7. The presentation is mainly limited here to basic concepts of the Bray± Moss±Libby (BML) formulation.1. td 1 2 cp |{z} |{z} fresh gases burnt gases tt k 1 as an approximation of the characteristic time of the integral length scales of the turbulent ¯ow ®eld. A one-step. Eqs. Introduction Known under the initials of its authors. incompressible ¯ows. where the ratio 1 /k is large (¯ame-holder wakes.1. t as a sum of fresh. Eddy-Break-Up modeling tends to overestimate the reaction rate. td cp 1 b x. Bray±Moss±Libby model 7.224 D. c. burnt gases and reacting mixture. BiModal Limit. t a x. _ 1999. When oxidizer is in excess. Moss and Libby. c 0 or c 1).e. of the chemical reaction is introduced where c 0 in fresh gases and c 1 in fully burnt gases.3.3. t. Champion and Libby). Veynante. t | {z } burning gases 129 128 0 where YF is the initial fuel mass fraction in the reactant stream. fresh gases. _ ~ ~ ~ d v =d c. The turbulent mixing time. A progress variable. L. at location x.). in terms of fuel mass fraction: 2 3 ~ ~ 1 YF YF v F 2CEBU r 12 0 _ 0 k YF YF 1 g x. turbulence leads to a break down of fresh gases structures. and then by Bray. Moss and Libby. etc. Following the Kolmogorov cascade. unity Lewis numbers. is in®nite both when c 0 and when c 1 (Borghi. Some adjustments of the model constant CEBU have been proposed to mimic where a . (77)). Accordingly. is used for practical simulations: 1 ~ _ v CEBU r c 1 2 c ~ 127 k or. Then. Assuming an in®nitely thin f ¯ame front (i. ®rst proposed in 1977 [116]. (125) and ~ (126) lead to inconsistencies because the c derivative of v . a corrected version of the Eddy-Break-Up model. unfortunately. this model has evidenced some special features of turbulent premixed combustion (counter-gradient turbulent transport. from the involved physical hypothesis. Combining a statistical approach using probability density functions and a physical analysis. The EBU model was found attractive because the reaction rate is simply written as a function of known quantities without any additional transport equation and is available in most commercial CFD codes. has been the subject of a large amount of work leading to many improvements (see papers by Bray. The basic idea of the BML formulation is to presume the probability density function of the progress variable c at a given location x. tdcp 1 130 leads to the following relations: a1b1g1 131 . etc. without the square root. The modeled reaction rate does not depend on chemical characteristics and assumes a homogeneous and isotropic turbulence. The reaction rate may be recast in terms of progress variable. x. ¯ame turbulence generation. b and g respectively denote the probability to have.

tdcp 1 132 with f 0 f 1 0: The balance equation for the progress variable c may be written: 2rc 1 7´ ruc 7´ rD7c 1 v _ 2t 133 function of the Favre average progress variable c : ~ 1 ~ ~ r c r c rcp P cp dcp rb b 0 138 where r b is the burnt gases density. (135)). assuming a constant pressure P. In this situation. (129)). t a x.3. x. tdcp _ 134 _ 0 rc ~ rc ~ and a 1 2 rb rb 139 The mean density r may be written: 2 3 1 c c p dcp aru 1 brb 1 2 re ru 1 re rb r rP c rb rb 0 140 One may also introduce the reaction heat release factor t . t . which is ®xed for a given chemical reaction). Accordingly. x. a @ g and b @ g ). This signal also satis®es: c2 c 1 O 1=Da and c 1 2 c O 1=Da 137 t ru T 21 b 21 rb Tu 141 leading to: ~ ru 1 1 trb r 1 1 tc 142 corresponding to the perfect gases state law. L. at the location x. t v c p P cp . t v cp f cp . BML model analysis This model is developed under the assumption that Re @ Da @ 1.D. t is: 1 v x.2. we will focus only on the case where Re @ Da @ 1. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 225 Fig. 7. Then at the point x. td 1 2 cp 1 O 1=Da 136 È which is `quasi-bimodal' for large Damkohler numbers. (129): 1 v x. t g x. t inside the reaction zone. g and the probability density function f. c looks as a telegraphic signal as displayed in Fig. The BML model involves a presumed pdf (Eq. corresponding to g ! 1. comparing turbulence and chemical time scales. 16. f cp . de®ned as: leading to Eq. Intermittency between fresh and fully burnt gases at a given location in the reaction zone. This signal corresponds to a bimodal (c 0 and c 1) probability density function. the turbulent Reynolds number Re. È Using the Damkohler number. 15. a and b are easily determined as a . and. Probability density function in premixed turbulent combustion.e. td cp 1 b x. The objective is now to determine the unknown functions a . g ! 1 (i. Starting from the conservative and non-conservative forms of the progress variable c balance equations: 2rc 1 7´ ruc 7´ rD7c 1 v _ 2t r 2c _ 1 ru´7c 7´ rD7c 1 v 2t and multiplying by c and adding these equations: 2rc 2 1 7´ ruc 2 7´ rD7c 2 2 2rD7c´7c 1 2cv 144 _ 2t Under this assumption. Da. 16. b This equation is averaged and the mean reaction rate. the combustion is controlled by the turbulent transport and the reaction zone may be assumed to be in®nitely thin. At a given location in the ¯ow. x. Then. Veynante. the probabilities a and b become: a ~ 12c . an intermittency between fresh gases c 0 and fully burnt ones c 1 is observed and the probability density function of the progress variable c reduces to: P cp . ~ 1 1 tc b ~ 1 1 tc ~ 1 1 tc 143 The probability density function P c is determined and ~ depends only on the mean progress variable c (and on the heat release factor. b . tdcp _ _ 0 135 All studies devoted to this line of models are based on such a formulation. x. the mean reaction rate cannot be calculated from the pdf since g was neglected in Eq. 1 0 Fig. however.

226

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266

Subtracting the balance equation for c 2 (Eq. (144)) to the balance equation for the progress variable c (Eq. (133)) leads to a balance equation for c 1 2 c : 2 rc 1 2 c 1 7´ruc 1 2 c 2t _ _ 7´ rD7c 1 2 c 1 2rD7c´7c 2 2cv 1 v 145

linking mean reaction rate and scalar dissipation may also be recovered from the general relations between modeling tools (Eq. (121)): _ v 1 krDu7cul s (

v _ u7cu

)

rx ~

s

153

Under the assumption of the BML model, the progress variable c is equal to zero or to unity. Accordingly, c
1 2 c 0 and the balance Eq. (145) reduces to: _ _ 2rD7c´7c 2cv 2 v leading to, after averaging: 2rD7c´7c
2c m 2 1v _ where a progress variable cm, de®ned as: 1 cv f
cdc _ 0 c m 1 v f
cdc _

0

146 147

To link mean reaction rate and scalar dissipation rate using the ¯amelet analysis, Eq. (153) requires estimates for v =u7cu _ and rDu7cu; averaging of these quantities along iso-surface c c p and integration over all the possible c p values. Assuming that in®nitely thin ¯ame elements may be viewed as one-dimensional, steady state, premixed laminar ¯ame propagating at a given laminar ¯ame speed SL, the balance equation for the progress variable c in this 1D ¯ame is:

r 0 SL 7´c 7´
rD7c 1 v _

148

154

where r 0 denotes the density in the fresh gases. Integrating Eq. (154) across all the ¯ame front from fresh gases to burnt gases 21 # x # 11 and up to a given location x0, corresponding to a progress variable c0 21 # x # x 0 ; leads to:

has been introduced and characterizes the chemical reaction. The mean reaction rate v becomes: _ v2 _

r0 SL

1 1

21

v dx _

rx 2cm 2 1

**1 v _ dc 0 u7cu x0
**

21

155

149 rD7c x0 r0 SL c0 2

where rx is the scalar dissipation rate of the progress variable c. e rx rx rD7c´7c rD 2c 2c 2x i 2xi 150

v dx _

156

! c0 v 1 v _ _ dc c 0 2 dc 0 u7cu 0 u7cu

The mean reaction rate v is then related to the dissipation _ ~ rate x ; describing the turbulent mixing, and to cm, characterizing the chemical reaction. A transport equation for the scalar dissipation rate may be written and solved [117], or one may postulate a linear relaxation of the ¯uctuations generated by micromixing (Section 9.2.2), leading to:

As the ¯ame front is supposed to be thin and planar, all c p iso-surfaces have the same surface and all quantities are constant along these surfaces, moreover the ¯ame is supposed locally one-dimensional and u7cu 7c: Then: 1 1

0 0

kv =u7cul s dcp _ krDu7cul s dcp

1

0 1 0

v =u7cudc _ rDu7cudc v dx _

0 0

rx

rc HH2 tt

151

where a turbulent time scale, t t is introduced. Assuming an intermittency between fresh and burnt gases (c 0 or f c 1), c HH2 is given by Eq. (126). Then: ~ 2 r c 1 2 c ~ _ v tt 2cm 2 1 152

1 1 1 1

21

v dx _

2

1

0

21

c dc 2

1 c 0

v =u7cudc dc 0 _

The Eddy-Break-Up model expression (Eq. (127)) is recovered (Section 7.2). The BML model may then be viewed as a theoretical derivation, where the assumptions made are clearly stated, of the Eddy-Break-Up (EBU) model, initially based on a phenomenological approach. 7.3.3. Recovering mean reaction rate from tools relations The analysis developed in Section 7.3.2 is the usual derivation of the BML model. However, expression (149)

v dx _ 2 1 1 1 1 2cm 2 1 2 cv dx 2 _ v dx _

21 21 21

1 1

157

**where cm is de®ned as: cv dx _ 2 cm 11 1 v dx _
**

21

1 1

158

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266

227

~ Fig. 17. The Reynolds average c of the progress variable cis plotted as a function of the Favre average c for various values of the heat release factor t , assuming a bimodal distribution (c 0 or c 1) of c (Eq. (165)).

Then, from Eq. (153): v 2 _

rx 2cm 2 1

159

of the progress variable c, which may be recast in terms of ¯ame surface density, exhibiting simple links between modeling tools. 7.3.4. Reynolds and Favre averaging Assuming a bimodal distribution of the progress variable ~ c, Reynolds
c and Favre
c averages are easily related. From: 1 c cP
cdc b
163

0

recovering Eq. (149). This second analysis leads to the following comments: ² The de®nition of cm is slightly different in Eqs. (149) and (159). In the ®rst one, the integration is performed over all possible values of the progress variable c at a given location x; t: In the second analysis, the integration is performed along a normal direction to the ¯ame front, assumed to be a laminar one-dimensional premixed ¯ame. The expressions (148) and (158) are identical when: 1 dc 21 f c 160 dl dx f where dc=dxf describes the inner structure of the ¯amelet. ² This second derivation clearly exhibits how a ¯amelet model works (the turbulent ¯ame is viewed as a collection of thin laminar ¯ame elements). The cm parameter is related to the inner structure of the ¯ame front and to the properties of the chemical reaction. ² Making use of the ¯ame surface density S as de®ned by Eq. (116), the expression for the mean burning rate becomes: _ v r 0 SL S where 161

rc r c ~

1

0

rcP cdc rb b rb c

164

together with expressions (142) and (143), one easily obtains: c ~ 1 1 tc ~ 1 1 tc 165

corresponding to a model for the density/progress variable correlations (Eq. (15)):

r H c H 2r

~ ~ tc 1 2 c ~ 1 1 tc

166

~ Reynolds
c and Favre
c averages of the instantaneous progress variable c are compared on Fig. 17 for various values of the heat release factor t . The discrepancy between the two quantities strongly increases with t . r H c H is plotted ~ as a function of c on Fig. 18. 7.3.5. Conditional averagingÐcounter-gradient turbulent transport The analysis of intermittency between fresh gases
c 0 and fully burnt gases
c 1 leads to the introduction of ~ conditional averaging. The Favre average Q of any quantity Q may be expressed as a function of the conditional averages for fresh gases
Qu and fully burnt products
Qb : 1 r Q rQ rQP
cdc aru Qu 1 brb Qb
167 ~

0

S

~ ~ c 1 2 c 2 1 2 1 rx < ~ r tt 2c m 2 1 r0 SL 2cm 2 1 r0 SL 162

² The BML model has been developed starting from a statistical analysis (two peaks pdf), the mean reaction rate was written in terms of the scalar dissipation rate

228

D. Veynante, L. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266

~ Fig. 18. Correlation 2r H c H =r as a function of the Favre averaged progress variable c for various values of the heat release factor t , assuming a bimodal distribution (c 0 or c 1) of c (Eq. (166)).

using Eqs. (142) and (143): ~ ~ ~ Q
1 2 cQ 1 cQ u b
168 where Qu and Qb are de®ned as: 1 1 Qu QPc
Quc 0dQ

21

**using the same formalism:
**

u b g ~ ~ ~ ~ i i j j u HH u HH
1 2 cu Hi u Hj 1 cu Hi u Hj 1 c
1 2 c
ub 2 uu
ub 2 uu i j | {z } intermittency

172 where one may note a weighted mean between the Reynolds À À uÁ bÁ stresses in the fresh u Hi u Hj and in the burnt gases u Hi u Hj representative of turbulent motions. The additional term represents the intermittency between fresh and burnt gases. 7.3.6. Extensions to partially premixed combustion Some attempts have been made to extend BML modeling concepts to partially premixed combustion, i.e. when reactants are not perfectly premixed before burning [118]. Two dif®culties are then encountered. First, in the de®nition of the progress variable c (Eq. (35)), unburnt and burnt gas u b temperatures and fuel mass fractions Tu, Tb, YF and YF are no longer constant and depend on local equivalence ratio. A balance equation may be still derived (see, for example, Ref. [119]) but it incorporates additional terms and is not obvious to close, especially when rich and lean zones coexist in the same ¯ow ®eld. Mixing should also be taken into account through a mixture fraction Z. The challenge is then to model the joint probability density function P
cp ; Z p ; x; t: Lahjaily et al. [118] write: P
cp ; Z p ; x; t a
x; td
cp Pu
Z p ; x; t 1 b
x; td
1 2 cp Pb
Z p ; x; t 1 g
x; tF
cp ; Z p ; x; tH
Z p 2 Zmin |{z}

burning zones

Qb

1 1

21

QPc Quc 1dQ

Pc Quc is the pdf of Q for the given value c of the progress variable (conditional pdf, see Section 6.4.1). ~ The components uk of the mean velocity vector u may be ~ written as a linear combination of their conditional fresh and burnt gases averages: ~ ~ i ~ i ui 1 2 cuu 1 cub Then: ~ i ~~ ~~ i ~ i ~ r u HH c HH r ui c 2 ui c r cub 2 ui c r c 1 2 c ub 2 uu f g i 170 which is the scalar turbulent ¯ux, generally modeled using a gradient assumption: ~ m 2c r u HH c HH 2 t g 171 i Sc 2xi The two expressions (170) and (171) may describe opposite ¯uxes: consider a left-traveling one dimensional turbulent ¯ame, because of thermal expansion the conditional veloi city in the burnt gases, ub ; is expected to be larger than the i conditional velocity in the fresh gases, uu : According to Eq. g (170), the turbulent ¯ux, u HH c HH is expected to be positive. On i the other hand, as the mean progress variable gradient is also g positive, Eq. (171) leads to a negative value of u HH c HH : This i situation, called `counter-gradient turbulent transport', is a key point of the BML analysis and will be further discussed in Section 8. g The Reynolds stresses u HH u HH may also be decomposed i j 169

**1 gm
x; tFm
cp ; Z p ; x; t1 2 H
Z p 2 Zmin |{z}
**

mixing without burning

(173)

where Pu Z p ; x; t and Pb Z p ; x; t represent the mixture fraction distributions in fresh and burnt gases respectively. F and Fm are the distributions within the reaction zones,

Algebraic expressions for the ¯ame surface density S Assuming intermittency between fresh and burnt gases Fig. Combustion is supposed to occur when Z p $ Zmin (H is the Heaviside function verifying H z . Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 229 Fig. Experimental burner. 7. A propane/air premixed ¯ow is injected in a rectangular burner through a grid.2.121].4. 19. Equivalence ratios f are in the range 0. Models based on the ¯ame surface area estimation 7. Their derivation and their histories differ. The ¯ame surface density is here introduced at the light of experimental data from Refs.1. (b) f 0:9: Flow rate: 35 g/s [121]. L. t: 7. ¯ame surface density pro®les are plotted as a function of the transverse coordinate for various downstream locations and for two equivalence ratios. The turbulent ¯ame is stabilized behind a small cylinder (blockage ratio of 6%) [120. .4. A turbulent premixed propane/air ¯ame is stabilized behind a small cylinder (blockage ratio of 6%). These models assume that the chemical reaction occurs in thin layers separating fresh gases from fully burnt ones (high DamÈ kohler number limit). [120. 20.4. Introduction Several ¯ame surface density models are now discussed. Veynante. 0 0 and H z $ 0 1). t and Pb Z p . Under BML assumptions. corresponding to inlet velocities between 10 and 30 m/s (turbulence levels from 5 to 10%). described in Section 6.3. Flow rates are about 35±100 g/s. [120±123]. (a) f 0:78. the two last contributions are neglected (g p 1 and gm p 1) and dilution effects are incorporated through Pu Z p . assuming to be ¯amelets. CH and C2 radical emission (reaction rate estimation) and high-speed laser tomography (¯ame front characteristics) measurements have been performed and are described in Refs. The experimental burner is displayed on Fig. Velocity (laser Doppler velocimetry). Transverse ¯ame surface density (S ) pro®les (m 21) plotted as a function of the transverse location for various downstream locations (mm downstream the rod).1.7±1. 20 (half burner). x. and within the non-reactive mixing zone between fresh and burnt gases.2. In Fig. but they are all based on similar concepts. 19. The reaction zone may then be viewed as a collection of laminar ¯ame elements called ¯amelets. x.121].D.

a submodel is required for the wrinkling length scale which increases with the downstream location in our experiment. The reaction rate per crossing ¯ame. The discrepancies are easily explained by the crude model used for the scalar dissipation of the progress variable to derive Eq. the mean reaction rate v is proportional to the intermittency _ between fresh and burnt gases. may relate the wrinkling length scale Ly to the thickness d B of the mean ¯ame brush. Then: C B g 1 n ´hni d 181 Ly < a unp ´nu s p s B Assuming that a . According to Gouldin [125. wF (Eq. np (see Section 6. (182). 16).3). On the other hand. the ¯amelet orientation factor s y (Eq. In Eq. (179)) and the ¯ame surface wrinkling factor J (Eq.3): B C 1 1 179 sy unp ´nu s Despite of the same normal vectors n and np involved. 22 where the ratio S=c 1 2 c. the ¯amelet orientation factor s y is directly linked to the vectors normal to the instantaneous ¯ame front. Veynante. (174)). d B/J measures the wrinkling length scale Ly of the ¯ame front. and inversely proportional to t t.3. the csignal is not exactly bimodal). (100) and (177). (162). SL and d l are respectively the speed and the thickness of the laminar ¯ame.230 D. n.126]. _ Estimating the local reaction rate per unit ¯ame area V c _ c r0 SL where r 0 is the from the laminar ¯ame speed SL (V fresh gases density) the mean reaction rate becomes. the ¯ame surface density S and the local reaction rate per ¯ame crossing (wF) are not obvious to estimate. determined from c 1 2 c ~ ~ or c 1 2 c. for example from time-resolved local temperature measurements (thermocouple). Moss. leading to: Sg c 1 2 c c 1 2 c Ju7cu < aJ s y Ly dB 180 n 2c 1 2 c ^ T 175 ^ where T is time scale for the ¯uctuations of c. as displayed on Fig. This analysis is attractive because the crossing frequency n may be easily obtained in experiments. Libby and their coworkers have proposed to describe the mean reaction rate v as the product _ of a ¯ame crossing frequency n and a local reaction rate per ¯ame crossing. L. generally assumed to be proportional to the integral length scale lt: n S 178 Ly Cl lt L uH where Cl and n are two constants of the order of unity [75]. leading to the algebraic expression [124] Sg c 1 2 c g 11t ~ ~ c 1 2 c s y Ly s y Ly 1 1 tc2 ~ 177 where g is a constant of order unity. Nevertheless.3. The previous expression is slightly different than the one proposed in Section 7. under the assumption of a in®nitely thin ¯ame front: Fig. combined with Eqs. when n1: ~ ~ g uH g u H 1 1 tc 1 2 c c 1 2 c r0 v r0 _ s y Cl lt s y Cl lt ~ 1 1 tc2 182 As tt lt =u H is a turbulence characteristic time scale. u H =lt is generally modeled from the r0 SL dt =tt 176 where r 0 is the unburnt gases density. and to the mean ¯ame brush. is displayed. 21. (Section 7. wF: v wFn _ 174 Because c(t) may be viewed as a telegraphic signal (Fig. 21 (in practice. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 ~ function of c: A submodel is required to describe the wrinkling length scale Ly. (165) has been used to replace c as a . g and s y are constant.3. This model has been latter rewritten in term of ¯ame surface density. (97)) are different quantities. is usually modeled as: wF where a is a model constant of the order of unity. Eq. De®nition of the ¯ame transit time tt is the ¯ame crossing frequency BML model. as shown in Fig. s y is a ¯amelet orientation factor measuring the mean angle of the instantaneous ¯ame front with the c surface and assumed to be an universal model constant s y < 0:5: Ly is a ¯ame front wrinkling length scale and Eq. The physical analysis leading to the Eddy-Break-Up model is recovered and a similar expression for the reaction rate is found. These relations. (175) is generally closed estimating ^ T from a characteristic turbulent time t t. The agreement between this BML model and the experimental data is very good. this crossing frequency is derived from a statistical analysis of the telegraph equation. Bray. The ¯ame transit time tt measures the averaged time spend by a point in the ¯ow to cross a ¯ame front and corresponds to the mean transition time between c 0 and c 1 levels of the progress variable. corresponding to g=s y Ly and assumed to be constant.

