You are on page 1of 26

COMBUSTION AND FLAME 4 8 : 1 - 2 6 (1982)

Calculation Methods for Reacting Turbulent Flows: A Review


W. P. JONES
Department of Chemical Engineering and Chemical Technology, ImperialCollege, London SW7 2By, England

and
J. H. W H I T E L A W
Department of Mechanical Engineering, ImperialCollege, London SW7 2By, England

The purpose of this review is to describe and appraise components of calculation methods, based on the solution of
conservation equations in differential form, for the velocity, temperature and concentration fields in turbulent
combusting flows. Particular attention is devoted to the combustion models used within these methods and to
gaseous-combustion applications.
The differential equations are considered first and the implications of conventional (i.e., unweighted) and densityweighted averaging discussed in the contexts of solution methods and physical interpretation. In general, it is concluded
that equations should be solved with dependent variables in density-weighted form and that the interpretation of
measurements requires special care to distinguish between conventionally averaged and density-weighted properties.
Finite difference approximations contained within numerical procedures for solving the equations relevant to two- and
three-dimensional recirculating flows, such as are to be found in combustion chambers, are considered briefly. It is
concluded that computer storage requirements very often preclude the possibility of reducing the numerical error to
entirely negligible proportions everywhere. Considerable care must therefore be taken in both specifying a sufficient
number and the distribution of mesh points to be used and also in the interpretation of computational results. Turbulence
models are also discussed briefly and deficiencies noted. Since many flows are partly controlled by mechanisms other
than diffusion and turbulence transport, these deficiencies are of major importance in a limited range of circumstances
which are discussed.
The various recent methods proposed to represent reaction in turbulent flames are reviewed in relation to diffusion and
premixed flames and to flames in which an element of both is present. The application of laminar flame sheet models,
chemical equilibrium assumptions, probability density functions of different forms, and truncated series expansion of
reaction rate expressions are considered together with the use of probability density function transport equations and
their (Monte Carlo) solution. The appraisal is made in relation to presently available results and future requirements and
possibilities.

1. INTRODUCTORY REMARKS
The calculation of turbulent, combusting flows
has received considerable attention in recent years,
as is demonstrated by the papers contained in the
reviews of Bracco (1976), the AGARD Conference on Combustion Modelling (1980), Jones
(1980), and Libby and Williams (1980). As a conCopyright 1982 by The Combustion Institute
Published by Elsevier Science Publishing Co., Inc.
52 Vanderbilf Avenue, New York, NY 10017

sequence of the growth in the related literature,


interest in the abilities and potential abilities of
the various calculation procedures has also increased, and the main purpose of this review is
to describe and appraise their constituent components and, as a result, to provide a basis for the
evaluation of resulting calculations. Attention is
limited to procedures based on the numerical

0010-2180182/10001 +26502.75

W.P. JONES and J. H. WHITELAW

solution of conservation equations in finite-difference form, not because they are clearly most
suitable for design purposes now, but because they
offer very wide ranging possibilities which need to
be carefully and critically observed and evaluated.
A calculation method for practical combustion
systems comprises a number of essential components. Included in these are the conservation equations, numerical solution methods, turbulence
models, and combustion models, and while it is
difficult to isolate these components completely,
they are described separately in Sections 2, 3, 4,
and 5, respectively. Gaseous combustion is assumed in the related discussion since two-phase
flows add considerable complexity and require
separate consideration. Radiation models are relevant to many combusting flows, and particularly
to large furnaces, but are also outside the scope of
this article. The final section attempts, partly by
the presentation and evaluation of published resuits, to indicate the usefulness of results which
can be obtained with presently available procedures.

2. CONSERVATION EQUATIONS
For gas-fueled flames, in the absence of radiation,
the local values of velocity, concentration of
chemical species and enthalpy can be represented
by the following conservation equations:
Momentum

aG

aG
axj

- Oxj

(ov,+ovj

Dt

Dxj P~ Dxj Pr
N

~c - 1

~C,~}
ha ~

~=1

Oxj

~4)

where Pr and Sc are, respectively, Prandtl and


Schmidt numbers, gi is the gravitational acceleration vector, Ui is the velocity vector,'Ca is the
mass fraction of species c~, h is the enthalpy, p is
the pressure, and ora the net mass formation rate
per unit volume through reaction of species a.
The equation of state for an ideal gas, viz.,

p =p

(5)

(6)

and
N

+ Pgt,

(1)

Continuity
_

(3)

0h
0h
P~+P~o~j

h =~

ap o r ,

= 1, N,

Enthalpy

h -- ha(T)

Oxi ~x i

--+--=O,
0t 0x~

Sc ~x~ l + pi~'

and auxiliary thermodynamic data to relate the


enthalpy of each species to temperature, viz.,

out+ w,
p-~ o~ o~j
Op

Chemical species

(2)

C~h~

(7)

0(ffil

complete the formulation.