D. ITNFS is one of the closure schemes for the ¯ame surface density balance equation (see Section 7. Champion. 23. Moss. reducing the effective ¯ame strain rate. (b) f 0:9 [121]. G K only slightly depends on the velocity ratio u H =S L : The ef®ciency function does not reach a constant level when lt/d l increases since increasing lt/d l. The mean reaction rate is then [118. Because the ef®ciency function G K tends to counterbalance the known trend of Eddy-Break-Up modeling to overestimate the mean reaction rate in highly strained regions. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 231 Fig. plotted as the function of the mean progress variable c for several locations downstream from the rod: (a) f 0:78. Veynante. S= c 1 2 c transverse pro®les in mm 21 corresponding to g= s y Ly in the Bray. The BML model proposes a simple algebraic expression to estimate the ¯ame surface density s and the corresponding reaction rate. as displayed on Fig. G K also decreases. (177)). dt SL 3 where s log10 2 3 s1 uH SL 2 4 2 H 31=3 53 2 1 u 1 2 exp 2 3 2 SL lt dl and 185 1 c 1 2 c k 183 where a is a model constant and the ef®ciency function G K depends on the length scale (lt/d l) and the velocity u H =S l ratio comparing the turbulence and the laminar ¯ame characteristics. corresponds to an implicit increase of the turbulent Reynolds number Re. this simple approach improves the accuracy of the numerical predictions. but Bray.128. L. 22.129]: 2 _ v ar 0 G k lt u H .3 and Table 5). The ef®ciency function has been ®tted from DNS data [127. Libby model (Eq.4. turbulence k± 1 model as 1=k: The ITNFS ef®ciency function G k [127] may be introduced in the turbulent time scale to account for the reduced ability of small turbulent structures to wrinkle the ¯ame front.130]: log10 G k 2 1 exp2 s 1 0:4 s 1 0:4 2 2 2 3 3 uH s 2 0:11 SL 184 1 1 2 exp2 s 1 0:4 s 1 When the length scale ratio lt/d t tends towards zero. Champion and Libby have mainly focused their attention on a careful description of the turbulent ¯uxes using the balance equations for the . keeping u H =SL constant.

232 D. CA. S3. The turbulent strain rate is probably overestimated. Veynante. E and K are model constants. The strain rate due to thep turbulent ¯uctuations is estimated from the time scale n=1 of the Kolmogorov structures.3. (89) and (90)) are unclosed and require modeling.132]. a. Flame surface density balance equation closures The previously described balance equations for S (Eqs. n t is the turbulent viscosity and s S a ¯ame surface turbulent Schmidt number. 7. S2 is the strain rate due to the turbulent motions (Eq. D describes the consumption of ¯ame area. the turbulent kinetic energy and its dissipation. Three versions are presented in Table 5. respectively. developed by Candel and his coworkers following the initial work of Marble of Broadwell [84]. Re is the turbulent Reynolds number. the turbulent ¯ux of ¯ame surface density is expressed using a classical gradient assumption. d. Despite the fact that the Kolmogorov structures contain the highest energy. C. [117]. D is the fractal dimension of the ¯ame surface.4. (94)) and the third source term. The ¯ame surface density may also be derived from fractal theories. l . In the Choi model (CH) [137]. The cut-off scales are generally estimated from the turbulence Kolmogorov lk and the integral lt length scales. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Table 5 Comparison of source (Si) and consumption (D) terms in the ¯ame surface density balance equation in different turbulent premixed combustion models (details in text). In Table 5. leading to [133] 1 Louter D22 S 186 L outer Linner where Linner and Louter are respectively the inner and outer cut-off length scales (the ¯ame surface is assumed to be fractal between these two scales). g . b 0. In the two . occurs only in the derivation proposed in Ref. u H denotes the rms turbulent velocity and ltc is an arbitrary length scale introduced for dimensional consistency and combined to a 0 as a single arbitrary constant Model CPB [134] S 1 km S A ik 2uk S 2xi S 2 kt S S3 D a 0 CA 1=2 1 S n a 0 SL R 2 1 e2aR 2 S 3 1 2 c CFM1 CFM2-a CFM2-b [130] MB [117] Aik Aik Aik E 2uk S 2xi 2uk S 2xi 2uk S 2xi 1 a0 S k 1 a0GK S k 1 a0GK S k p 1 a 0 Re S k F 1 H H 2c uc S L k i 2xi u Hi u Hk 2uk S k 2xi CD [136] CH1 CH2 [137] 1 S a0 l S for k t # a0 K L k dL 1 1=2 a0 S 15n a0 uH S ltc 1 2 c1 SS L k p S 1C k 2 b0 L S 12c p S 1C k 2 b0 L S 12c p S 1C k 2 b0 L S c 1 2 c p SL Re b0 S2 S 2g c 1 2 c 1 1 d pL k S b0 L S 2 12c b0 b0 SL S2 c 1 2 c SL S2 c 1 2 c Reynolds stresses r u HH u HH and the scalar ¯uxes r u HH c HH to take gj g i i into account the possible occurrence of counter-gradient transport and ¯ame turbulence generation [131. a 0. G k is the ef®ciency function in the ITNFS model [127] and depends on the length scale (lt/d l) and the velocity (u H /SL) ratios comparing the turbulence and the laminar ¯ame characteristics. Five main closures are summarized: ² The CPB model [134] is derived from the exact transport equation for the ¯ame surface density. various closures found in the literature are compared where S1 is the strain rate acting on the surface and induced by the mean ¯ow ®eld. the strain rate due to the turbulent ¯uctuations is estimated from the characteristic time of the integral length scale (k/1 ). their lifetime is too short (because of viscous effects) to actually affect the ¯ame front [65]. In the initial version (CFM1). The modeled balance equation is rewritten as: 2S n ~ 1 7´ U S 7´ t 7S 1 S 1 1 S2 1 S3 2 D 187 2t sS In this expression. ² The coherent ¯ame model (CFM). k and 1 denote. L.

ltc. (184) [127]). at least to describe the ¯ame front propagation in a homogeneous and isotropic turbulent ¯ow ®eld. (205)). 1. the expression of the turbulent strain rate acting on the ¯ame front is improved from direct numerical simulations and multi-fractal analysis (ITNFS model. [135] has compared CPB. Veynante. this term S3.It was found that only the CFM-2 formulation is able to predict the so-called bending phenomenon. However. 10. A slight overestimate (respectively underestimate) of temperature is pointed out for CPB (respectively MB) and is probably due to the expression for the strain rate. (184)) as a function of the length scale ratio lt/d l for several values of the velocity ratio u H /SL (0. where the turbulent ¯ame speed decreases before the occurrence of ¯ame extinction when the turbulence level increases. Eq. as shown below. a ¯ame is able to sustain higher strain rates than expected. Such a term was tested in the coherent ¯ame model but the choice of the critical strain rate is somewhat arbitrary. based on the Kolmogorov turbulent time scale. L. ITNFS ef®ciency function G K (Eq. In the case of a statistically one-dimensional turbulent ¯ame propagating in a frozen turbulence. 11). The MB closure [117] is found to be more sensitive to the inlet turbulent quantities than CFM1 but CFM2 models. all these closures exhibit strong similarities. CFM1. Moreover. The transport equation is rewritten as a ¯ame surface density transport equation under the ¯amelet assumption (see Section 6. A comparison between these models to predict turbulent ¯ame speed ST may be found in Ref. the strain rate is only proportional to the turbulence intensity u H and an arbitrary length scale. as experimentally evidenced [110]. MB and CPB provide reasonable predictions of mean velocities and temperatures. ² The CH model [137] has been devised for spark-ignited engines to recover experimental data obtained in a closed vessel [138]. In CH2. In a ®rst analysis. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 233 Fig. In fact.5. (116)). succeeding formulations (CFM2). incorporating an ef®ciency function . This approach leads to a different expression for the source term S1 and an additional source term S3 is found. this critical strain rate cannot be deduced from planar strained laminar ¯ame studies because. the consumption term D is always proportional to S 2.e. A recent work. The CD predicts extremely high temperatures whereas CFM1. is incorporated in the model constant. The ®rst expression (CH1) corresponds to the closure in CPB model. ² The MB model [117] is based on an exact equation for the scalar dissipation rate x : r x rD ~ 2c HH 2c HH 2x i 2xi 188 assuming a constant density r . S3 corresponds to an anisotropic contribution of the turbulent strain rate aT (see Eq. For example. The destruction term differs in these two last formulations (CFM2a and CFM2b). a KPP (Kolmogorov± Petroski±Piskunov) analysis was used to analytically determine the turbulent ¯ame speed ST as a function of the model parameters.1.D. which does not depend on the available ¯ame surface density S. this analysis does not hold because the ¯ame surface density balance equation is derived assuming an established ¯ame surface. Eq. [130]. The consumption term D is similar to the one proposed in CFM2-b model whereas two formulations of the strain rate induced by turbulent motions are proposed. An additional term is proposed to take into account ¯ame extinction under excessively high strain rates.0. S3 seems to be negligible in practical simulations. MB and CD models to predict a turbulent premixed jet ¯ame. due to curvature and unsteady effects. might be viewed as an ignition term [135] involving a gradient of the mean fuel mass fraction or of the mean temperature (i. 23. ² The CD model [136] is similar to the ®rst version of the coherent ¯ame model (CFM1). fresh gases are ignited by heat transfers). Despite these comments.

On the other hand.141]: knl s S MS 27c 197 ~ where c and c are related using the BML relation r c r b c ~ (Eq. (90)): 2S ~ 1 7´ uS 1 7´ ku HH l s S 2t ~ 7´u 2 knnl s : 7u S 1 k 7´u HH 2 nn : 7u HH l s S | {z } | {z } AT aT This relation. see for example [75.4.2. AT. Assuming that the conditional velocity on the ¯ame ~ surface. These authors have also shown that the isotropic assumption made in Ref. Strain rate induced by the mean ¯ow ®eld. M. AT. Veynante. a relation between orientation factors and Reynolds stresses is indirectly proposed: kni nj ls g u HH u HH i j 2k 199 191 where c denotes the c-level chosen to de®ne the ¯ame front.2) and the summary of the most popular closures (Section 7. (169)): ~ ~ ~ u 1 2 cuu 1 cub leads to: ~ ~ ku HH ls kuls 2 u cp 2 c ub 2 uu 193 192 where k is the turbulent kinetic energy. Flame surface density models are extended to non-isenthalpic ¯ows (premixed reactants injected in a co-¯ow of air [139]). the aim of this section is to carefully analyze this balance equation.3.121]. its contribution to the strain rate AT induced by the mean ¯ow ®eld is not that important.140]) have shown from a simple BML-type approach ~ that the turbulent ¯uxes of the mean progress variable c and of the ¯ame surface density (S ) are closely related. corresponding to the mixing layer shear stress. In the ¯ame-holder stabilized turbulent ¯ame investigated in Ref. The BML relation (Eq. 7.4. have not been tested. the orientation factors become: knnls MM 1 kmmls 196 Using the de®nition nu7cu 27c and assuming an in®nitely thin ¯ame front leads after averaging to [134.4. A counter-gradient turbulent transport will be observed simultaneously for these two scalar ®elds. Then. [134] is clearly wrong for the turbulent ¯ame stabilized downstream a small rod. The only unclosed quantities in the strain rate due to the mean ¯ow ®eld. A pioneering work has been ®rst conducted [118] extending the algebraic ¯ame surface density BML model to non-isenthalpic ¯ows. L. [117] leads to very good results [121]: 2 f u HH k g u HH u HH i j k±i kni ni ls . kul s is a linear function of the conditional fresh uu and burnt gases velocities: ~ kuls uu 1 cp ub 2 uu p dij 1 2 Mk Mk 3 198 ² In Ref. (138)). Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 including the inability of small vortices to wrinkle the ¯ame surface. a much more complicated model is proposed that will not be described here because its practical implementation is probably not so easy. is incorporated to account for dilution phenomena. The turbulent ¯ux of ¯ame surface density is generally expressed using gradient transport: n 190 ku HH l s S 2 t 7S sS Bidaux and Bray (1994.4.4. Analysis of the ¯ame surface density balance equation Following the description of the exact balance equation for the ¯ame surface density S (Section 6. Turbulent transport. [134] one assumes an isotropic distribution of the ¯uctuating components of the normal vector n: kmi mj ls each unclosed term may be investigated as follows: 7. (170) the ¯ame surface density turbulent ¯ux becomes: f ku HH ls S ~ HH c p 2 c gHH u c S ~ ~ c 1 2 c 194 . A presumed pdf for a mixture fraction Z. ² From experimental data [120]. kni nj±i ls 200 4k 2k These orientation factors have very important effects and may lead to surprising results. m: n M 1 m with knl s M and kmls 0 Then. because of the low value of the corresponding orientation factor. Following [134]. determined from mean and rms values of Z. the vector normal to the ¯ame front.4. are the orientation factors knnls. n may be split into a mean component. using similar concepts.1. [117]. This analysis may be based on direct numerical simulations [17] or on experimental data [120. Several closure schemes have been proposed: 195 2 7´kwnl s S 1 kw7´nl s S |{z} |{z } propagation curvature 189 ² In Ref. and a ¯uctuation.234 D. con®rmed by direct numerical simulations ~ [140]. unpublished work.3). shows that turbulent ¯uxes of c and S are closely related. Starting from the `propagative' form (Eq. only the ¯uctuation cross products kmmls remain unclosed.4. [121]. a slight modi®cation of Ref. But. 7. the main velocity gradient is the transverse gradient of the mean axial velocity. The strain rate AT is dominated by From Eq.

This ®nding explains why. this term is generally assumed to be proportional to the inverse of a turbulent time scale. aT is always modeled being isotropic despite the orientation factors nn occurring in expression (202). 7. n. either the Kolmogorov time scale (CPB model) or the integral time scale 1 /k (CFM. G k.D. 24. (b) Components of the source term ATS due to the strain rate induced by the mean ¯ow ®eld. (184)). as illustrated in Fig. The source term due to the strain rate aT is: B a T S k7´u HH 2 nn : 7u HH l s S d ij 2u HH 2u HH i 2 ni nj i 2xj 2xj C S s In most models. Notice also that the strain rate AT cannot be reduced to the simpli®ed expression ®rstly proposed [84]: ~ AT u7Uu 201 ~ Recent works [142] have shown that u7Uu may probably be viewed as a model for the total strain rate AT 1 aT and not only for the strain rate AT due to the mean ¯ow. Using the previous splitting of the normal vector to the ¯ame front. Nevertheless. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 235 Fig. the axial gradient of the mean axial velocity. L. combined with the geometrical relation (197). the mean strain rate AT was not included because the simple expression (201) clearly overestimates AT. Veynante. Strain rate aT due to turbulent motions. This turbulent time may corrected with the ef®ciency function. in previous versions of the coherent ¯ame model where the orientation factors kninjls were not incorporated.3. of the ITNFS closure (Eq. Data are plotted as a function of the transverse coordinate y for x 80 mm and f 0:9 [121].4. induced by the thermal expansion due to the combustion heat release. 24. MB and CD models). leads to: B C B C 2u HH 1 2u HH 2c 2c i a T S d ij i S 2 2xj s S 2xj s 2xi 2xj B B C C 2u HH 2c 2u HH 2c i i 1 mj 1 mi 2xj s 2xi 2xj 2xj B C 2u HH i 2 mi mj S 203 2xj s This interesting relation decomposes the source term due to 202 .4. (a) Transverse pro®les of the mean velocity gradients and the corresponding strain rate AT.

The previous analysis was based on both a theoretical analysis of the exact S -balance equation and experimental . as displayed in Fig. Mantel and Borghi [117] have shown that this term is negligible in numerical simulations of a ¯ame propagating in an homogeneous and isotropic turbulent ¯ow ®eld. the case when the mean progress variable c is close to zero (see Fig.8. A priori. the previous expression may be reduced using: B C B HH C 2u HH 2u i i mj kmj ls 0 204 2xj s 2xj s from the mj de®nition. The links between these closures are further discussed in Section 9. [117] from the derivation of a balance equation for ~ the scalar dissipation rate x of the progress variable c. Propagation and curvature terms. this last part is not proportional to the available ¯ame surface density S . The propagation/curvature term becomes: kw7´nl s S S L k7´nl s S 207 where the only unknown is the mean ¯ame front curvature k7´nls The curvature is positive when the ¯ame front is convex towards the fresh gases. a priori.121] in a V-shape turbulent ¯ame show that it is not always negligible and is of the same order as the curvature term. 25). On the other hand.2). etc. the strain rate induced by turbulent motions. the normal propagation term becomes: 7´kwnl s S S L 7´knl s S S L 7´27c 2S L 7 2 c 206 where the geometrical relation (197) is used. This term S3 corresponds to an anisotropic contribution depend ing on 7c and on the strain rate due to turbulent motions.4. a simple linear model may be proposed: k7´nl s < cp 2 c Ly 208 205 This simple analysis explains the third source term. curvature effects. Replacing Ly by its value as a function of the ¯ame surface density S from BML modeling (Eq. These two terms are analyzed together because they derive from the laminar ¯ame propagation and are related to the ¯ame front displacement speed w. curvatures are probably negative (¯ame convex towards the burnt gases) when c < 1: Assuming that the mean curvature is of the order of the inverse of the wrinkling length scale Ly. Nonetheless. Veynante. into two parts: a contribution depending only on turbulent characteristics that may be assumed to be isotropic (®rst and last terms) and an anisotropic contribution. which is. [120.). S3 found in Ref. we have: lim k7´nl s c30 Fig. (208)). Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 This stands as the main dif®culty of the propagative approach for the ¯ame surface density balance equation because various effects are incorporated in w (strain rate. Assuming a constant displacement speed w equal to the laminar ¯ame speed SL. 26. The modeling of these terms requires the description of the displacement speed w that may have values far from the laminar ¯ame speed SL [78]. This term is positive in the fresh gas side of the turbulent ¯ame brush and becomes negative in the burnt gas side. aTS .2) or for the pdf balance equation (Section 9. third and fourth terms). The expression differs from classical closure where kw7´nls is always negative (term D in Table 5).4.236 D. This term is generally neglected in models (see Table 5). where the mean ¯ame surface orientation occurs through the gradient of the mean progress variable (second. Flame front curvature analysis. 7.121].4. L. A linear relaxation is retained to model the mean curvature k7´nls (Eq.8. Assuming a gradient type closure for the turbulent transport of the fresh gases. This type of closure is common in turbulent reacting ¯ows and possesses similarities with the relaxation model used for the scalar dissipation rate (Section 9. Then: B C B C B C 2u HH 2u HH 1 2u HH 2c 2c i i i aT S d ij S 2 m i mj S2 2xj s 2xj s S 2xj s 2xi 2xj |{z } |{z} isotropic contribution anisotropic contribution 1 Ly and lim k7´nl s 2 c31 1 Ly Then. 25. When velocity and normal vector ¯uctuations are supposed uncorrelated (this assumption is well veri®ed in direct numerical simulation data). This ®nding is probably questionable in other con®gurations and further investigations are required. (177)) leads to the following model for the curvature term in the S -equation: kw7´nl s S < bS L cp 2 c 2 S c 1 2 c 209 where b is a model constant. S3 is proportional to 7c´7c and appears as a model for the anisotropic contribution of the turbulent strain rate.5. Experimental data from Refs. This trend is in agreement with the ®ndings of direct numerical simulations [17] and experimental measurements [120.

27. (141). increases [144].3. these models are unable to describe the ¯ame stabilization because the S equation (Eq. 26. where G k is the ITNFS ef®ciency function [127]. Flame surface density (half top) and mean reaction rate. In their formulation.5. (85). L. to recover the observed results. de®ned by Eq. have the same sign in some regions and cannot be approximated with Eq. this expression is useful to explain counter-gradient turbulent transport: Because of the thermal expansion due to combustion heat release. (170) Section 7. This surprising ®nding was explained by the work of Bray.5): ~ i i g r u HH c HH r c 1 2 c ub 2 uu ~ 212 i Even though conditional velocities are not obvious quantities. 2c=2xi . whereas the mean reaction rate remains low. Transverse pro®les of propagation (27´[knlsS ]). This ®nding displays one of the dif®culties of ¯ame surface density models. A related approach: G-equation As described in Section 6. 7.3. Turbulent transport in premixed combustion 8. Orientation factors.D. fresh and burnt gases are separated by an interface (high surface densities) where Fig. In this zone. the burnt gas conditional velocity. x 80 mm [121]. 7. (211)). the premixed turbulent ¯ame may be described using a level set approach. The turbulent ¯ux ku HH ls S is described using a classical gradient expression. (85)). An improved version of the closed balance equation may be proposed: 2S ~ 1 7´ uS 1 7´ ku HH l s S 2t ~ ~ 7´u 2 knnl s : 7uS 1 G k 1 bSL cp 2 c 2 u H lt .4.1. 7 2 c 1 b cp 2 cS 2 =c 1 2 c where b 0:4 and cp 0:5 are also displayed.6.121]. extracted from two different data sets. 8. the turbulent ¯uxes of the progress variable ~ c have the same sign as the mean gradient 2c=2xi : Countergradient transport also increases when the heat release factor t . Moss and Libby discussed in Section 7. in opposition with the modeled expression (Eq. 27. In addition. reaction rate and ¯ame surface density are roughly proportional. is likely to be greater than the fresh gas conditional i i velocity.1.3. . the turbulent ¯uxes of the progress variable c are directly connected to the conditional mean velocities in i i fresh uu and burnt gases ub (Eq. (211). All source terms in this equation are proportional to S or to S 2 and the equation cannot generate ¯ame surface when there is no initial ¯ame surface. an initiation effect must also be incorporated to account for ignition Ç time delay in the local reaction rate per unit surface. excepted close to the stabilization rod.4. estimated form CH radical emission (half bottom) ®elds are compared for f 0:90: Flame surface density data. Introduction data [120. In their present formulations. curvature (k7´nlsS ) and combined (k7´nlsS 2 7´[knlsS ]) terms (m 22) in the ¯ame surface density balance equation plotted as a function of the mean progress variable c: The consumption term modeled as 2S 2/ (1 2 c) and the new proposed closure. V i (Eq.20] have evidenced in some turbulent premixed ¯ames counter-gradient turbulent transport where the turbulent ¯uxes r u HH c HH and the mean g i ~ ~ progress variable c gradient. kninjls may be described using a formulation previously discussed. This phenomenon is known as counter-gradient turbulent transport or counter-gradient turbulent diffusion. ub . are g i generally modeled using a gradient transport hypothesis as for inert scalars (Section 7): ~ m 2c g r u HH c HH 2 t 211 i Sct 2xi where mt is the turbulent viscosity given by the turbulence model and Sct is a turbulent Schmidt number. According to Eq. Flame stabilization modeling The ¯ame surface density ®elds obtained from laser tomography and the mean reaction rate estimated from CH radical emission are compared in Fig. r u HH c HH . Theoretical [143] and experimental studies [19. the ¯ame surface density S is high. are not available from 30 to 70 mm downstream the rod [121]. SL dl 3 1 S 1 SL7 2c k (210) S2 c 1 2 c Turbulent ¯uxes of the progress variable c. f 0:90. Most of the modeling issues discussed above are then recast in terms of G-equation and modeling for the turbulent burning velocity ST [89]. (189)) is derived assuming that the ¯ame does exist. uu : Then. combustion has started but is not yet fully established. Veynante. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 237 Fig.

2. whereas the second one corresponds to a counter-gradient situation.146]. 8. 29. The ¯ame front is progressively wrinkled by turbulent motions. leading to a counter-gradient turbulent transport in agreement with expression (212). . simulations of twodimensional ¯ame/turbulence interactions were reported [140]. of the instantaneous balance equations. in the Rutland database. the burnt gas con ditional velocity. a planar laminar premixed ¯ame is ®rstly superimposed on a homogeneous isotropic turbulent ¯ow ®eld (Fig. Positive (respectively negative) values of ru HH c HH denote a counter-gradient (respectively gradient) turbulent transport. This result is ®rst surprising: ub is expected to be larger than uu because of thermal expansion due to the heat release. 30 for the two DNS simulations. Turbulent ¯ux ru HH c HH displayed as a function of the mean progress variable c (bold lines) in the tri-dimensional direct numerical ~ simulations from CTR (Ð) and Rutland (. L.2. The corresponding turbulent ¯uxes estimated using the BML expression (Eq. The main discrepancy between the two databases lies in the initial turbulence level which is higher in the CTR simulation u H0 =SL 10 than in the Rutland database u H0 =SL 1: Mean and conditional average velocities across the turbulent ¯ame brush are displayed in Fig. This section is devoted to these results.2. is higher than the fresh gas condi tional velocity.. in the CTR database. (211)) for these ¯uxes. (212)) are displayed for comparison using thin lines. this observation is Fig. and the turbulence decays in time. Introduction Gradient and counter-gradient turbulent scalar transport was observed in DNS of premixed ¯ame/three-dimensional turbulence interactions [145.-). Numerical con®guration. obtained from a direct solution without any closure models. However. The reduced costs of 2D simulations allow the investigation of a large range of ¯ame and turbulence parameters. These results lead to the following comments.238 D. Results Turbulent ¯uxes ru HH c HH extracted from the direct numerÂ ical simulations conducted by Trouve [145] (denoted CTR) and Rutland [146] are displayed in Fig. retaining a simple gradient closure (Eq. 8. Velocities are made dimensionless using the laminar ¯ame speed SL [140]. The various versions of this BML model propose an algebraic closure for the mean reaction rate of the progress variable c. uu .2. Direct numerical simulation analysis of turbulent transport 8. A plane laminar premixed ¯ame is initially superimposed to an homogeneous and isotropic turbulence. 29. As the numerical con®guration is statistically onedimensional in the propagating direction. 28). ² On the other hand. The ®rst database clearly exhibits gradient turbulent transport. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. The results demonstrate the power of DNS to help in the modeling of turbulent combustion. In these simulations. 28. whereas other models may lead to more sophisticated reaction rate formulations. the fresh gas conditional velocity is higher than the burnt gas con ditional velocity. ub . the ¯ame may be assumed to be in equilibrium with the turbulent ¯ow ®eld and relevant modeling information is extracted from DNS. quantities such as mean progress variable or mean turbulent ¯uxes may be extracted from averaging in the perpendicular direction. The occurrence of counter-gradient turbulent transport have been analyzed using direct numerical simulation (DNS) [140]. Then. ² As expected. After a time of the order of the eddy-turnover time of the largest turbulence structures.1. Veynante. but focus attention on the scalar turbulent transport (closure schemes for the turbulent ¯uxes r u HH c HH balance g i equations).

ub 2 uu . (194)).-). 30. Note that Eq. fresh gases conditional velocity. L.4. the progress variable c is not fully bimodal (c 0 or c 1).4).D. Physical analysis A simple physical analysis is now reported to derive a criterion predicting the occurrence of counter-gradient turbulent diffusion [140]. Conditional velocities across the turbulent ¯ame brush displayed as a function of the mean progress variable c in the Rutland (top) and the CTR (bottom) databases. i are directly connected to the surface-averaged ¯uctuating g velocity. Thus. ku HH ls (Eq. is determined primarily by thermal expansion and its .and two-dimensional direct numerical simulations lead to very similar results for this problem. The two-dimensional simulations conducted in Ref. The `increase'. the ¯ow ®eld is assumed to be statistically one-dimensional and only the turbulent transport in the g propagating direction. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 239 ~ Fig. in agreement with expression (212) and a gradient turbulent transport. [140]. Favre averaged velocity. u HH c HH . In both cases. u (Ð ). 31 have been used to analyze the occurrence of counter-gradient turbulent transport. uu (´´´). referenced as CRCT in Fig. 32: Low turbulence level: The ¯ame front remains smooth and the velocity jump between fresh and burnt gases. The gradient turbulent transport is clearly enhanced by an increase in the turbulence level u H /SL and decreasing values of the heat release factor t . kuls (.3. of the line delimiting gradient and counter-gradient turbulent transport when lt/d l decreases is due to the reduced ability of small scale turbulence motions to wrinkle the ¯ame front. u HH c HH will be described.. a model for u HH c HH may be i i deduced from a model for ku HH ls involving the conditional i fresh and burnt gases mean velocities: ~ ~ i i ku HH ls kui ls 2 ui cp 2 c ub 2 uu i 213 In the following. 29) because. ub (±) ~ Ð and ¯ame front conditional velocity. in terms of velocity ratio u H /SL. turbulent ¯uxes of the progress variable c. (212) overestimates the g turbulent ¯ux u HH c HH (see Fig. due to the ¯ame thickness that must be resolved in the numerical simulation. Following Bidaux and Bray (1994) (unpublished work already presented in the BML model context Section g 7. the Bray±Moss±Libby model remains able to predict the turbulent ¯ux type (gradient or counter-gradient). Velocities are made dimensionless using the laminar ¯ame speed SL [140]. Three. Our analysis is based on the two limiting cases pictured in Fig. 8. This phenomenon has already been discussed [65] and included in the ITNFS model [127]. Veynante. burnt gases conditional velocity.

Also plotted are the DNS conditions of the Rutland (CR. Then. (213) becomes: ~ ku HH ls cp 2 ctSL 214 ~ ² At the leading edge of the turbulent ¯ame (near c 0). and length-scale ratio. Premixed turbulent combustion diagram. value is close to the one obtained in a plane and laminar ¯ame ub 2 uu < tSL : Eq. the ¯ame front motions are assumed to be dominated by the turbulence properties taken upstream of the ¯ame. Two limiting cases: counter-gradient transport promoted by thermal expansion (left). ² At the trailing edge of the ¯ame brush c < 1. In 2D DNS. The Classical Klimov±Williams criterion and the criterion from Ref. lt/d L. Veynante. the ¯ame ~ front is convected by turbulent motions towards the burnt Fig. gradient transport due to turbulent motions (right). ub 2 uu < 2u H 215 High turbulence level: Due to strong viscous dissipation of turbulent eddies in the hot burnt gas. . NB . As the turbulence is decaying in the CTR simulation. tsL =2au H 1 (Eq. u H /SL.240 D. 32. where u H denotes the rms velocity in the fresh gases. The transition criterion. CTR conditions are displayed as an almost vertical line. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. The symbols W t 3 and A t 6 correspond to the CRCT DNS. 31. the turbulence decay is smaller and is not represented. 32). the ¯ame front is convected towards the fresh gas with a mean velocity estimated by 2u H (see Fig. L. The DNS ¯ame-¯ow conditions are plotted as a function of the velocity ratio. separating CGD (below) from GD (above) is plotted for t 3 and t 6 [140]. Filled (open) symbols denote gradient (counter-gradient) turbulent diffusion. (220)). t 2:3) and CTR t 3 simulations. [65] are given to show the domain of validity of ¯amelet combustion.

the thermal expansion due to heat release dominates and the ¯ame is able to impose its own dynamics leading to a counter-gradient turbulent transport.1. modeling ku ls as a sum of these two contributions leads to: ~ ku HH ls cp 2 c tSL 2 2au H and the turbulent ¯ux becomes: g ~ ~ u HH c HH c 1 2 c tSL 2 2au H 219 218 HH 216 and expression (219) may be used to derive a criterion delineating gradient and counter-gradient regimes. balance equations for turbulent scalar ¯uxes) to stagnating ¯ames in the limit of small turbulence intensities [131]. the possible occurrence of countergradient transport is generally neglected in modeling. Then. 33). ² The mean progress variable gradient may be estimated as: ~ ~ ~ c 1 2 c 2c < db 2x 221 This simple model is well veri®ed in direct numerical simulations [140] and has also been recovered when applying a second order modeling (i. the gradient type contribution in Eq. or close to the transition between gradient and countergradient regimes. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 241 Fig. The factor 2 has been introduced assuming cp < 0:5: Then. in practical applications turbulent transport may be counter-gradient. the turbulent transport is analyzed as follows: for a suf®ciently high turbulence level. (21)) as a function of the length scale ratio lt/d l comparing the integral turbulent length scale and the laminar ¯ame thickness [140]. a is expected to be of order unity for large turbulent length scales and vanishes when turbulent eddies are too small to affect the ¯ame front. 31. the heat release factor t is about 5±7 leading to a transition between gradient and counter-gradient situations when u H =SL is of the order of 3. as shown in Fig. where combustion instabilities occur and wrinkle the ¯ame front. Nevertheless. similar to the ITNFS model [127]. L. the ¯ame is unable to impose its own dynamics to the ¯ow ®eld and the turbulent transport is of the gradient type.D. Veynante.149] have reported regimes. (219) corresponds to a Prandt±Kolmogorov turbulence modeling: p 2c ~ ~ ~ 2c 1 2 cdb au H < 2adb k 2x 222 p where k is the turbulent kinetic energy and u H k. This criterion is well veri®ed by direct numerical simulation results. 8. The turbulent ¯ux may be viewed as the sum of two contributions acting in opposite directions. Recent experimental results [147] have con®rmed these ®ndings. Counterg gradient turbulent diffusion occurs when u HH c HH is positive where a length scale d b characterizing the ¯ame brush is introduced. DNS estimated ef®ciency function a (Eq.3. as for any inert scalar. The ef®ciency function a has been estimated from DNS (Fig. Comments ² Following Fig. 31 and criterion (220). corresponding to low turbulence levels. gases with a mean speed estimated by 1u H (see Fig. one induced by turbulent motions and the other by thermal expansion. 33. Then. 32): ub 2 uu < 1u H leading to the simple linear model: ~ ku HH ls 22 cp 2 cau H 217 where a is an ef®ciency function. ² Recent works [148. acting in . In many cases. On the contrary. The Bray number: NB tSL 2au H 220 is greater (respectively lower) than unity when a countergradient (respectively gradient) turbulent transport is expected.e. when the turbulence level remains low. accounting for the weak ability of small scale turbulent motions to wrinkle and convect the ¯ame front.

0 . ² A corrected Bray criterion (Eq. such a simpli®ed formulation is unable to predict a gradient transport in ~ the vicinity of the leading edge c < 0 always observed.4.5.0 g v HH c HH G G . ² Expression (219) has been derived to analyze the occurrence of counter-gradient transport. 34). 0) 2y . External pressure gradient effects The previous analysis is completed by investigating the effects of externally imposed pressure gradients on turbulent premixed ¯ames. In Refs.0 g u HH c HH G CG vb 2 vu (when y .e. 34 and 35. it is experimentally shown that counter-gradient transports are enhanced in ducted ¯ames. 36. (b) favorable imposed pressure gradient± counter-gradient turbulent transport. L.151] by differential buoyancy effects of pressure gradients on pockets of heavy and cold fresh gases and on pockets of light and hot burnt gases. even in counter-gradient situations. the same type of DNS as [140] was conducted. [20]. a gradient and a counter-gradient turbulent diffusion Zone 1 2 ub 2 uu . [121]): in the ®rst part of the chamber (region 1). and the turbulent ¯ame speed ST (Fig. In many combustion systems. the Fig. These instabilities should also be included in combustion models [150]. a pressure decrease from unburnt to burnt gases. [148.0 ~ 2c 2x . The thermal expansion modi®es the ¯ame dynamics as is clearly apparent on high speed tomography ®lms (scheme in Fig. an opposite direction than counter-gradient diffusion reducing the ¯ame front wrinkling. and enhance classical gradient turbulent transport. [153]. the turbulent ¯ame speed is modi®ed by a buoyancy term linearly dependent on both the imposed pressure gradient and the integral length scale lt. For instance. As proposed in Ref.0 ~ 2c (when y .0 . Counter-gradient transportÐexperimental results The V-shape turbulent ¯ame was described in Section 7. 34. 36). i. respectively. ¯ames are ducted and submitted to strong pressure gradients due to thermal expansion and leading to ¯ow accelerations. Counter-gradient turbulent transport was ®rst explained [19. introducing an externally imposed pressure gradient (in fact.0 . The main conclusions. 8.1. see [152]). (a) No imposed pressure gradient±gradient turbulent transport. A planar laminar ¯ame separating fresh gases (left) from burnt gases is initially superimposed to an homogeneous and isotropic turbulent ¯ow ®eld u H0 =SL 5 [152]. 35). All these results suggest that counter-gradient turbulent transport should be expected in most ducted turbulent ¯ames and some experimental observations are now reported. ² On the other hand.4. In Ref. displayed in Figs. 0) . 35). due to technical reasons. but is not suited for numerical simulations. the ¯ame brush thickness.0 .242 D. We do not give more details because this modi®ed criterion needs to be validated and improved (length scale effects are not so clear and various analyses are possible. It also promotes counter-gradient turbulent transport. Veynante. is found to decrease the ¯ame wrinkling (see Fig. Table 6 Geometrical analysis of the scalar turbulent transport in a V-shaped premixed ¯ame (see Fig. (220)) has been proposed to account for the pressure gradient effects. see also Ref. Superimposed instantaneous temperature and vorticity ®elds at the same time. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 8. G and CG denote. because of the rod wake. are: ² A favorable pressure gradient. a constant acceleration).152]. adverse pressure gradients tend to increase the ¯ame brush thickness and turbulent ¯ame speed (Fig.

-). Reduced turbulent ¯ame speed ST/SL plotted as a function of the reduced time t=tf SL t=dL . (a) initial turbulent level u H0 =SL 5. no pressure gradient (Ð) and increasing favorable pressure gradient (´´´ and . To include counter-gradient turbulent transport in modeling Except for the work of Bray. Libby and their co-workers [19. The approach of Bray et al. but the change in rotation of the coherent structures corresponds to a transition between gradient and counter-gradient transport for g the downstream turbulent ¯ux u HH c HH : The turbulent ¯uxes are of gradient type just behind the rod.129. very few works have been devoted to the effective modeling of countergradient transport.D. .. 8. except for their symmetry due to pressure waves. is mainly based on second order modeling. coherent structures embedding the ¯ame front turn clockwise (counter-clockwise) in the upper (lower) ¯ame sheet. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 243 Fig. Moss.-). 35. using balance equations g for the turbulent ¯uxes u HH c HH : These equations are easily i Fig. L. Veynante. Markers correspond to the Libby theory [152. 36. no pressure gradient (Ð) and increasing adverse pressure gradient (´´´ and . the upper (lower) coherent structures start to turn counter-clockwise (clockwise). Â Â as in classical Von Karman vortex streets. because of thermal expansion in burnt gases (region 2). When the centerline velocity increases. (b) initial turbulent level u H0 =SL 2. for different values of the externally imposed pressure gradient. as expected to ensure the stabilization of the ¯ame.. Accordingly. the transverse turbulent ¯ux g v HH c HH is always of gradient type. and becomes of counter-gradient type further downstream.151].6. This phenomenon may be recast in terms of turbulent transport using a simple geometrical analysis based on the Bray±Moss±Libby relation (212) and summarized in Table 6. where tf dL =SL is a ¯ame time.132.128.153]. A rough scheme of the coherent structures dynamics observed using laser tomography [121].

8. ² Mean progress variable gradient terms (IV) needs g Reynolds stress modeling for u HH u HH : Then. model closures and implementation dif®culties. ´ ´´ and y 12 mm. Some comments may be made: g ² The u HH c HH turbulent transport (III) is generally modeled i using a classical gradient expression (gradient turbulent transport at the third order). however this assumption is not supported by DNS results [152]. 37. very promising results were obtained. Axial velocity u histograms for four transverse locations 80 mm downstream the rod in the V-shape turbulent ¯ame (Fig.244 D. and closed. [128]. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 derived from momentum and progress variable balance equations: 2r uj u HH c HH ~ g 2r u HH c HH g i i 1 2t 2xj |{z} | {z } I II ² The mean pressure gradient term (VI) is easily known under a BML assumption (Section 7. new balance i j equations are derived. This balance equation is.151]. reaction zone (y 5 mm. unburnt gases (y 18 mm. u HH c HH . ² The mean velocity gradient term (V) is closed because g turbulent ¯uxes u HH c HH are provided by their balance j equations. ² The ¯uctuating pressure term (VII) is more dif®cult to understand and to model.3). this term is neglected. the action of mean (VI) and ¯uctuating (VII) pressure gradients. Recently. -´-).7. Ð ). c-gradient effects (IV). [151]. Veynante. [128. to turbulent ~ transport of r u HH c HH (III). The mean pressure term (VI) and the ¯uctuating pressure term (VII) probably need to be modeled together as c HH 2p=2xi : In Ref. and six for i HH u gHH ). products and thin ¯amelets was proposed [132]. particularly in predicting experimental results [20]. a model based on a partitioning of each pressure ¯uctuation covariance into contributions from reactants. The comparison of this new model with DNS results are encouraging and con®rm the importance of the intermittency between the conditional mean pressure gradients in reactants and products. mean velg i ocity gradient effects (V).152]. Making use of Eq. Fig. 19): burnt gases (y 0 mm. ² Turbulent dissipation terms (VIII and IX) are generally expressed together using small scale dissipation rate assumptions [151]. for these quantities (second order turbulence model). Towards a conditional turbulence modeling? The BML formulation (Section 7.132.3. r u HH c HH turbulent dissipation g i (VIII and IX) and reaction rate (X). L. very few simulations have been conducted using the second order formulation. respectively. unclosed and each term may be extracted from direct numerical simulations [132. the pressure gradient term closure is carefully discussed.5) directly distinguishes between the properties of fresh and burnt gases. (165) yields: ~ c HH c 2 c t ~ ~ c 1 2 c ~ 1 1 tc 224 2 2r u HH u HH c HH f ~ ~ 2c 2p j i g 2u 2 r u HH u HH gj 2 ru HH c HH i 2 c HH i j 2xj 2xi 2xj 2xj |{z} |{z} | {z } | {z } III IV V VI 2p H 2Jk HH 2tik 2 c HH 2 u HH c 1 r u HH v g i i _ |{z} 2xi 2x 2xk |{z} | } |{z} {zk X VII VIII IX (223) where the RHS terms correspond. . The discussion of the closure schemes for this equation is beyond the scope of the present review and the reader may ®nd relevant information in Refs. Because of very high computaReynolds stresses.140. of course. In Ref. Ð) [121]. u i j tional costs. Recently [129]. Second order closures for turbulent ¯uxes and Reynolds stresses require nine additional balance equations in 3D simug lations (three for progress variable ¯uxes.

Yn Single-step or complex chemistry Chemical source is closed (only for pdf methods) this was achieved using conditional averaging (Eq. (170) and (172)). Veynante. [55]) Major assumptions In®nitely fast chemistry Chemical equilibrium 1D strained reaction zones Diffusion balances chemistry as in a 1D counter-¯ow ¯ame Single-step or complex chemistry Constant Le i 1 or complex transport properties Lei ± 1 Diffusion is captured via micromixing modelling Flame structure in mixture fraction space Species are function of Z: Yi Yi Z p [35] Species are functions of Z and x : Yi Yi Z p . The velocity distribution is almost Gaussian in the unburnt and the burnt gases.148. according to Eq. L. tt PaSR _ Solve for Monte-Carlo particles: ~ dYik =dt Y i 2 Yik =tt 1 v k _i p ~ p to get P Y1 . This approach is very attractive because turbulence characteristics may differ in fresh and in burnt gases as shown in Fig. (168)).191] _ v i k Z v PaSR Z p .103] ~ ~ Y i 1 Y C Z p P Z p dZ p 0 i Presumed pdf 1 PaSR modelling: MIL/PEUL [68. an integrated amount of heat release. x Du7Zu 2 its scalar dissipation rate. ¼. tp dZ p dtt _i t t Pdf methods: p p p ~ ~ p Y i Y1 ¼ Yn Yip P Y1 . This bimodal distribution does not correspond to the assumptions made in the derivation of most turbulence models. ¼. PaSR stands for partially stirred reactor. (172).1. Z is the mixture fraction. but becomes clearly bimodal in the reaction zone denoting an intermittency between fresh and burnt gases. x st dj 21 _ j coordinate across the ¯amelet [88] Turbulent combustion model Presumed Pdf: f ~ ~ P Z p from Z and Z HH2 ~ Y i Z Yi Z p P Z p dZ p Flamelet modeling: ~ ~ Y i Z x Yi Z p . A conditional approach to determine the conditional averages of a quantity Q in fresh Qu and burnt Qb gases is probably a promising way leading to a straightforward description of turbulent transport (Eqs. P denotes pdfs. Introduction Much work has been devoted to the numerical modeling of non-premixed combustion systems. Reynolds averaged models for non-premixed turbulent combustion 9.154]. where tt < k=1 is an eddy break-up mixing time. xp P Z p . The objective is then to derive balance equations for quantities such as: ~ rcQ r cQb . tp k P Z p . a conditional moment. r 1 2 cQ r 1 2 cQu ~ which is easy from c and Q balance equations [141]. but no clear conclusions can be found in the literature. 37. Yn dY1 ¼dYn [93] Simple transport properties Lei 1 Solve for representative PaSR: ~ ~ dYi =dZ Y i 2 Yi 1 tt v PaSR = Z 2 Z _i to get v i Z. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 245 Table 7 Ç ~ Modeling strategies for non-premixed turbulent ¯ames. mainly assuming a chemistry much faster than mixing and molecular diffusion. V . This `mixed is burnt' regime is easily described from the . S . such as k± 1 . All these works devoted to turbulent transport suggest that turbulent combustion modeling might probably be greatly improved by advancing the description of turbulent transport itself. This point motivates large eddy simulation (LES) for turbulent combustion modeling [155].1) are displayed.4. where velocity histograms obtained using laser Doppler velocimetry in the V-shape turbulent ¯ame (Section 7. ~ turbulent micromixing was expressed with a linear relaxation (IEM-LMSE) D7 2 Yi < Y i 2 Yi =tt . These equations remain to be closed and a few attempts have been conducted in this direction [91.D. 9. other formulation exist [96] (from Ref. xp t Steady [63] or unsteady [177] Total heat release is a function of x : _ V x st 1 1 v T j. the density of ¯ame surface and Y iC . xp dZ p dxp [165] Coherent ¯ame model: Solve for the density of ¯ame surface S _ v T V x S [84] _ Solve for conditional moments: Y iC Z p Yi uZ Z p Conditional moment closure: presumed pdf P Z p [102.

Fuel/air mixing modeling 9. the premixing of the reactants is not always complete at the molecular level and some partial premixing may be observed. pdf method. A simple description of turbulent mixing is thus obtained f ~ from the two ®elds: Z and Z HH2 : Introducing the classical gradient transport closure for the turbulent ¯uxes (Section 6). followed by non-premixed and partially premixed combustion. and as (II) and (III). Diffusion is then addressed using turbulent micromixing modeling. 9. (41)): 2 3 HH HH 2r u j x ~ ~ 2r x ~ 2 HH 2Z 2Z 1 1 rDu j 2t 2xj 2xi 2xi 2xj 22 2 3 2 3 ~ ~ 2 2Z HH 2Z 2Z 2Z HH 2u HH j rDu HH 22 rD j 2xj 2xi 2xi 2xi 2xj 2xi |{z} |{z} I II 22 3 ~ ~ 2uj 2Z 2Z HH 2u j 2Z HH 2Z HH rD 22 rD 2xj 2xi 2xi 2xi 2xi 2xj |{z} |{z } III IV 2 3 HH 2 2u HH 2Z HH 2Z HH j 2 2 rD 2xi 2xi 2xj |{z} V 2 3 2 3 2 2Z HH 2 2Z HH 2 2 D2 r 2xj 2xi 2xj 2xi |{z} VI 2 2 3 2 2Z 2r 2 2Z D rD 2xj r 2xi 2xi 2xj |{z} VII (227) The term (I) on the RHS describes curvature effects of the mean mixture fraction ®eld. Sometimes. Depending on the simpli®cations made for these mechanisms. L. Veynante. As for premixed combustion (Sections 7 and 8).4. r x rDu7Zu 2 was discussed for ~ premixed turbulent ¯ames. ~ 9.1. ² Even in burners operating in the premixed regime.4). (VI) is the dissipation rate of the scalar . various approaches for laminar ¯ames and models for turbulent ¯ames are obtained (see Table 7). the structure and the properties of the ¯ame depend f on Z HH2 . BML and S models (Section 7). the second in Eq. implies the accurate determination of the position in the ¯ow where combustion starts and the control of pollutants emission. ² Finite rate chemistry assuming a local diffusive±reactive budget similar to the one observed in laminar ¯ames (¯amelet assumption). (226) is the production of ¯uctuations by ~ the mean gradients. in the EBU. Section 6). and more generally for gas turbines operating in the non-premixed regime. contains correlations between the Z ®eld and the velocity ®eld.246 D. these correlations vanish for suf®ciently large Reynolds numbers. There exist strong motivations for improving non-premixed and partially premixed turbulent combustion modeling: ² The development of new combustion technologies for aircraft engines. Introduction ~ The mean value of the mixture fraction Z gives an indication of the local mean fuel/air mixing in turbulent ¯ows. Hypotheses formulated to construct models for non-premixed turbulent ¯ames may be organized into three major groups: ² Assumption of in®nitely fast chemistry (mixed is burnt). The proposed closures are all based on a particular description of turbulent mixing. while chemical sources can be dealt with in an exact and closed form (only for pdf). Balance equation and simple relaxation model for x ~ A transport equation may be derived for x from the Z balance equation (Eq. (V) is the strain rate of the scalar Z ®eld. measuring the degree of mixing between reactants. the last is the scalar dissipation rate x remaining unclosed. the modeling of turbulent diffusion ¯ames relies on simplifying assumptions for both chemistry and transport.2.2.2) and mixing problems have been the subject of many discussions [156±158]. In addition. already discussed for ¯ame surface density modeling in premixed ¯ames (Section 7. Compared to other terms.2. partial premixing is even desirable to limit pollutant emissions (strati®ed charge engines). Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 turbulent mixing of conserved scalars (Section 3. hence the basic concepts useful to capture fuel/air turbulent mixing are ®rst presented. (IV) corresponds to the correlations between mean ¯ow motion and ¯uctuations.2. ² Finite rate chemistry with treatment of molecular and heat transport separated from chemical reaction (CMC. one may write: r ~ 2Z ~ 1 r u´7Z rn t 7 2 Z ~ ~ 2t 225 r f 2Z HH2 f ~ 1 r u´7Z HH 2 rnt 7 2 Z HH2 1 2rnt u7Zu 2 2 2r x ~ g ~ |{z} | {z } 2t Production Dissipation 226 The ®rst term on the RHS is the turbulent transport. Crucial points which cannot be addressed using the in®nitely fast chemistry hypothesis. ² Many practical systems include liquid injection of the fuel.

(231) is dif®cult to justify.D. IFCM is an interesting ®rst guess to provide the global ¯ame structure for the maximum of heat that can be released. Veynante.2). a valuable result for certain aspects of design. Table 1): IFCM IFCM YF YF Z. It is very popular and is usually coupled with a low Mach number solution of the turbulent ¯ow [161]. Eddy dissipation model The eddy dissipation model (EDC) is a direct extension to non-premixed ¯ame of the eddy break up (EBU) closure. 1=tk . tdZ p 233 0 ~O Y IFCM ~ T IFCM 1 0 IFCM ~ YO Z p P Z p . for very large . Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 247 dissipation rate and (VII) is negligible when density gradients are kept small. f ~ The knowledge of Z and Z HH2 from their closed transport ~ equation may be used to presume the pdf of Z. x. for instance. The mean ¯ame structure is then known from an in®nitely fast chemistry model (IFCM) without solving balance equations for mean thermochemical quantities. for large a . (231). x. then. Note that the piecewise relations are exact when the chemical time is in®nitely small. this species may be the reaction products. this model does not respect the response of diffusion combustion in mixture fraction space and may generate mean fuel mass fraction values ~F lower than Y IFCM : Actually. (229) is a very simpli®ed description of micromixing. L. Because the internal structure of diffusion ¯ames is easily obtained from condi tional statistics as Yi uZ Z p (Section 3. Rl 2xi 2xi 2xj k=1 tk where Rl is a Reynolds number based on the Taylor ~ microscale. [159]. Then. x. Eq. i. the linear relaxation is thus a ®rst approximation that may be improved. x. with a direct depenf dence on Z HH2 21 is anticipated for the dissipation rate of the scalar dissipation rate: 2rD2 2 2xj 2 2Z HH 2xi 3 2 2xj 2 2Z HH 2xi 3 .4.e. Considering the piecewise relations (Fig. so that the thermodynamics of the ¯ow is fully known from Eq. Eq. t dZ p |{z} p Flame structure Mixing 230 The simplest of these models invokes the in®nitely fast chemistry assumption. To account for the existence of burnt gases bringing the energy to ignite the fresh reactants. The fuel burning rate is calculated according to: 2 3 ~ ~ 1 YO YP ~ .2. . Rl rx2 ~ f Z HH2 232 The equilibrium condition (Eq. 1=n1=2 : The remaining part of the j ~ strain rate term is proportional to x .3. YO YO Z. 9. tdZ p z }| { ~ Yi uZ Z P Z p . t 1 0 relating the fuel and oxidizer mass fractions. Therefore. [35] which assumed an in®nitely fast single step chemical reaction (Section 3. A quadratic behavior in x .3. with a b -shape (see Section 6. Presumed pdf: in®nitely fast chemistry model One of the ®rst descriptions of non-premixed combustion was given in Ref.2). x. P Z p . _ k s 1 1 s where a and b are adjustable parameters of the closure. T T IFCM Z 228 the ¯uctuating velocity gradient is assumed to be of the order of the inverse of the small scales characteristic time.1. To derive a linear relaxation closure for the scalar dissipation rate. many models (¯amelet.3. mean quantities may be directly obtained with a b -pdf (Section 6. IFCM is therefore a two-equations model for non-premixed turbulent combustion. a homogeneous Z steady ®eld in equilibrium (production dissipation) is ®rst considered. 22rD 2u HH 2Z HH 2Z HH rx ~ rx ~ j . A priori. In Eq.2). 5. (230): 1 IFCM ~F ~ Y IFCM YF Z p P Z p . 2u HH =2xi . tdZ p 234 235 1 0 ~ T IFCM Z p P Z p . t.b 231 rv F ar min Y F . the reaction rate is limited by a de®cient species. The scalar dissipation rate depends on the detail of the characteristics of the turbulence (velocity and scalar energetic spectrum). The inputs of this f ~ `mixed is burnt' closure are Z and Z HH2 determining the b pdf chosen to capture the detail of fuel/air mixing. (228)) leads to: ~ x f Z HH2 k=1 229 recovering the widely used linear relaxation model.2) and Eq.2).4. by solving a modeled balanced equation for x : Such a closed equation may be ~ found in Ref. Eq. CMC) incorporate a b -pdf to calculate means quantities: ~ Y i x. and the temperature to the mixture fraction in the case of in®nitely fast chemistry. (227) reduces to a balance between the straining rate of Z and the dissipation rate of the scalar dissipation rate: 22rD 2u HH 2Z HH 2Z HH 2 j 2D2 r 2xi 2xi 2xj 2xj 2 2Z HH 2xi 3 2 2xj 2 2Z HH 2xi 3 9. Models assuming in®nitely fast chemistry 9. initially devoted to turbulent premixed combustion [160] (Section 7. (235).

Flamelet models are therefore constructed from an asymptotic view of diffusive-reaction layers as given by Fig. but when the joint pdf P Z p .4. in®nitely fast chemistry may be replaced by a chemical equilibrium condition [162]. 38).2). x. In these measurements and calculations. IFCM is an accurate description of a turbulent diffusion ¯ame. t is . Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. [164]. ¯amelet models are derived assuming that the local balance between diffusion and reaction is similar to the one found in a prototype laminar ¯ame for the same values of Z and x . quenching or even small ®nite rate chemistry effects exist. To handle multi-step chemistry. [163] in H2/argon±air ¯ame.165]. 9. but not in®nitely fast (see for example [163. (b) Raman scattering measurements of Barlow et al. The two control parameters of planar and steady laminar strained ¯ames are used: the mixture fraction Z and its scalar dissipation rate x (Section 3. For a given state of mixing in the turbulent ¯ow. 6. È Damkohler numbers (Section 3. thus given values of Z and x .1. these two quantities ¯uctuate in ~ space and time.4. Solid lines indicate chemical equilibrium. Introduction Experiments in jets ¯ames and direct numerical simulations suggest that there exist situations in burners where the chemistry is fast. Models have been proposed for such ¯ames [63. the response of the ¯ame in mixture fraction space lies in the vicinity of the curves given by YiIFCM (Fig.248 D. L. 38. xp . Scatter plots of major species mole fractions and temperature as functions of mixture fraction: (a) Data from three-dimensional direct numerical simulations of Montgomery et al. This model lacks any prediction capacities when ignition. Unfortunately such a zero order model does not exist for premixed turbulent combustion. Flamelet modeling 9. Veynante.164]). In a turbulent ¯ow.2).

x.e. Z HH2 ~ to which x is added. YiSLFM Z p . xp . then. Consider the simple case of a jet ¯ame. since it uses the same formalism. extinction may not be completed.4. on the air side. Two issues emerge: 1. for a given value of x corresponding to local micromixing conditions. As shown in Ref. and various x provides a Flamelet Library Yi(Z. turbulent diffusion mixes these products with air. on the fuel side. SLFM stands for steady laminar ¯amelet model. a trend that is not fully reproduced by the ¯amelet solution in physical space where the input parameter is the strain rate [171]. YiSLFM Z p . ² There are other issues related to the multi-dimensional character of diffusion ¯ames. it is not obvious how to draw conclusions since one would need to measure the scalar dissipation rate in turbulent ¯ames. one supposes that the mixing ®eld may be reduced to a steady one-dimensional structure. This transverse loss or gain of heat modi®es the structure of the ¯amelet in the normal direction to the stoichiometric surface [55]. when the Damkohler number Dap tc xst 21 is large. Moving downstream. Z. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 249 known. IFCM is recovered. choosing a given laminar ¯ame prototype. the mean properties of the ¯ame may be calculated as: ~ ~ YiSLFM Z p . Veynante. Z HH2 and x ). t must be presumed using the mean values ~ f ~ available to quantify fuel/air mixing (i. Time-dependent ¯amelets have been used to include unsteady effects [172±174]. The hypotheses involved in SLFM can be disputed. xp is the local ¯ame structure in mixture fraction ~ space and P Z p .2): 2 2 3 2Yi x 2 Yi vi 1 _ 2t Lei 2Z 2 2 2 3 N hn v n 2T _ 2 T 2 1x 2t Cp 2Z 2 n1 where x l=rCp u7Zu 2 : These equations were used to discuss laminar diffusion ¯ames (see Eqs.2. (237) has been obtained neglecting diffusion in the direction tangential to the iso-Zst surface. with a limiting frequency at which the ¯ame almost behaves like a steady state ¯ame.D.x ). xp YiSLFM Z p .167]. For a given chemistry. leading to ¯amelet interactions when the distribution of x is nonuniform on the iso-Zst. x. ~ f The inputs of SLFM are similar to those of IFCM: Z. xp features mixing without reaction (Fig. xp must be determined and tabulated under particular hypothesis. 4 [63]). or even quenching when x becomes too large (Fig. and with fuel. Further downstream. This situation is strongly enhanced in ¯ows where recirculation zones are found to stabilize combustion. This model may be viewed as a direct improvement of the in®nitely fast chemistry model (IFCM). Flame structure in composition space. the gradients measured along the stoichiometric surface are much smaller than those in the perpendicular direction (Section 3. one has to solve for: v 2 _ x 22 Yi Lei 2Z 2 237 The solution of this equation for given concentrations and temperatures boundary conditions. 9.2). a ¯ame con®guration widely studied experimentally [166. Thus. xp . It is usually observed that the decay of OH towards equilibrium is predicted by the ¯amelet solution in mixture fraction space using the scalar dissipation rate as a control parameter sensitive to species boundary conditions.171]. ² Reference states at in®nity used to tabulate the ¯amelets may have to account for partial premixing [170. and up to now. the validity of Eq. P Z p . but more works are required to conclude on the dimensionality of scalar micromixing in ¯ames. when a high value of the scalar dissipation rate is imposed to the ¯ame for a suf®ciently short period of time. YiSLFM Z p . Assuming within the turbulent ¯ow thin quasi-one-dimensional structures convected and stretched by the ¯uid motions. 6). ~ 2. (47) and (50) and Section 3. t captures the statistics of fuel/air mixing.2). x. xp . this shortcoming of SLFM is overcome introducing transient ¯amelets. the reactants feeding the reaction zone are not likely to be either pure fuel or pure air. where close to the nozzle inlet. [171]. when using Eq. arguing that when the mixing element is suf®ciently thin and features weak curvature. and neglecting higher order terms. [175]. the characteristic time required to balance diffusion and reaction is assumed to be much smaller than any . 5). Straining cannot be uniformly distributed along the ¯ame sheet. A variety of techniques are available to build these libraries [168]. In SLFM. Some experimental results are available [169]. thereby including ®nite rate chemistry effects. In Ref. L. An increase in x is followed by ®nite rate chemistry effects. In consequence. but with an additional parameter: the scalar dissipation rate x . (237) to describe a prototype ¯ame of turbulent ¯amelets. the equations for the species and temperature become (Section 3. xp P Z p . pure fuel and pure air react to form products. tdxp dZ p 236 Yi Zp other ¯ow time scale of the problem. leading to various improvements: ² Eq. Omitting the time derivative (steady ¯amelet). Unsteady ¯amelets were used [176] to simulate extinction and re-ignition xp YiSLFM Z p . ² Unsteadiness is also an important aspect. xp may be tabulated from solutions of counter-¯ow diffusion ¯ames (Fig. (237) in a turbulent ¯ow also depends on the properties of micromixing. for which reference states at in®nity vary according to the value ~ of a progress variable c: ² Flamelet libraries can be calculated in physical space or in mixture fraction space. and therefore a given chemiÈ cal time t c.

178±181]. These numbers are then È compared with Dap . Again using DNS. This work was pursued [177] introducing as an additional control parameter. This is more generally related to questions arising concerning the determination of quenching limits along with the accuracy of quenching predictions using ¯amelet theory a problem which can be addressed numerically. As expected for these ¯ames featuring strong unsteadiness effects. A critical Damkohler number at which extinction occurs is determined as a function of a ¯ame thickness parameter.250 D. Left. Below this value. de®ned as the ratio between the rms mixture fraction and the reaction zone width in mixture fraction space. bold: iso-reaction rate. Right. ² Another related point of interest in ¯amelet theory is the response of the turbulent ¯ame when quenching zones develop. Fig. triangle: bottom left 2D ¯ame. L. The value of x when the ¯ame extinguishes is in perfect agreement with x q measured in a laminar ¯amelet library . In Fig. even in the simple con®guration of ¯ame/vortex interaction. the quenching Damkohler derived q from asymptotic analysis [30]. Three successive times of two-dimensional unsteady quenching of a planar diffusion ¯ame pinched by a pair of vortices. This observation was explained by statistical variability. the diffusion time needed to exchange mass and energy over a distance DZ in mixture fraction space. Consequently. show that when strong unsteadiness and/or curvature effects appear. history effects were included using a Lagrangian time measured along the stoichiometric line. 39. q q Eq. q Other DNS results have shown that the scalar dissipation rate controlling the growth of the ¯ame hole is lower than the one that should be applied to ®rst quench the ¯ame [38]. used to construct a combustion diagram ([70]. are easily derived from the simulations to ®nd the limit conditions where unsteady (DaLFA) and extinction (Daext) effects become important. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. one may discuss the assumption that the occurrence of local extinction at some points does not prevent the use of a ¯amelet model for the remaining part of the turbulent ¯ame. Two critical Damkohler numbers. dotted: vorcity. in a turbulent jet ¯ame. Numerical simulations of ¯ame/vortex interaction. 39. constant density ¯ames near extinction were È studied [182]. In other words. an È extinction may be observed with Damkohler numbers larger or smaller than Dap : In this last study. for instance by post-processing DNS databases to study the reactive/diffusive layer in terms of ¯amelets [70. Veynante. con®guration studied in Ref. twice the asymptotic quenching value in these q simulations. the lower q value of Daext differs from that of steady ¯amelets and exceeds the value Dap given by ¯amelet theory.74. a DNS database is further analyzed [53] in which a diffusion ¯ame is pinched by a pair of vortices. the time evolution of the reaction zone is fully dominated by unsteady effects. diamond: top left 2D ¯ame. [53]. solid line: 1D ¯amelet library. the laminar ¯amelet assumption does not always predict quenching (Fig. square: middle left 2D ¯ame. zoom: region of the ¯ow where the centerline response of the ¯ame versus inverse mixture fraction dissipation rate is plotted. It is found that DaLFA =Dap < 2 while Daext =Dap < 0:4: Therefore. DaLFA and Daext. (237) is an interesting approximation for Da p larger than Dap . È 10). the value of the critical È extinction Damkohler number was different from the one predicted by laminar ¯ame theory. 10).

x. xp . t [63]: 1 1 ~ P xp . (Fig. the scalar dissipation rate xqEd at the quenching location is smaller than x q and varies depending on the structure of the edge ¯ame. These values are representative of the existence of burning in a premixed regime at the extremity of the diffusion ¯ame. Comparison between experiments and ¯amelet modeling [177]. xqEd decreases and a partially premixed front develops at the edge of the burning zone.4. ² When partially premixed ¯amelets are expected in the combustion system. tP xp . Therefore. since diffusion is directly included and coupled with chemistry in Eq. t: Most of the ¯amelet models explicitly suppose that the mixture fraction and its dissipation rate are two uncorrelated quantities: ~ ~ ~ P Z p . When the chemistry is fast enough with mixing elements featuring 238 erfc denotes the reciprocal of the complementary error function and x 0 is the maximum value of the scalar dissipation rate. t2 2s x. . x. that determines the equivalence ratio of the ¯amelet whose progress of reaction is given by a progress variable c. top). Mean and conditional scalar dissipation rate are then related by: ~ f ~ ~ x st x F Z. Veynante. 40) and further re®nements of ¯amelet models are under development [184]. Bottom: radial pro®les of mean temperature and OH mole fraction. L. x. t (see Section ~ 6. the turbulent micromixing of chemical species does not need any particular treatment. It is also worthwhile to note that with SLFM. But once a hole exists in the ¯ame. Top: Favre-averages mean values of temperature and OH mole fraction along the centerline of a jet ¯ame. Another alternative is to estimate a mean scalar dissipation rate under stoichiometric condition [89]. Starting from the solution of a steady strained planar counter-¯ow ¯ame: r j p 1 a r j p p Z h erfc h= 2. the fuel/air ~ turbulent mixing is captured via the joint pdf P Z p . 39. x. xp . Z HH2 where F is given by: ~ f F Z. t 2px s x. for x . models based on Eq. t p p exp 2 2 ln xp 2 g x. t < P Z p .4. the quenching scalar dissipation rate x q is the relevant quantity to describe the quenching of a burning ¯amelet. j the coordinate normal to the stoichiometric plane and the assumption r 2 D < cst r2 D0 was 0 used.2) and a log-normal distribution is used for P xp . The corresponding scalar dissipation rate is given by: a exp 22erfc21 2Z2 x Z 2p x0 exp 22erfc21 2Z2 x0 F Z 21 Fig. 40. (237). middle). Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 251 sheet-like properties. x. Bradley and coworkers have proposed ¯amelet closures along these lines where chemical sources are tabulated using one-dimensional premixed ¯ames [183]. The chemical sources are parametrized in terms of the mixture fraction Z. Z HH2 1 0 239 F Zst ~ F Z p P Z p dZ p 240 and to retain in the ¯amelet library the pro®les ~ ~ YiSLFM Z p . [63]. x. s 2 2 is proposed. Within the family of closures for non-premixed turbulent ¯ames. (236)). for instance at the base of a lifted turbulent jet-¯ame. x. Latter after quenching. where h j dj 2 D 0 r0 a is the strain rate. 39. 39.3. x st for averaging with P Z p (Eq. (237) represent great progress compared to the in®nitely fast chemistry hypothesis.D. it may be interesting in the modeling to use premixed ¯amelets instead of diffusion ¯amelets. Mixing modeling in SLFM In steady laminar ¯amelet models (SLFM). bottom). 9. then the reaction rate reaches values above the laminar ¯amelet behavior (Fig. The mean burning rates are then obtained by averaging the source terms with a presumed form for the joint-pdf of Z and c. When the diffusion ¯ame and its extremity have reaction zones of the same thickness. t The two parameters s and g are provided by the ®rst and second moments of the scalar dissipation rate: 2 3 s2 f ~ ~ and x exp g 1 x HH2 x 2 exp s 2 2 1 2 In Ref. they provide interesting results (Fig. The ¯uxes of heat along the stoichiometric surface are responsible for this departure from one-dimensional ¯amelet behavior. x q the response of the reaction rate follows the laminar ¯amelet one (Fig. but diffusion ¯ame quenching leads to edge ¯ame combustion (partially premixed ¯amelets). t ~ A b -function is assumed to presume P Z p .

In ¯ame surface density models. an exact but unclosed. The balance equation for the ¯ame surface density S was based on phenomenological considerations starting from balance equation for a material surface where combustion effects have been intuitively added [84.187. The integrated burning rate _ V i may also be expressed as a function of the species molecular ¯uxes. where x is the averaged scalar dissipation rate. V i is estimated as V i V i x . (53) in Section 3. one may write: DZF < 2Zst which provides an interesting approximation of this reaction thickness. The introduction of effects of unsteadiness was realized in Ref. jdj P xp dxp _ 21 | {z } 1 1 _ V i xp The reaction rate per unit of ¯ame area should. Conclusion The laminar ¯amelet model presumes the conditional mean of species rYi uZ Z p averaging over the response of laminar prototype ¯ames: ~ rYi uZ Z p r YiSLFM Z p .2). (229)): ~ x Cx f Z HH2 k=1 where Cx is a constant depending on the topological and spectral properties of the mixture fraction ®eld. identifying the ¯ame surface to the stoichiometric iso-surface Z Z st : The ¯ame surface density is then de®ned as: S u7Zud Z 2 Z st u7ZuuZ Z st P Zst 241 where P Zst denotes the probability of ®nding Z Zst : Starting from this de®nition and the balance equation for the mixture fraction Z (Eq. The mean burn_ ing rate v i is then expressed as the product of S by V i the _ local reaction rate per unit of ¯ame area. [172] in the context of SLFM.4.5. jP c xp uZ p dxp dj _ 246 x ! v i xp . therefore. 9. be estimated from the integrated reaction rate of a ¯amelet submitted to a scalar dissipation rate x with the distribution _ _ _ ~ ~ P xp : In practice.4. xp Pc xp uZ p dxp _ x 244 lead to: _ Vi 1 1 u7Zu 0 In our generic classi®cation of turbulent combustion modeling (Section 6) and Eq. under a ¯amelet assumption. SLFM may be viewed as a presumed pdf technique involving conditional mean values determined from laminar ¯ame solutions. 9. the stumbling block is the estimation of the ~ mean scalar dissipation rate x : The linear relaxation model may be written (Eq. This last expression was successful to model turbulent ¯ame lift-off [185]. using Eq. Another alternative is to build models where quantities are estimated introducing a direct treatment of micromixing and small scale diffusion (see Sections 9. According to Eq.2. The coherent ¯ame model. (121) ( vi _ Cx < 2. x v i Z p . (243): _ Vi < < 1 1 21 x v i xp . Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Here again. In diffusion ¯amelets.188] have provided an exact balance equation. xp P xp dxp xp vi _ u7Zu ) s _ u7Zu V i S 242 _ The local burning rate V i is estimated from a one-dimensional laminar ¯ame as: _ Vi ( vi _ u7Zu ) s < 1 v 1 1 _i v i dj _ dZ 0 u7Zu 21 243 where j is the coordinate along the normal to the ¯ame front.186]. L. More recently.8). coherent ¯ame model Flame surface density concepts were ®rstly introduced for the coherent ¯ame model (CFM) [84]. Veynante. a ¯ame time response is then intro_ duced when calculating V i: . S accounts for _ ¯ame turbulence interaction and V i for the chemistry. (230). xp Pc xp uZ p P Z p dxp dZ p 245 _ Then. (41)). then. Flame surface density modeling. The reaction zone is centered at the location Z Zst : As the stoichiometric surface is close to the air side for hydrocarbon ¯ame Zst p 1. (91) and the relation: v i uZ p _ v i Z p . it was proposed to choose [33]: ~ x Cx DZF 2 k=1 balance equation for the ¯ame surface density S may be derived (Section 6. Recent works [88. it is assumed that the interface between fuel and oxidizer in non-premixed burners behaves. where v i depends on the mixture _ fraction Z and the scalar dissipation x . corresponds to a simpli®ed version of the steady laminar ¯amelet model. for any values of Z. [186].3. expressions (96).3.252 D.6±9. see Eq.2). following the analysis proposed in Ref. as S Z p < u7Zu S (Section 6. DZF is the thickness of the reaction zone of these laminar ¯ame measured in mixture fraction space.

n=11=2 is the Kolmogorov time. t 1 1 1 0 tig Z p . x. t ig or x . Eq. t are known. x. (248) describes a local ¯ame element. 42 shows an example of numerical simulation of a lean premixed prevaporized (LPP) combustion chamber using MIL. t 1 1 tig Z. . The pdf of the mixture fraction P Z p . Since t ig depends on position in mixture fraction space and mean temperature. The transition between mixing without combustion and a `mixed is burnt' regime is controlled by the position of an ignition time delay. 41). t is determined so that Y MIL calculated along the MIL trajectory in mixture fraction space (Fig. Y F . may be observed in the premixed tube.x. Sketch of MIL in mixture fraction space. MIL model Á MIL (`Modele Intermittent Lagrangien' or Lagrangian Intermittent Model) was proposed by Borghi in Ref. The LPP system is divided into three parts: ®rst. weighted by ~ the b -pdf. x ig) prevent ignition and contribute to pure mixing. Z: This look-up table is easily derived from ~ partially stirred reactor (PaSR) calculations. t q 2 1 q21 tq k q 2 1 tt q exp 2 q tk ~F q x. For the highest temperature and pressure conditions. [68]. Veynante.x. tp P Z p . t ig. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 253 tion occurs only when ignition is faster than turbulent fuel/air mixing. like autoignition or ¯ashback. t is the pdf of the micromixing times having a mean value equal to the integral length scale time (k/1 ). within the distribution of fuel/air turbulent mixing times tk . L. liquid fuel is injected and vaporized (prevaporized tube). corresponding to t t . x. the mean burning rate of fuel is estimated as: ~ _ v F x. t: ~ 4. The pdf of turbulent mixing times P tt . x. corresponding to fast mixing. 41). is introduced to calculate _F v MIL Z p . For a given value of tig < x21 . x. x. Arguing that mixing occurs without reaction for large values of the scalar dissipation rate. t was proposed [189. x. reproduces the mean Y F : Fig. combining mixing and chemical effects. where molecular diffusion. 41. leading to: t p ~ dYF Y p 2 YF 1 tp v MIL t _F F ~ dZ Z 2 Z p 248 This technique is close to the one chosen in Ref. where T ~ and Y F are the mean conditions in the PaSR. tdtt Fig. some undesirable effects. the ¯ame structure is constructed dynamically from two possibilities of diffusion combustion: Mixing without reaction (before ignition) and in®nitely fast chemistry (after ignition) (Fig. tp [68]. The ignition delay t ig is associated to a scalar dissipation rate xig tig < x21 : The main objective of MIL is to account ig for unsteadiness in the coupling between small scale diffusion and chemistry. t (equivalent to the pdf of the inverse of x ). 2. where tk . x.t P tt .191]: t P tt . Whereas for x . Detailed chemistry can be used to construct such a t ig table [189]. tt . x ig. accounting for auto-ignition and transient burning in turbulent ¯ames. Combustion is expected to reduce pollutant emission in a lean premixed regime. ~ 3. D7 2YF. a is also a function of Z and of the local ¯ow conditions through posi ~ tion and time. then is mixed with air in the premixing tube before entering the main combustion chamber. 11.t ~ t v MIL Z p . is modeled as a linear relaxation term ~ Y F 2 YF =tt : Solving Eq. Assuming that combus- where P tt . 9. fuel mass fraction and mean temperature. Once P tt . t and P Z p . tp : _F t A generic shape for P tt . These ®ndings are reproduced with turbulent combustion closures like MIL describing strongly coupled micromixing and ignition phenomena [191]. tdtp dZ p _F t t 247 A partially stirred reactor (PaSR). [190] for modeling NOx production in turbulent jet ¯ames. The mean concentration of fuel Y F : ~ The output is the mean burning rate v F to calculate _ ~ new Y F values. tP tp .6. mixing is slower than chemistry and combustion develops. by means of the spectral distribution of micromixing times. mixing ig times smaller than t ig (t t . t ig. MIL incorporates the ¯ame structure in Z space. x. x. tig ~ ~ tig T.D. the probability to ®nd ignition at a particular location is: a Z. The inputs of MIL are: 1. The premixing tube operates for a large range of inlet temperatures (from 300 to 800 K) and a large range of pressures (from 1 to 40 bar). As in the steady laminar ¯amelet model (SLFM) and the conditional moment closure (CMC). A table of ignition delay times t ig depending on mixture fraction. (248) with boundary conditions similar to the ones of ¯amelet libraries provides v MIL Z p . x.

but as shown from the pdf balance equation (Eq. 1: Vaporized tube.7. the pdf of the mixture fraction was assigned a b -shape. the conditional average rxuZ Z p needs closure. In non-premixed ¯ames. 9. The ¯uctuations .103.8. ² Within a homogeneous ®eld.8. Instead of getting Qi rYi uZ Z p = ruZ Z p from laminar ¯amelets. Similarly. and therefore. The following points have been addressed: ² The approximation of the conditional chemical source ~ using Eq. This conditional mean is an important ingredient of the pdf transport equation and micromixing models are available (Section 9. or conditional moment closure (CMC). neglecting two terms (Eq.105]. (237)) replacing Qi by Yi [107]. These two turbulent combustion models have similarities in term of their general formalism. the CMC equation reduces to: 2Q 22 Q i i rxuZ Z p ruZ Z p 1 v i uZ Z p _ p2 2t 2Z 251 This last formulation may be compared with the ¯amelet equation (Eq. and MIL suggest that non-premixed turbulent ¯ames may be conveniently studied using conditional averaging in the mixture fraction space. Once the number M of conditional values required to capture the conditional means of the N chemical species is determined. the ¯amelet assumption deals with a coupled representation of diffusion and reaction as in a laminar ¯ame for a given value of x . the CMC approach proposes to solve a balance equation for Qi [102. this method is expected to provide interesting results for conditions not too far from in®nitely fast chemistry. (110)). Conditional moment closure In®nitely fast chemistry (IFCM). (114)): 2Q 2Q i i 2 rui uZ Z p ruZ Z p 2t 2xi |{z} I of species concentrations and temperature for a given level of mixture fraction may be included in the conditional burning rate by using second order CMC.4. Y N but implies a weak level of ¯uctuations for a given value of Z. that may not be correctly captured with a presumed pdf. YN uZ Z p _ _ 250 As for any combustion models. the conditional value of may be approxix ~ mated from P Z p to obtain rxuZ Z p [192]. L. 9. When calculating the chemical source. but their underlying physical assumptions strongly differ. the detail of rxuZ Z p is linked to the detail of the shape of the pdf. 42. the conditional pdf Pc Yi uZ p is approximated by a peak located at Yi uZ Z p : Accordingly. using in®nitely fast chemistry (IFCM).2). The objective of PDF modeling is to relax all hypotheses concerning the shape of pdfs. Once a methodology has been developed to calculate pdfs. CMC incorporates diffusion as turbulent micromixing and decouples it from chemistry. ¼. the scalar dissipation rate is unknown (II). Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. 3: ¯ame tube. when the pdf is known via its presumed b -shape (Section 6. BML and related closures assume a bimodal probability density function (pdf) of the progress variable. ² The estimation of rxuZ Z p from a b -pdf is feasible ~ for simple shapes of P Z p .1). Combustion starts in the premixed tube.254 D. CMC modeling requires the solution of N £ M balance equations. ¼. (250) is much less restrictive than to write v i < _ ~ _ ~ v i Y 1 . In the simplest version of CMC. modeled and solved [105]. a given thickness of the laminar diffusive zone. micromixing (II) and the chemical source (III) are unclosed. steady laminar ¯amelet (SLFM). This leads to a variety of modeling re®nements where more transport equations for conditional quantities are studied. On the other hand. which is written. Veynante. steady laminar ¯amelets (SLFM). 2: premixed tube. Snapshot of temperature in a TURBOMECA-ONERA LPP engine computed with MIL [191]. CMC has been the subject of many works and re®nements during the past years [105]. it is possible to construct turbulent combustion closures in which all the values taken by 22 Q i 1 rxuZ Z p 1 v i uZ Z p _ p2 2Z | {z } | {z } II III 249 In addition to turbulent transport included in (I). CMC was also extended to reacting particles [193]. With CMC. Pdf modeling In premixed turbulent combustion modeling. the ¯uctuations of chemical source in mixture fraction space are neglected leading to: v i uZ Z p v i Y1 uZ Z p .

Differential diffusion is easily introduced in SLFM when tabulating counter-¯ow ¯ames. Using pdfs and CMC. (Eq. these preferential diffusion effects may be more dif®cult to introduce in CMC. This term represents turbulent micromixing [95. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 255 species is resolved (Eq.199]. the impact of the molecular dissipation is less than the effect of viscous dissipation. Veynante. some works have been made in this direction [194. resulting from a discrete Monte Carlo simulation of the continuous pdf equation. t Y p 1 v p P Z p .195]. (111) in Section 6): r 2 ~ p P Z . 5 and 43). When the Schmidt number of Z is greater than unity (n . For those small scales (lZ . Two levels of complexity may be chosen: ² One may solve the balance equation for the joint pdf of the thermochemical variables.1. 43. 9.198. temperature.D. L. x. then the balance equation for the joint pdf of velocity and . The main advantage of the pdf method lies in the possibility of treating complex chemical sources directly (Section 6). suggesting that the cascade mechanism is limited by the mixing frequency fZ(lZ) of the largest scales. t 2t N i 2 h 2u´7Y i uY x.96].8. Two-dimensional pdf of mixture fraction and reaction progress variable for a round jet non-premixed ¯ame [104]. ² A new numerical treatment is required and leads to the decomposition of the pdf into a large number of computational particles required to succeed in the computational implementation of pdf methods [93. x. There are two major issues with pdf modeling: ² Models must be proposed for capturing micromixing and pressure ¯uctuation and viscous effects (the latter only when the velocity ®eld is incorporated into the pdf). then the smallest scales of Z are smaller than the Kolmogorov scales. the scalar dissipation rate appears as a key quantity in both premixed and nonpremixed turbulent systems. one needs to express the conditional mean value of the scalar dissipation rate (or the conditional mean value of diffusion). etc. The chemical composition and velocities of these particles evolve according to simple differential equations. species and temperature in mixture fraction space may be reached (Figs. Using turbulence modeling. Turbulent micromixing From the discussion of fundamentals properties of ¯ames and of turbulent combustion models. t Y p P Z p . As in ¯amelet modeling and CMC. A global picture of mixing should therefore distinguish between macro and micromixing [200]: The macromixing is organized from a cascade mechanism between large and small scales. Recent developments have also included time scale information into the pdf [196. t 2xk 2xk 252 ² Velocity ®eld statistics may also be included into the pdf. a mean mechanical time tt < k=1 and a microscale time tk < n=11=2 may be estimated. Di). t _i ~ 2Yip i1 N N 2 2 2Yip 2Yjp i1 j1 2r 42 D 2r 5 3 2Yi 2Yj ~ Y x. (113)).197]. species. x. such as the k± 1 closure. in the context of pdfs. The study of this cascade in a fully developed turbulence shows that the mixing frequency fZ is a function of the characteristic length lZ [201]: 1=2 2 32=3 1 hk fZ lZ . h k) the mixing frequency is of the order of a constant fZ lZ 1=n1=2 : This cascade process from large to small scale does not change the energy in the f system and therefore does not affect Z HH2 : The molecular diffusion acting at smaller scales makes Fig. lz n The speed at which Z is mixed increases when the characteristic length lZ decreases. a detailed description of chemistry may be incorporated in the modeling.

t 2t 0 254 ~ When P Z p .2. The mixture fraction is chosen as a shadow ®eld to estimate micromixing. t 2 h ~ 2 p D7 2 ZuZ Z p P Z p . this process starts at the Kolmogorov scale. 44 shows the conditional mean of the diffusive budget of the mixture fraction Z measured in a DNS of a mixing layer. nD2 =11=4 Sc21=2 hk All these approximations are given for non-reactive ¯ows to exactly quantify the effect of combustion is an arduous task. with Z The extension of this linear relaxation to conditional diffusion provides the linear mean square estimation (LMSE) or interaction by exchange with the mean (IEM) model. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. a b -pdf is used to reproduce the correct behavior of conditional diffusion. and a ®rst approximation is given by: 2 3 2 3 Z 2 Zp Z 2 Zp 2 ZuZ Z p < < 253 D7 tt k=1 This simple model cannot capture the time evolution of the pdf in homogeneous mixing problems and has many shortcomings. the scalar dissipation f ~ rate of the mixture fraction was expressed as x Z HH2 =t : t at Z 0 and Z 1). This dif®culty is also encountered in ¯amelet or ¯ame surface modeling. with a slope that can be approximated using the inverse of tt < k=1: Since the gradient of Z must relax to zero for Z 0 and Z 1. the gradients smoother and then induces the micromixing. Fig.204].8. this closure is also called a `Beta-Mapping-Closure'. x.2. this equation determines D7 2 ZuZ Z p . Various techniques have been proposed [95]. phenomenon required to put the reactants in contact with the reaction zone. x. 9. the Batchelor scale becomes the relevant limiting length: lB . Left: Scatter plot of D7 2Y. for instance it relaxes the ¯uctuations without modifying the shape of the pdf [94±96. A quasi-linear response is observed close to the mean. Veynante. Right: D7 2 YuY x. the behavior at the edges of the sample space is not of prime importance to capture the major properties of micromixing. When n D. GIEM model An improvement of the IEM model was recently discussed [205]. t Y p : These two terms were measured in DNS of an inert mixing layer [234]. this quasi-linear response is combined with the conditions D7 2 ZuZ 0 D7 2 ZuZ 1 0: For a `Gaussian-type pdf' (no peaks . x. when n ± D. 44. t 2t 2Z since D7 2 ZuZ 0 0.256 D. IEM was also developed and used in chemical engineering applications by Villermaux [203].8. In pdf modeling. but appears in a different form. Linear relaxation model. In a homogeneous ®eld.3. First proposed in the pdf context by Dopazo and O'Brien [202]. x.2). one may write: 2 p 3 1 2 Z ~ 1 1 2 ZuZ Z p 2 D7 P Z . t is given by a b -function parameterized f ~ and Z HH2 . x. the time evolution of the mixture fraction pdf is: i ~ 2P Z p . or mimic its impact on the pdf. the micromixing closure must estimate the term D7 2 Yi uY Y p describing the molecular diffusion of each chemical species. L. 9. tdZ ~ P Z p . IEM/LMSE In b -pdf modeling (Section 9.

that may be expressed in terms of a normal displacement speed wn and a curvature term D7´n. 9. Fig.208. (109) may be recast as: 2 3 1 7´ rD7ZuZ Z p r 2 3 1 2 p 2 rDu7Zuu7ZuuZ Z ru7Zu 2n 2 D7´nu7ZuuZ Z p ) ! 1 2 rDu7Zu 1kD7´nl s u7ZuuZ Z p ru7Zu 2n s 2kvn ls 1 kD7´nl s u7ZuuZ Z p 2 258 The molecular diffusion is decomposed into a normal contribution.205. Interlinks PDF/¯ame surface modeling The unclosed molecular diffusion term in the pdf balance Eq.95±97. but overcomes some of its drawbacks. (38). Their applications to ¯ames can be disputed easily since mixing described via stochastic processes does not always account for the presence of reaction zones.D.196. 46 illustrates the impact of micromixing modeling on temperature and CO concentrations predictions. tYi 2D7 2 Yi < 0 3 BYi t 2t C Z. and strong efforts have been made in this direction. much work has been done to improve these models which have then been quite successful in the calculation of jet ¯ames [97]. The dif®cult point lies in the choice of the transition probability T.5. (254) and (256). Z 11 3 Z p d Z p 2 2 257 256 where C(Z p. a full review is beyond the scope of this paper. For the same calculations. 45 shows pdfs of the temperature calculated from Eq. C(Z p. relaxing the constraint ~ BZ t Z of IEM. Fig. tP Z . Z 11 3 Z p) is the probability that a volume of concentration Z 1 interacts with a volume at Z 11 to evolve into Z p. Z 11 3 Z p dZ 1 dZ 11 2 P Z p meaning that the two volumes interact to generate a volume with Z p Z 1 1 Z 11 =2: Eq. Z Z p C Z p . It allows for reproducing the correct behavior at the edges Z p 0 and Z p 1 and its time dependence gives the correct behavior during the relaxation of the pdf. t) is a pseudo-mixing frequency depending both on the position in mixture fraction space and time.4. (110) [207] using the coalescence-dispersion model [206] and the k±1 closure for the velocity ®eld [208]. f As Z HH2 is an input of the closure. the reduced chemical scheme of Ref. using the same pseudo-mixing frequency (an assumption made in most micromixing models): D7 2 Yi uYi Yip .210±236]. However. t 2b ~ 1 ~ 11 P Z . t 2t tZ 3 ~ Â T Z 1 . Detailed discussion of mixing models in the context of pdfs may be found in the following list of references [93. These closures are clearly linked to a purely stochastic view of micromixing and are well suited to Monte Carlo simulations [93]. Many types of mixing models were proposed. The diffusion coef®cient D is assumed constant and kD7´nl s Dk7´nl s : The ¯ame surface averaged curvature may be modeled from geometrical considerations ( b is a parameter of the model and T(Z 1. When only diffusion is acting. BZ(t) denotes the point where the diffusive ¯ux is zero. t This attractive model conserves the simplicity of linear relaxation modeling. Multiple choices are possible and have been made (see Ref. The same treatment is applied to any reactive species Yi. the simplest one consists of writing: 2 3 Z 1 1 Z 11 T Z 1 . the generic form of these models is written [94]: 2 ~ 2P Z p . L. 9. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 257 The LMSE/IEM model leads to: 1 ~ D7 2 ZuZ Z p < Z 2 Z p tt Making a direct analogy GIEM proposes: D7 2 ZuZ Z p < C Z p .206.194. [95] for a . t) and BZ(t) are determined from Eqs. (257) was ®rst proposed by Curl [206] to simulate the production of droplets of a given diameter by coalescence of two droplets of various sizes and re-dispersion of the two droplets. Micromixing modeling is obviously one of the greatest challenges of turbulent combustion modeling.8. These closures are usually presented in the context of non-continuous interactions of volumes of ¯uid. This approach has similarities with the estimation of the conditional values of x in CMC using b -pdf [192] and is not a full closure of micromixing per se. t BZ t 2 Z p 255 review). Veynante.8. the mean scalar dissipation ~ rate x should be previously modeled. Stochastic micromixing closures Micromixing modeling may also be developed by choosing an ad hoc stochastic process mimicking the relaxation of pdfs due to small scale diffusion. t BYi t 2 Yip The parameter BYi t is then estimated from the fact that micromixing by itself should not change the mean: 2 3 2Yi C Z. [209] has been utilized with the LMSE (IEM) and Curl's coalescence re-dispersion formulation. according to Eq.104.

Joint velocity/concentrations pdf modeling When the joint pdf of velocity and species enters the problem. Ly. [88.258 D. Expressing the mixing time from linear relaxation hypof ~ thesis. Pdf calculations by Jones and Kakhi [208] using the Curl mixing model [206]. where u7Zu is expected to relax to zero. Veynante. . the need for modeling scalar dissipation rates. where u7ZuuZ Z p characterizes laminar ¯ame elements. the IEM model provides an interesting estimation of the micromixing term around Z Z p when the turbulent mixing time is correctly calculated [234] but becomes de®cient when Z p tends towards 0 or 1. [207] `L' jet ¯ame (location x=D 20). experiments have 2 In non-premixed turbulent ¯ames. L. when neglecting the normal displacement speed kwn ls (Section 9. tt < Z HH2 =x and using Eq.2): 2 2 3 1 Z 2 Zp 7´ rD7ZuZ Z p r tt 260 It was shown in (Section 9. Mean curvatures are supposed positive (i. evidencing. This is not the case in premixed combustion. convex toward Z 0 region) and negative for Z 1. one needs to model additional viscous and pressure terms related to velocity ¯uctuations (see Eq. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 Fig. ensuring zero values at either side of the ¯ame brush.4. where the IEM±LMSE model requires a correction to account for kwn ls that is of the order of the laminar ¯ame speed [237]. (258) and (259) shows that the geometrical assumption used to model the mean curvature corresponds to an LMSE±IEM model of the molecular diffusion.121] as in Section 7. This shortcoming is attenuated in the ¯ame surface density context where the surface averaged curvature k7´nl s is multiplied by the ¯ame surface density (Eq.8. (116) under a ¯amelet f assumption yields Ly < Z HH2 =u7Zu: Wrinkling length scales depend on scalar ¯uctuations given by Eq.2) that for a non-reacting scalar. (90)). (208)): k7´nl s < 2 Z 2 Z1 Ly 259 where Ly is a wrinkling length scale of the ¯ame surface and Z 1 is the Z level where mean curvatures are equal to zero. In turbulent non-reactive ¯ows. are then related as Ly < D u7ZuuZ Z p tt . (226).e. (113)). Comparison between measurements (left) and numerical predictions (right) of unweighted temperature pdf at different locations in the Masri et al. 9. 45.4 (Eq. once again.6.8. the mean normal displacement speed kwn ls is expected to be negligible against curvature effects [88] because such ¯ames are unable to propagate. The turbulent mixing time t t and the ¯ame wrinkling length scale.8. The comparison of Eqs.

tP Z p 2 p A Z p . Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 259 Fig. solid line: coalescence±dispersion micromixing Curl model [206]. t 2up i 261 The physical properties of this closure are included in the tensor Gij. the Langevin model [238] for the pressure and viscous unclosed terms of Eq. As already indicated. Veynante. [207] `B' jet ¯ame. Z p . Z Z r 2xi 5 2 n7 2 u Hi uu up . Radial pro®les of CO mass fraction and temperature at different locations in Masri et al. L. The pdf ®eld is decomposed into a set of stochastic particles whose evolution simulates the turbulent ¯ame. The simplest model considers a linear relaxation for the velocity ®eld. x. Using this equivalence. dashed line: LMSE micromixing model (IEM). tP Z p 1 2t 2Z 2 22 Z p is well reproduced by a stochastic process. each particle . (113) may be written: 42 3 1 2p H p p uu u . Basic properties of stochastic methods [239] show that an equation with a diffusive term for a pdf P: 2P Z p 2 1 22 B Z p . as done in Ref. Z Z p P up . tdW t where dW t is a white noise (Wiener process). t Gij up 2 uj P up . tdt 1 B Z t.D. Z . a numerical solution for the pdf implies the development of accurate Monte Carlo solutions [93]. then Gij 1=2 1 3C0 =4 1=kdij . pdf calculations [208]. Reduced chemical scheme of Jones and Lindstedt [209]. More sophisticated methods have been proposed [240]. Closures for the velocity pdf may then be obtained using a Langevin type model. x. shown that the velocity ¯uctuations have pdfs that may be approximated with Gaussian functions. where C0 is a constant. Z p . [238]. x. 46. where a parallel is drawn between Reynolds stress closures (modeled equations for u Hi u Hj ) and the tensor Gij [241]. Z(t). When a Lagrangian context is adopted [242]. similar to the Langevin process: p dZ t A Z t. Two mixing models have been used: ®lled circles: measurements. t 1 C0 1 j 2 p p P u . a simple stochastic model reproducing this velocity distribution.

There are many open questions and challenges left. Keyes DE. [5] Roberts WL. [7] Smooke MD. Driscoll JF. are instantaneously identi®ed at the level of the resolved grid.) are added to the complexity of constant-density turbulent ¯ows. [6] Mueller CJ. In most of Monte Carlo simulations. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 n is characterized by its position xp t. Vorticity generation and attenuation as vortices convect through a premixed ¯ame. a key quantity representing micromixing of the reactants occurring at the small scales where chemical reactions develop. mainly devoted to feasibility tests (two-dimensional simulations. simply re-written in terms of Favre averaged quantities.67:85±122.) are not intentionally included in the calculations. So far LES modeling is essentially based on the same modeling strategies as the ones described here for RANS turbulent combustion modeling. Numerical solution of two-dimensional axisymmetric laminar diffusion ¯ames. 10. whereas counter-gradient turbulent transport is known to appear in some situations. (261) corresponds to the system for the discrete particles: V b dxin u b b in b dt b b 2 2 3 b b ` duin 1 2p 2 ui dWi 2 1 Gij ujn 2 uj 1 C0 11=2 1n r 2xi 2xj 2xj dt b dt b b b b b dY p ~ i 2 Yip Y b i b X 1 v i Yip _ k=1 dt 262 The last dif®culty is the determination of the mean pressure ®eld 2p=2xi : Methods have been proposed to calculate it from Monte Carlo simulations [101]. [93±95]. A laminar vortex interacting with a premixed ¯ame: measured formation of pockets of reactants. Reuss DL. Dynamics of combustion fronts in premixed gases: from ¯ames to detonations. Reynolds stresses are generally described from well established turbulent models (mostly k. All these various models are related via a scalar dissipation rate.26:321±52. Premixed combustion and gasdynamics. which illuminate unexpected links between wellknown closure schemes. Few studies have focussed on species turbulent transport. etc. Driscoll JF. 1993. Williams FA. various dif®culties (strong heat release. Fundamental aspects of combustion. Annu Rev Fluid Mech 1994. Ludford G. its velocity up t and n n p concentration(s) Yn t: The Langevin model given in Eq. 1 ). Conclusion Modeling turbulent combustion is a challenging task. References ÄÂ [1] Linan A. Oxford: Oxford University Press. Combust Sci Technol 1989. LES is at a very early stage for combustion applications and only few works have been done in this direction. Cambridge: Cambridge University Press. etc. ² LES may directly provide part of the description of turbulence/combustion interactions because zones of fresh and burnt gases. However. the results of these preliminary tests suggest that LES will rapidly become a complementary way to carefully simulate and understand turbulent combustion systems. generally modeled using a gradient expression. but more work is needed to reach a level where liquid atomization. vaporization together with ¯ames are properly described.87:245±56. having different turbulence characteristics. [2] Clavin P. Three main ingredients must be modeled: g ~ ~ ² Reynolds stresses: ui uj 2 ui uj : g ~ ~ ² Turbulent transport of species mass fractions: ui Yk 2 ui Y k : ² Mean reaction rate of species: v k : _ Most works have focus on re®ned descriptions of the mean reaction rate v k : To provide this reaction rate. Veynante. 1993.260 D. In turbulent ¯ames. Proc Comb Inst 2000. Theory of laminar ¯ames. [4] Clavin P. Mitchel RE. etc. However. large range of time and length scales. Drake MC. A review of the most classical Reynolds (or Favre) averaged Navier±Stokes (RANS) models has been proposed. which are based on a pure statistical or more oriented geometrical view of turbulent ¯ames. the Lagrangian simulation is coupled with Eulerian calculations of means values providing the mean pressure ®eld and the dissipation rate of the velocity ¯uctuations. ² LES appear as a promising tool to capture combustion instabilities. Rosalik ME. Major models are then directly derived from the proposed expressions.). The precise knowledge of these relations may be useful to understand the exact implications of the underlying physical hypothesis. L. So far. Fully detailed reviews on pdf methods may be found in Refs. ² Flames are mainly driven by mixing. tools _ have been proposed. Combust Flame 1991. DNS with a two-way coupling between a dilute spray and a carrier phase were recently performed to progress in this direction [243] and closures exist [243±245]. combustion effects on the ¯ow (¯ame induced turbulence generation. [3] Buckmaster J. constant density ¯ows. complex chemistry. full pdf methods including a model for the turbulent frequency following a ¯uid particle have also been developed with success [197]. Combust Flame 1998. Nonetheless. One of them is the understanding of the coupling between spray and combustion including detailed chemistry. LES is a good candidate to capture unsteady turbulent mixing. . higher viscous dissipation.112:342±58. RANS combustion models will remain useful for the next few years. large eddy simulations (LES) stands as very promising technique for turbulent combustion: ² Combustion ¯ow ®elds generally exhibit large scale motions [26]. The links and the similarities between combustion models have been demonstrated here.28:569±85.

Ind Engng Chem 1928. Experiments on the scalar structure of turbulent CO/H2/N2 jet ¯ames. Linan A.20:998±1004. Philadelphia: SIAM. [45] Hartley LJ. Prog Energy Combust Sci 1996. editors.21:531±76. [46] Buckmaster J. Mansour MS. Vervisch L. [36] Kim JS. Chen J-Y. p. Â [24] Lesieur M. [40] Phillips H. Pittsburgh: The Combustion Institute. Jou WH. editor. Verzicco R.75(1±3):53±72. 1996 p. Moss JB. New Tools in Turbulence Modelling. In: Metais O. ÄÂ [37] Clavin P. Fiechtner GJ. Berlin: Les Editions de PhysiqueÐSpringer. editors. Ferziger J. [16] Poinsot T. Laminar ¯amelets in turbulent ¯ames. ÄÂ [30] Linan A. Combust Flame 2001. Rogg B. [42] Plessing T. 1997. editors. Veynante D. Turbulent reacting ¯ows. Twenty-Sixth Symposium (International) on Combustion.415:227±60.25:127±40. 1969. Annu Rev Fluid Mech 1998. [43] Kioni PN.28:45±82. PA: Edwards. Bray KNC. Large eddy simulations of combustion instability in an axisymmetric ramjet combustor. Combust Sci Technol 1991. Asymptotic and numerical study of diffusion ¯ames with variable Lewis number and ®nite rate chemistry. J Fluid Mech 1994. Effects of heat release on triple ¯ame.116:192±206.80:23±46. Edge-¯ames and their stability. Principles of combustion. 1982. 105±40. Peters N. 1996. [15] Kuo KK. Trouve A. Combust Flame 1998. 1986. The evolution equation for the ¯ame surface density.6(7):1447±54. Direct numerical simulation of nonpremixed turbulent ¯ame. Veynante. [34] Cuenot B. In: Moin P. Direct numerical simulation for turbulent reacting ¯ows. New York: Plenum Press. 2nd ed.384:107±32. [21] Piomelli U. Equations fondamentales de Â l'aerothermochimie. p. Greenhalgh DA. An introductive course. ÄÂ [31] Linan A. Reading. Poinsot T. [25] Veynante D. Flame spread in laminar mixing layers: the triple ¯ame. p. Poinsot T. Experimental and numerical studies of a triple ¯ame. Poinsot TJ. 5±18. Stanford University. 269±336. [9] Poinsot T. Paris: Editions Technip.104(1):111±37.22:427±509. Prog Energy Combust Sci 1989. Chasnov JR. Williams FA. Acta Astronaut 1974. Theoretical and numerical investigation of a symmetrical triple ¯ame using a parabolic ¯ame tip approximation. Higuera FJ. 1985. [32] Barlow RS. Kumar A. Flame propagation in a nonuniform mixture: analysis of a slowly varying triple ¯ame. Berlin: Les Editions de PhysiqueÐ Springer. 2001. editors. Bray KNC. In: Libby PA. [12] Vervisch L. In: Â Metais O. Baum M. Orlandi P. Metais O. Poinsot T. Vervisch L. 423±31. Moin P. [18] Favre A. New trends in large-eddy simulations of turbulence. 63±113. Combustion in high speed ¯ows. 1983. ÄÂ [41] Kioni PN. Large eddy simulation: an introduction and Â perspective. Pittsburgh: The Combustion Institute. 219±32. 1994. [22] Ferziger J. Tenth Symposium (International) on Combustion. Combust Sci Technol 1981. [28] Ghosal S. J Comp Phys 1995. Prud'homme R. The turbulent burning velocity for large-scale and small-scale turbulence. p. In: Verlarde MG. Applications of direct numerical simulations of premixed turbulent combustion. Vervisch L. Using direct numerical simulations to understand premixed turbulent combustion. 1997. Combust Flame 1993. Dordrecht: Kluwer Academic Publishers. Coherent structures in combustion. Candel S. 1965. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 [8] Givi P. In: SIAM. Effects of nonunity Lewis numbers in diffusion ¯ames. Henningson DS. Ferziger J. [23] Lesieur M. . Combust Flame 2000. Schumann TEW. Flame propagation in a nonuniform mixture: analysis of a propagating triple-¯ame. Combust Flame 1989. Á [13] Barrere M. 1994. Theoretical and numerical combustion. Annu Rev Fluid Mech 1996. Moss JB. Masuya G. Combust Flame 1996. editor. Model free simulations of turbulent reactive ¯ows.D. Theory of gaseous combustion. Reynolds WC.118:24±37. Berlin: Les Editions de PhysiqueÐ Springer. [33] Bray KNC. Large eddy simulations: theory and È applications. Jackson TL. Prog Energy Combust Sci 1996. Recent approaches in large-eddy simulations of Â turbulence. The asymptotic structure of counter¯ow diffusion ¯ames for large activation energies. Ignition and ¯ame spread in laminar mixing layer.95:276. Rogg B. Ferziger J. Turbulence and transition modelling. New tools in turbulence modelling. Combust Flame 1998. Â [11] Poinsot T. Combust Flame 2000. In: Metais O. [48] Bourlioux A. p. 1973. Terhoeven P. Combust Sci Technol 1991.76:71±88. editors. Turbulent transport in con®ned premixed ¯ame. p. Edge ¯ames and partially premixed combustion in diffusion ¯ame quenching. 19th Symposium (International) on Combustion. Statistical equations of turbulent gases.115(3):335±53. Â [17] Trouve A.1007(1). [27] Menon S. [14] Williams FE. The basic equations for the large eddy simulation of turbulent ¯ows in complex geometry. Poinsot T. MA: Addison-Wesley. [44] Dold JW. [20] Shepherd IG. 231±66. 1996. Combust Sci Technol 1996. Diffusion ¯ames. J Fluid Mech 2000. [19] Bray KNC. Paris: Masson. Alfredsson PH. [10] Baritaud T. London: Academic Press. Extinction of diffusion ¯ames with nonunity Lewis numbers. Asymptotic and numerical study of the stabilization of diffusion ¯ames by hot gas. 1994. 461. 1997. Philadelphia. ÄÂ [39] Linan A.31:101±18.278:1±31. 1±28. Stanford. 29±47. Bray KNC. New tools in turbulence modelling. p. Linan A. [26] Coats CM. Dold JW. An experimental and numerical study on a laminar triple ¯ame. J Engng Math 1997.115:41±68. L. Pittsburgh: The Combustion Institute. New York: Wiley. CA: CTR.15:1±107. Reynolds averaged and large eddy simulation modeling for turbulent combustion. [47] Ghosal S. Turbulence production in premixed turbulent ¯ames. [38] Favier V.120(1/2). [35] Burke SP. Problems of hydrodynamics and continuum mechanics. J Fluid Mech 1999. Peters N. editors. Carter CD. p. Libby PA. In: Hallback H.30:655±92. 261 [29] Peters N. Dordrecht: Kluwer Academic Publishers. Cuenot B.125(1/2):788±803. p. Phys Fluids 1995. Johansson AV. Poinsot T. Combustion theory. Non equilibrium cooperative phenomena in physics and related ®elds. Studying turbulence using numerical databasesÐV. ÄÂ [49] Ruetsch GR. Linan A. Williams FA. In: Buckmaster J.120(4):549±69. Flame in a buoyant methane layer.

A capturing±tracking hybrid scheme for de¯agration discontinuities. 1996. 1978. J Fluid Mech 1991. [58] Nieuwstadt FTM. Veynante D. [55] Vervisch L. editors. Lee. 1231±50. Combust Flame 1995. One-dimensional propagation of a premixed turbulent ¯ame with a balance equation for the ¯ame surface density. Veynante D. Nieuwstadt FTM. [64] Tennekes H. 233±40. Van Kalmthout E.70:1±15. An experimental and numerical study on a laminar triple ¯ame. Kleiser L. Combust Theor Model 2000. p. Cook A. Laminar ¯amelet concepts in turbulent combustion.102(3):285± 97. Twenty-First Symposium (International) on Combustion. Field equation for interface propagation in an unsteady homogeneous ¯ow ®eld. Mahalingam S. Echekki T. [53] Favier V. Pittsburgh: The Combustion Institute. Combust Sci Technol 1993:90. Williams FA.37(7):2728±31. 1994. Mansour MS. p. chemical and physical principles. p. Poinsot TJ. 115. Int J Engng Sci 1988. A ®rst course in turbulence. Candel S. Pittsburgh: The Combustion Institute. [70] Cuenot B. Bray KNC. Proc Combust Inst 2000. Paris 6. Veynante D. Nonpremixed reacting ¯ows near extinction. The evolution of surfaces in turbulence. Moser V. Direct and large eddy simulation II. 1972. Combust Sci Technol 1991. 1239±45. [63] Peters N. Chen JH. 1988. p. New York: Springer. Large-eddy simulation of air pollution dispersion: a review. Williams FA.14:245±92. [59] Meeder JP. Subgrid-scale segregation of chemically reactive species in a neutral boundary layer. Piana J. [66] Roberts WL. 1997. Using numerics to help understanding of nonpremixed turbulent ¯ames. Chaffee Hall. Veynante. Finite-rate chemistry and transient effects in direct numerical simulations of turbulent nonpremixed ¯ames. Goss LP. 1997. p. Â [56] Borghi R. Chen JH. Turbulent reacting ¯ows. Smiljanovski V. In: 29th Aerospace Science Meeting. Poinsot T. [68] Borghi R. J Comp Phys 1996. Large eddy simulations of turbulent premixed ¯ames using a capturing/tracking hybrid approach. 1991. Poinsot T. Twenty-Second Symposium (International) on Combustion. Direct and large eddy simulation II. 1997.129(2):263±83. 301±10. Pitz RW. Pittsburgh: The Combustion Institute. [61] Borghi R. Cambridge: MIT Press. Libby PA.114(1/2):231±45.262 D.143:25±62. Candel S. Candel S. In: Buckmaster JD. In: Metais O.228:561±606. 1998. Combust Theor Model 1997.26(5):445±69. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 combustion. Turbulent combustion: fundamental aspects and a review. Philadelphia: SIAM. Direct numerical simulation of a turbulent reactive plume on a parallel computer. Kollmann W. Kerstein AR. Louisiana 1996. [57] Villasenor R. Veynante D.28:11±24.115(3):335±53. TwentySixth Symposium (International) on Combustion. 2±61. 1986. 44. New Orleans. Meeder JP. Voke PR. Combust Sci Technol 1990. Williams FA. Interaction between chemical reaction and turbulence in supersonic non-premixed H2 ±air combustion. Project Squid Headquarters. p. Peters N. Cuenot B. Broadwell JE.7(10):2496± 503. Driscoll JF. Pittsburgh: The Combustion Institute. Vervisch L. TwentySixth Symposium (International) on Combustion. Kleiser L. Vervisch L. Lumley JL. In: Chollet JP.94:58±69. 1998. Mongia HC. [60] Bray KNC. Pittsburgh: The Combustion Institute. Terhoven P. Fichot F. Phys Fluids A 1995. editors. editor. In: Libby PA. [62] Borghi R. p. [67] Bilger RW. [69] Y. Bourlioux A. Combust Flame 1998. Vervisch L.4(1):77. The mathematics of combustion. Purdue University. Surface density function in premixed turbulent combustion modeling. 1977. [51] Plessing T. 1994. Phys Rev A 1988. Dordrecht: Kluwer Academic Publishers. 1996. Effects of unsteadiness in edge-¯ames and liftoff in non-premixed turbulent combustion. Effects of curvature and unsteadiness in diffusion ¯ames. Combust Flame 1998. PhD Thesis. Kim W-W. Ashurst W. PhD Â Thesis.124(4)646±55. Klein R. Turbulent combustion. Williams FA. vol.81(1±3):45. Sixth International Conference on Numerical Combustion. Destriau M. Combust Flame 1993. L. 1980. p. Structure and propagation of methanol± air triple ¯ames. Project Squid TRW-9-PU. Topics in applied physics. Williams FA. Flame stretch and the balance equation for the ¯ame area. Pope SB. In: Libby PA. Stability diagram for lift-off and blowout of a round jet laminar diffusion ¯ame. Quenching processes and premixed turbulent combustion diagrams. Twenty-Fifth Symposium (International) on Combustion. [52] Echekki T. Lacas F.1(2):183±215. In: Chollet JP. Paris: Editions Technip. The challenge of turbulent combustion. Egolfopoulos FN. Moser V. The coherent ¯ame model of nonpremixed turbulent combustion. Combust Flame 2001. Dordrecht: Kluwer Academic Publishers. Poinsot T. editors. Reno (Nevada). [65] Poinsot T. Ferziger J. Klein R. Drake MC. similarities between probability density function and ¯ame surface approach. Combust Sci Technol 1999. Marble FE. Candel S. 264±80. Bray KNC. On the structure and morphology of turbulent premixed ¯ames. Universite Pierre et Marie Curie. Berlin: Les Editions de PhysiqueÐSpringer. editors. Riley JJ. Cornell University. Reactions chimiques en milieu turbulent. New tools in turbulence modelling. Bidaux E. AIAA paper 91-0375. [54] Ghosal S. 1985. p. implications for turbulent diffusion [71] [72] [73] [74] [75] [76] [77] [78] [79] [80] [81] [82] [83] [84] [85] [86] [87] [88] . Mungal MG. Images of the quenching of a ¯ame by a vortex: to quantify regimes of turbulent combustion. Menon S. 1994. Pittsburgh: The Combustion Institute. Triple ¯ames and partially premixed combustion in autoignition of nonpremixed mixtures. Turbulent ¯ows with premixed reactants in turbulent reacting ¯ows. Rec Adv Aerosp Sci 1985:117±38. Vervisch L. Direct numerical simulation analysis of the G-equation in premixed combustion. TwentySeventh Symposium (International) on Combustion.Y. A study of the laminar ¯ame tip and implications for premixed turbulent combustion. editors. Chen JY. 321±30. Vervisch L. Prog Energy Combust Sci 1988. Direct numerical simulations [50] Domingo P. An unsteady laminar ¯amelet model for nonpremixed combustion. London: Academic Press. 116±51. Combustion and ¯ames. Turbulent combustion modelling. Poinsot T. Voke PR. The structure of turbulent non premixed ¯ames. Large-eddy simulation of a gas turbine combustor ¯ow.

Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 analysis of ¯ame surface density models for non premixed turbulent combustion. Abramowitz M. Spalding DB. Veynante D. Experimental analysis of ¯ame surface density model for premixed turbulent combustion. Gouldin FC. 111±36.5(2):327±34. Pdf method for turbulent reacting ¯ows. Etude theorique et experimentale d'une ¯amme Â turbulente stabilisee par un barreau cylindrique. Conditional moment closure of mixing and reaction in turbulent nonpremixed combustion. Prog Energy Combust Sci 1999. Pittsburgh: The Combustion Institute. 1997. Twenty-Second Symposium (International) on Combustion. p. Ecole Centrale Paris. Pittsburgh: The Combustion Institute. Â Poinsot T. 541±63.D. Abdel-Gayed RG. Cambridge: Cambridge University Press. Proceedings of the 1996 Summer Program. Phys Fluids A 1997. 40. Twenty-Third Symposium (International) on Combustion. The pdf approach to turbulent ¯ow. A uni®ed statistical model of the premixed turbulent ¯ame. Kollmann W. Stanford. Williams FA. Lundgren TS. Pope SB. Said R. p. Turbulent reacting ¯ows. Lecture Notes in Engineering. Analyse de modeles. Yakhot V. Phys Fluids 1993. Chemical closure and burning rates in . Orszag CG. 1984. 569±77. 1990.80:304±12. In: Libby PA. Im HG. L. Veynante D. Â Â Duclos JM. Gouldin FC. London: Academic Press.116(4):519±45. Duclos JM. Twenty-Sixth Symposium (International) on Combustion. Mantel T. 505±12. Annual research briefs.99:581±90. Libby PA. 52±5. p. The development of a new ¯ame area combustion model using conditional averaging. Ecole Centrale Paris. Imperial College of Science Technology and Medicine. Bilger RW. USA: Dover. Piana J. Development of turbulence models for shear ¯ows by a double expansion technique. Twenty-Sixth Symposium (International) on Combustion. editors. Pittsburgh: The Combustion Institute. Duclos JM. 185. 1988 p. Borghi R.9(12):3826±33. Pittsburgh: The Combustion Institute. Murphy SN. Turbulent reacting ¯ows. Computations of turbulent combustion: progress and challenges. Combust Flame 1994. 1998. Prog Energy Combust Sci 1985. Stegun IA. Williams FA. 649±57. Experimental analysis of ¯amelet models for premixed turbulent combustion. 1996. 213±308. Berlin: Springer. Turbulent combustion. editors. Weller HG. Scott MJ. Fluid Dynam 1990. Comparison of prediction and measurement in non-premixed turbulent ¯ames. Borghi R.10(9):2347±68. Peters N. p. 1996. Lewis number effects on turbulent burning velocity. Statistical description of the turbulent mixing of scalar ®elds. Piana J. 1994. Ruetsch G. 23rd Symposium (International) on Combustion. Marooney CJ. Martel C. Acta Astronaut 1977. p. 1990. Valino L. A simulation with a cellular automaton for turbulent combustion modelling.10(5):969±75.4(7): 1510. 2000. 20th Sympo- 263 [89] [90] [110] [111] [91] [112] [92] [113] [114] [93] [94] [95] [96] [97] [115] [116] [117] [118] [119] [98] [99] [100] [101] [102] [103] [104] [120] [121] [122] [123] [124] [105] [106] [107] [108] [109] [125] [126] sium (International) on Combustion. Phys Fluids 1967. The interaction between turbulence and chemistry in premixed turbulent ¯ames. Reynolds WC. Recent developments in pdf methods. Fuego F. Bray KNC. vol. UK. Hamid MN.25(6):595±687. Klimenko AY. PhD Thesis. Gosman AD. Center for Turbulence Research.76:213±8. Libby PA. A new model of premixed wrinkled ¯ame propagation based on a scalar dissipation equation. 1996. 1988. 85±99. NY. Bradley D. 1994. Twenty-Second Symposium (International) on Combustion. Lawes M. Conditional moment closure for turbulent reacting ¯ow. Turbulent ¯ame propagation in partially premixed ¯ames. In: Williams FA. Veynante D. Turbulent reacting ¯ows. p. Veynante. The probability density function (pdf) approach to reacting turbulent ¯ows.135:153±73. 375±474. Â Martel C. Mixing and chemical reaction in steady con®ned turbulent ¯ames. Int J Mod Phys B 1997. Pittsburgh: The Combustion Institute. editors. Bradley D. p. Weller HG. editors. Smith NSA. p. CA: CTR. editors. 1993. Gouldin FC. Assessment of combustion submodels for turbulent nonpremixed hydrocarbon ¯ames. Etude experimentale de la combustion turbulente Â Â Â Á premelangee. Phys Fluids A 1998. 381±8. Turbulent reacting ¯ows.4:291±319. Phys Fluids 1992. Combust Flame 1999. Chen JY. Lawes M. Large eddy simulation of turbulent front propagation with dynamic subgrid models. Ferziger JH. Pittsburgh: The Combustion Institute. Dopazo C. 1993. TwentyFifth Symposium (International) on Combustion. Combustion regimes and the straining of turbulent premixed ¯ames. Turbulent premixed combustion: further discussions on the scales of ¯uctuations. Stanford University.25:327. 13th Symposium (International) on Combustion. Pdf modeling of chemical nonequilibrium effects in turbulent non-premixed hydrocarbon ¯ames.11:119±95. Pope SB. Miles PC. p. Thangam S. Moss JB. Swaminathan N. Combust Flame 1989. A new spectral method for calculation of the time-varying area of a laminar ¯ame in homogeneous turbulence.96(4):443. Trouve A.11(25). Champion M. 1980. p. London: Academic Press. 1989. 413±20. Bradley D. Bruel P.1:285±349. London: Academic Press. Pittsburgh: The Combustion Institute. Combustion intensity and burning rate integral of premixed ¯ames. Joint Meeting of the British and German Sections of the Combustion Institute. Combust Flame 1990. Pittsburgh: The Combustion Institute. O'Brien EE. Theor Comp Fluid Dynam 1990. Chen J-Y. Gatski TB. Kollmann W. Lahjaily H. Bilger RW. Combust Sci Technol 1998. Bray KNC. Pittsburgh: The Combustion Institute. Review of closure models for the ¯amelet regime of premixed turbulent combustion. Queens' College. Distribution function in the statistical theory of turbulence. Combust Flame 1994. In: Moin P. PhD Thesis. In: Libby PA. In: Borghi R. Speziale CG. p. Borghi R. Dopazo C. Introduction of dilution in the BML model: application to a stagnating turbulent ¯ame. Kollmann W. Handbook of mathematical functions. 1996. Abdel-Gayed RG. Thermo-Fluids Section Report TF/9307. p. 1970. Cambridge. Bilger RW. Karmed D. Lund TS. Klimenko AY. Champion M. 629±36. 1994. 1971.

Poinsot T. 719±29. Bray KNC. Gore JP. 27th Symposium International on Combustion 1998. Kalt PAM.96(4):351± 70.99:687±96. 1986. Modelling of ¯amelet surface to volume ratio in turbulent premixed combustion. AIAA J 1981. Stanford. Effects of heat release in turbulent. Dimotakis PE.120(1/2):1±18. J Fluid Mech 1997. Â Bailly P.117(3):514±28. 95±124. Bruel P. Domingo P. Combust Flame 1999. Garreton D. Combust Flame 1998. In: Moin P. In: Center for Turbulence Research. Turbulent transport in premixed ¯ames. Non gradient theory for oblique turbulent ¯ame with premixed reactants. 1998. Libby PA. editor. Gradient and counter-gradient scalar transport in turbulent premixed ¯ames. CA: CTR. In: Zierep J. Bray KNC. Proceedings of the Summer Program. Choi CR.114(3/4):336±48. Premixed ¯ames in stagnating turbulence part iv: a new theory for the Reynolds stresses and Reynolds ¯uxes applied to impinging ¯ows. Evaluation des termes de transport et de Â dissipation de surface de ¯amme par simulation numerique directe de la combustion turbulente. J Fluid Mech 1997. 809±15. Combust Sci Technol 1989. Pittsburgh: The Combustion Institute. 1994. Poinsot T. Computational methods for ¯uid dynamics. Sixteenth Symposium (International) on Combustion. Â Ferziger JH. L.68:15±33. Combust Flame 2000. PhD Thesis. London: Academic Press. Dopazo C. 1996.121(4):555±74. Turbulent transport in premixed turbulent ¯ames. Chemical closure model for fractal ¯amelets. Theoretical analysis of the effect of gravity on premixed turbulent ¯ames. Trouve A. Priddin CH. Poinsot T. Experimenal investigation of turbulent scalar ¯ux in premixed ¯ames. Sivathanu YR. Sanquer S. [127] [128] [146] [129] [147] [148] [130] [131] [149] [150] [132] [151] [133] [134] [152] [153] [154] [135] [136] [137] [138] [139] [155] [156] [157] [158] [159] [140] [141] [142] [143] [144] [145] [160] [161] [162] . 1996. Pittsburgh: The Combustion Institute. Turbulent combustion of premixed ¯ames in closed vessels. Bray KNC. Libby PA. A comparison of ¯amelet models for premixed turbulent combustion. Cheng WK. Pope SB. Bray KNC. Effects of pressure gradients on turbulent premixed ¯ames. Chen JY. Coupland J. Deshaies B. Boger M. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 premixed turbulent ¯ames. Study of premixed turbulent combustion including Landau±Darrieus instability effects. In: Durst F. Studying turbulence using numerical databasesÐV. Â Bailly P. Meneveau C.116(1±2):1±14. Stanford University. Mixing and chemical reactions in turbulent liquid mixing layer. Hermanson JC. University of Sydney. Gouldin FC.199:297±332. Boughanem H. Pittsburgh: The Combustion Institute. Combust Flame 1995.19:205±13.264 D. Jones WP. 23rd Symposium (International) on Combustion. Development of a coherent ¯amelet model for a spark ignited turbulent premixed ¯ame in a closed vessel. TwentySixth Symposium (International) on Combustion. Turbulent reacting ¯ows. p. p. Contribution l'etude de l'interaction turbulenceÐ Â Â Á combustion dans les ¯ammes turbulentes de premelange a Á Â l'aide de modeles du second ordre. Dimotakis PE. On the mathematical modeling of turbulent combustion with special emphasis on soot formation and combustion. Stanford. The production of premixed ¯ame surface area in turbulent shear ¯ow. p.199: 333±75. Bray KNC. 1483±91. An evaluation of ¯ame surface density models for turbulent premixed jet ¯ames. International Congress and Exposition. PhD Thesis. J Fluid Mech 1986. An evaluation of combined ¯ame surface density and mixture fraction models for nonisenthalpic premixed turbulent ¯ames. Trouve A. Prasad ROS. Modeling of wrinkled laminar ¯ames with intermittency and conditional statistics. Veynante.81:433±8. Bray KNC. Countergradient diffusion in premixed turbulent ¯ames. Moss JB. editors.77:241±59. Combust Flame 2000. Pittsburgh: The Combustion Institute. p. Masuya G. Berlin: Springer. Huh KY. p. Combust Flame 1993. p. 1982. Bray KNC. The coupling between ¯ame surface dynamics and species mass conservation in premixed turbulent combustion. Mantel T. Â Trouve A. Schmidt FW. Eighth European Turbulence Conference.170:83±112. Stretching and quenching of ¯amelets in premixed turbulent combustion.86:311±32. Modelling of ¯ow and combustion in a production gas turbine combustor. Veynante D. Combust Flame 1994. 1996. Libby PA. Lumley JL. Lumley JL. Duplantier S. editors. Veynante D. Paul RN. Gore JP. Combust Flame 1994. J Fluid Mech 1989. Champion M. Bray KNC. Combust Flame 1989. Reynolds WC. Germany: University of Karlsuhe. Libby P. On conditioned averages for intermittent turbulent ¯ows. Detroit. 1999.332:263±93. Barcelona 2000. Checkel MD. 1976. 1994. Peric M. p. Williams FA. Laminar ¯amelet expressions for pressure ¯uctuation terms in second moment models of premixed turbulent combustion. reacting shear layer. 1994. Rutland C. Â Trouve A. Experimental and numerical study of a premixed ¯ame stabilized by a rectangular section cylinder. Bray KNC. 1990. 1996. Convective transport and instability phenomena. The occurence of ¯ame instabilities in turbulent premixed combustion. Orleans. Champion M.19:205±13. J Fluid Mech 1977. Universite de Poitiers. Twenty-Sixth Symposium (International) on Combustion. Diringer JA. Â Boughanem H. Thomas A. Stanford: Center for Turbulence Research. Effects of heat release on the large-scale structure in turbulent mixing layer. Combust Flame 1999. Oertel H. 75±94. Gouldin FC. Mantel T. Whitelaw JH.100(1/ 2):201. 923±30. Turbulence modelling and numerical solution methods for variable density and combusting ¯ows. 1987. Launder BE. Combust Flame 1991. Pittsburgh: The Combustion Institute. In: Libby PA. SAE paper 910268. J Fluid Mech 1989. McMurty PA. editors. Veynante D. 1991. Koochesfahani MM. p. Magnussen BF. Cant RS. Turbulent shear ¯ows 5th. Veynante D. AIAA J 1981. Riley JJ. 21st Symposium (International) on Combustion. PhD Thesis. Paul RN. Duclos JM. Â Veynante D. Berlin: Springer.353:83±114. Metcalfe RW. 259±66. Cant RS. France.95(1/2):101±18. Libby PA. Large eddy simulations of a turbulent premixed V-shape ¯ame. p. editors. 309±68. Prasad ROS. 310±23. Numerical modelling of SI engine combustion with a ¯ame sheet model. University of Rouen. Chen JY. Simonin O. Hjertager BH.

[168] Rogg B. [188] Van Kalmthout E. Peters N. Combust Sci Technol 1988. A linear-eddy model for turbulent scalar transport and mixing. Grenoble. Appendix C. Borghi R. NV. [183] Bradley D. Pope SB. Pope SB. A Monte Carlo treatment for reacting and coalescing dispersed phase systems. Wang H. Stanford University. Reno.6:67±87. Berlin: Springer. Trygvason G. Chen JY. 1995. in a temporally developing shear layer. 1996. [192] Girimaji SS. Ruetsch GR. Pittsburgh: The Combustion Institute. Kosaly G. 1998. 1057±64.11(6):1550±71. Dahm WJA. An unsteady laminar ¯amelet model for non-premixed combustion. On the structure of nonsooting counter¯ow ethylene and acetylene diffusion ¯ames. 1099±106. CA: CTR. Veynante D. Dahm WJA. Ferziger JH. In: AIAA Paper 93-0248. Candel S. Law CK. p. [170] Bish ES.27:1199± 206. Lagrangian model simulations of molecular mixing. Peters N. p. Rogg B. Pittsburgh: The Combustion Institute. Studying turbulence using numerical databasesÐVII. [186] Fichot F. Twenty-Fifth Symposium (International) on Combustion. [197] Van Slooten Jayesh PR. 39±60. Combust Flame 1994. [167] Sung CJ. [184] Cuenot B.7(5):1082±94. Chen M. France 1997. Combust Sci Technol 1995.102:481±92. 1994. p. Measurements of scalar dissipation in the reaction zones of turbulent nonpremixed H2 ±air ¯ames. 1994. Pittsburgh: The Combustion Institute. A ®eld Monte Carlo formulation for calculating the probability density function of a single scalar in turbulent ¯ow. Peters N. Gonzalez M. Pope SB. Nilsen VN. Combust Sci Technol 1991. A lagrangian simulation of ¯amelet extinction and re-ignition in turbulent jet diffusion ¯ames. Ecole Centrale Paris. Nonpremixed turbulent reacting ¯ow near extinction. Chem Engng Sci 1965. Gu XJ. Combust Sci Technol 1992. Stanford.3(5):1300±11. The Cambridge universal laminar ¯amelet computer code. Reduced kinetic mechanisms for applications in combustion systems. Blint RJ. [191] Ravet F. 1993. Brown TM. Peters N.101:501±28. [177] Pitsh H. TwentySecond Symposium (International) on Combustion. Combust Flame 1998:114. Poinsot T. Simulation and modeling of reactive particles in turbulent nonpremixed combustion. Strain rate modelling for a ¯ame surface density equation with application to non-premixed turbulent combustion. Veynante D. [199] Valino L. p. Structural response of counter¯ow diffusion ¯ames to strain rate variations.91:179±86. p. 952357. Pittsburgh: The Combustion Institute. Zurich. Riley JJ.4:77±97. Breitbach H. [189] Obounou M. Bray KNC.60:391±421. 1988. 1±10.99:775±83.10(1): 246±65. Combust Theor Mod 2000. Phys Fluid 1991.4:2529. NV 1998. È [185] Muller CM. Reno. SAE Paper. [200] Fox RO. Flamelet modelling of stabilization in turbulent non-premixed combustion. [166] Sun CJ. Law CK. Swiss Federal Institute of Technology. [171] Ferreira JC. Combust Flame 1990(82). 1993. Combust Flame 1995. PhD Thesis. Vervisch L. PhD Thesis. The spectral relaxation model of the scalar dissipation rate in homogeneous turbulence. Oefelein J. ETH. Phys Fluids 1992. 36th Aerospace Sciences Meeting and Exhibit AIAA paper 98-1027. 1998. Western States Section Fall Meeting. 589±97. [195] Fox RO. Flow Turb Combust 1998. [179] Delhaye B. Kosaly G. 25th Symposium (International) on Combustion. Reynolds WC.105:211±27. Levenspiel O. Advances in pdf modeling for inhomogeneous turbulent ¯ows. [187] van Kalmthout E. Analysis of ¯ame surface density concepts in non-premixed turbulent combustion using direct numerical simulation. Simulation and modeling of reactive shear layers. Phys Fluids 1995. 1990. Pittsburgh: The Combustion Institute. Lucht RP. Combust Flame 1995. [190] Broadwell JE. p. A lagrangian model for predicting turbulent diffusion ¯ames with chemical kinetic effects. [164] Montgomery CJ. p. Â [178] Mell WE. The lagrangian spectral relaxation model for differential diffusion in homogeneous turbulence. Sung CJ. 1107±13. Twenty-Fifth Symposium (International) on Combustion. Veynante D. Prog Energy Combust Sci page 1984:319. Etude des ¯ammes de diffusion turbulentes. Rogg B. Proc Combust Inst 1998. 693±8. Gaskell PH. A strained dissipation and reaction layer formulation for turbulent diffusion ¯ames. Dibble RW. In: Peters N. [198] Spielman LA. Pitz RW. Veynante D. Direct numerical simulation investigation of the conditional moment closure model for nonpremixed turbulent reacting ¯ows. [180] Delhaye B. Unsteady ¯amelet modeling of turbulent hydrogen±air diffusion ¯ames. editors. Theoret Comput Fluid Dynam 1994. [173] Pitsch H. including ®nite rate chemical reactions. Lutz A. Eleventh Symposium on Turbulent Shear Flows. [175] Darabiha N. [165] Peters N. 2-d Simulation of turbulent autoignition with transient laminar ¯amelet source term closure. [174] Zhang Y. Pittsburgh: The Combustion Institute. Phys Fluids 1998. Keller D. editors. Computational methods for turbulent reacting ¯ows . 1994. L.107(4):321±35. Phys Fluids 1999. [181] Chang CHH. Delhaye B. Wan YP. Partially premixed turbulent ¯ame propagation in jets ¯ames. [193] Smith NSA. Pittsburgh: The Combustion Institute. Â Simulations directes et modelisation. Candel S. 1994. Candel S. The importance of time-dependent ¯ame structures in stretched laminar ¯amelet models for turbulent jet diffusion ¯ames. A turbulent jet chemical reaction model: NOx production in jet ¯ames. Combust Flame 1996. [169] Nandula SP. Riley JJ. 265 [182] Lee YY. Direct numerical simulation analysis of ¯ame surface density equation in nonpremixed turbulent combustion.60(2):157±72. Twenty-Sixth Symposium (International) on Combustion.20:247±54. Direct numerical simulation of turbulent H2 ±O2 combustion using reduced chemistry.D. Drake MC. Transient behaviour of laminar counter¯ow hydrogen±air diffusion ¯ames with complex chemistry. [196] Fox RO. [172] Haworth DC. 1993. The mathematical modeling of liftoff and blowoff of turbulent non-premixed methane jet ¯ames at high strain rate.86:163±81. In: Moin P. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 [163] Barlow RS. Modeling non-premixed turbulent combustion in aeronautical engines using pdf-generator. Twenty-Third Symposium (International) on Combustion. Veynante. Twenty-Seventh Symposium (International) on Combustion. [194] Kerstein AR. 1996. Egolfopoulos FN. [176] Mauss F. Pittsburgh: The Combustion Institute. 31st Aerospace Sciences Meeting and Exhibit. Liu JB.

Part 7. Pdf modeling of ®nite rate chemistry effects in turbulent nonpremixed jet ¯ames. AIChE J 1989. Phys Fluids A 1991. Improved Fokker±Planck model for the joint scalar. Ros J.231:361±94. Dutta A. Phys Fluids 1991. Vervisch L. Kollmann W. Dopazo C.115(1/2):210±29. Gao F.73:233±49. Linear-eddy modelling of turbulent transport. Norris AT. Phys Fluids A 1991. Part 6. Mapping closure and non-gaussianity of the scalar probability density functions in isotropic turbulence. Monte Carlo implementation and analytic solution of an inert-scalar turbulent-mixing test problem using mapping closure. Valino L. Revue de L'Institut FranÂ cais du Petrole 1996. On velocity-conditioned scalar mixing in homogeneous turbulence. Bilger RW. Pope SB. A binomial sampling model for scalar turbulent mixing.51(2).1(6):334±48. Phys Fluids 1979. Â [234] Reveillon J. Global reaction schemes for hydrocarbon combustion.3(9):2191±8. Jones WP. Pdf modeling and analysis of thermal NO formation in turbulent nonpremixed hydrogen±air jet ¯ames. L. J Scient Comput 1988. Pope SB. [231] Masri AR. Combust Flame 1998. Closure of the transport equation for the probability density function of turbulent scalar ®elds. Dopazo C.37:521±35. Ecole Centrale Paris. Phys Fluids 1994. [232] Fox RO.26:23. Handbook of stochastic methods. Combust Flame 1992.4(5):1028±35. Calculations of premixed turbulent ¯ames by pdf methods. Theory and effects in simple reactors. Direct numerical simulation of the turbulent mixing of a passive scalar. 1731±8. Phys Fluids 1991. An analytical solution for the scalar probability density function in homogeneous turbulence. [201] [202] [203] [204] [205] [206] [207] [208] [209] [210] [211] [212] [213] [214] [215] [216] [217] [218] [219] [220] [221] [222] [223] .4(47). A generalized Langevin model for turbulent ¯ows. p. Vervisch L. [244] Hollmann C. University of New York. Combust Flame 1991. AIChE J 1964. 1998. Pope SB. Janicka J. Finite rate chemistry and multi-stream mixing.35(12). Combust Flame 1987:67. Kollmann W. Â [243] Reveillon J. Turbulent non-premixed ¯ames of methane near extinction. [239] Gardiner CW.36(3):336±41. Theoret Comput Fluid Dynam 1991. Closures for probability distributions. On the existence of a stochastic Lagrangian model representation for secondmoment closures. Combust Flame 1998.31(3):506±20.3(6). Micromixing phenomena in stirred reactors. Masri AR. Bull Am Phys Soc 1989. Twenty-Sixth Symposium (International) on Combustion. Chem Engng Process 1998. Stony Brook. AIAA J 1998.29:208±28. A mapping closure for multispecies ®ckian diffusion. Spray vaporization in non-premixed turbulent ¯ames: a single droplet model. [236] Subramaniam S. Dopazo C. Eswaran V. Subramaniam S.34:2298.3(10):2438±44. Microstructure of diffusive scalar mixing ®elds. [240] Pope SB. Phys Fluids A 1991. Phys Fluids 1994. Annu Rev Fluid Mech 1994.83:27±42. Villermaux J. 1997. Ind Engng Chem Res 1998:6. [225] Gao F. Combust Sci Technol 1982. A binomial langevin model for turbulent mixing. Phys Fluids 1986.2:255±70. PhD Thesis. Combust Flame 2001. Linear-eddy modelling of turbulent transport. A mixing model for turbulent reactive ¯ows based on Euclidean minimum spanning trees. Curl RI.3(4):511±3. Berlin: Springer. Dopazo C. [235] Subramaniam S.240:289±313. On the relationship between lagrangian micromixing models and computational ¯uid dynamics. AIChE J 1963. Non-isothermal turbulent reactive ¯ows: stochastic approaches. scalar gradient pdf. Pope SB.175(9). 1996. Pope SB. Ë Corrsin S. Pittsburgh: The Combustion Institute. Pope SB. J Fluid Mech 1991.10:870±7.6:973±85.3(1):25±43.28:131±5. [228] Jiang T-L. Turbulent mixing model based on ordered pairing. Leonard A.117(4):732±54. Arbitrary Schmidt number. [233] Fox RO. Phys Fluids 1996. Lindstedt RP. Hill J. Part II.88:397±412. Mapping closure for turbulent mixing and reaction. J Fluid Mech 1992.22(20). The bmc/giem model for micromixing in non-premixed turbulent reacting ¯ows.121(1/2):75±90. Valino L. 1973.73:261±8. A mixing model to improve the pdf simulation of turbulent diffusion ¯ames. Closure models for turbulent reacting ¯ows. [242] Pope SB. Phys Fluids 1982. Combust Flame 1988. Valino L. Encycl Fluid Mech 1986:707. PhD Thesis. Givi P. Combust Flame 1998.8(7):1702±4.266 D. Gao F. Phys Fluids A 1992. [238] Haworth DC. An improved turbulent mixing model. [230] Fox RO.8(10):2678±91. Binary and trinary scalar mixing by Fickian diffusionÐsome mapping closure results. Relaxation of initial probability density functions in the turbulent convection of scalar ®elds. Pittsburgh: The Combustion Institute. Á Â Â [245] Calimez X. [229] Kerstein AR. [237] Anand MS. Direct numerical simulation of turbulent ¯ows with chemical reaction.115(4):487±514. Joint probability density function of a scalar and its gradient in isotropic turbulence. Dispersed phase mixing. Consistent modeling of scalars in turbulent ¯ows. Kakhi M. Peeters TWJ. Tarbell JM. Roekaerts D. 1996. [227] Chen J-Y. Kraichnan RH. Phys Fluids 1988. On the relationship between stochastic Lagrangian models of turbulence and second-moment closures. Fox RO. Modeling of turbulent spray diffusion ¯ames including detailed chemistry. The isotropic turbulent mixer. [241] Wouters HA.2(7):1204±12. Gutheil E. Tsai K. Kolbe W. Vervisch / Progress in Energy and Combustion Science 28 (2002) 193±266 in the chemical process industry. Pope SB. Combust Flame 1988. Gao F. [224] Gao F. Simulation a petite echelle par une methode VOF Â Â d'ecoulement diphasiques reactifs.3(3).3(12):3034±7. Phys Fluids A 1990. Phys Fluids 1991. Twenty-Sixth Symposium (International) on Combustion. Pope SB. Veynante. [226] Kerstein AR. Dibble RW.26:404±8. Pope SB. Pope SB. J Non-Equilib Thermodyn 1979. Jones WP. Dopazo C. O'Brien EE. Comparison of mixing model performance for nonpremixed turbulent reacting ¯ows. Phys Fluids 1996. O'Brien EE. Lagrangian pdf methods for turbulent ¯ows. Subgrid mixing modeling: a dynamic approach.