The reaction mechanism and associated rate
constants must be known to obtain the formation
rate per unit volume. For example, in the idealized
irreversible reaction where species A and B corn-

REACTING TURBULENT FLOWS

bine to form species C, i.e.,

A + B - , c,

(8)

the forward rate constant is, typically, kt = a


exp ( - E / R T ) , and the net formation rates, which
appear in the species mass fraction equations, are
P

r~ = - - - k ~ C a C B ,
MB

i'B = - - -

MA

rc = +

cation is that, while molecular processes are responsible for mixing at a microlevel, the rate at
which they occur is independent of the values of
molecular transport coefficients and indeed of the
precise mechanism whereby molecular diffusion
occurs. Hence the assumption that we make is not
that the actual molecular processes are as implied
in Eqs. (3) and (4) but rather that turbulence is
dominated by inertial effects. Thus, the models
to be described take no explicit account of molecular transport at all and Eqs. (1), (3) and (4) are
rewritten in the more general forms

kfC a G,

PMc
MAMB

au~
au~
p ~
r pUj axj kfCa CB.

ap ar~j.+
aX i "1- aXj Pgt,

(10)

(9)

Equations (3) and (4) imply that the diffusion


of heat and mass can be adequately described by
the laws of Fourier and Fick. For multicomponent
systems the constitutive equations for the diffusion of heat and mass are more complicated since the
species mass flux depends in a complex manner on
the concentration gradients of all substances present and thermal-diffusion (Soret) and diffusionthermometric (Dufour) effects can be important.
Thus, although the set of equations (1)-(4) may
be applied to laminar flames, appreciable error is
in this case likely: Dixon-Lewis (1970), for example, has demonstrated that the thermal diffusional and ordinary diffusional fluxes are of the
same order over much of a flat laminar hydrogenair flame. Some justification for the use of Eqs.
(3) and (4) in turbulent flow is therefore required.
In turbulent flows, fluctuations in transported
quantities such as velocity and mass fraction are
sustained via interaction between the turbulence
and the gradients of mean quantities. These fluctuations have length and time scales of the order
of those of the mean flow and, providing the Reynolds number is large, are many orders of magnitude larger than the fine scales at which the
molecular diffusion becomes significant. The turbulence acts to reduce the scales of the fluctuations
from the largest scales down to the point where
molecular diffusion becomes significant by an inertial process which is rate controlling. The impli-

p -~-

pUs - - = -axj
axz

pSu(p, ~),

(11)

where represents the scalar set Ca and h. In the


case of enthalpy bp/at is omitted on the basis that
the Mach number is low and the volumetric source
pS will then be zero unless radiation is to be considered.
The flow in practical combustion systems is invaribly turbulent, and the temporal and spatial
variations in the dependent variables encompass
such a wide range of time and length scales as to
preclude the direct numerical solution of the
governing equations. The computer storage and
time requirements are well beyond the capacity of
any existing or planned computer.
To circumvent this difficulty, the dependent
variables are decomposed into mean and fluctuating
components, and the resulting equations averaged
to convert them into statistical equations describing the evolution of mean quantities. (In general,
the averaging process should involve ensemble
averaging, but for the stationary flows considered
here this is indistinguishable from time averaging.)
However, as a consequence of the nonlinearity of
the equations, averaging results in a loss of information, so that the equations are no longer closed
and closure assumptions are necessary before solution is possible.
For variable density flows two types of decomposition can be used: either the unweighted form

W.P. JONES and J. H. WHITELAW

commonly used for constant density flows or the


density-weighted decomposition suggestedby Favre
(1969). The unweighted decomposition and averaging are represented by

u=U+u';
where

1 [ to+r

U = lim-

Udt

r -'--~~ T

and

~'=0,

4tO

and density-weighted averaging by


p U = - ~ + p u " =- -flU+ pu",
=

+ pc~' -

+ pey',

where

-~J = lira

1 f to+l"
-

,r..+ oo T

pU dt

and ffff" = 0

to

but

~" 4= O.

With density-weighted averaging, the equations of


continuity and conservation of momentum and a
scalar may be written, for high-turbulence Reynolds numbers, as
a
--

(12)

(~O~) = 0

axt
a

~_
(

a
--

utus)=

axt

axs

u s ),

(13)

a
--~-~xi(Pui (o)

P,~a(~).

(14)

gi

,,-~, ,,

The use of unweighted averaging results in similar


equations but with two important differences.
First, the density-weighted quantities represented
by the tilde (~) over the quantity are replaced by
unweighted-averaged quantities with the overbar.
Second, correlations involving fluctuating density
correlations such as p'u' and ~
appear in the
equations.

The choice of decomposition to be adopted is


to some degree arbitrary though the densityweighted decomposition and averaging process is
to be preferred on the physical grounds that it
provides equations describing the variation of the
mean values of those quantities which are conserved. For example, the application of densityweighted averaging to Eq. (1), which is a statement
of the conservation of momentum per unit volume
in the xl-direction, gives an equation which describes the variation of mean momentum in the x idirection, i.e., pUi = pUi, not ~O i. In addition the
resulting equations are of simpler form and have
terms which are more easily interpreted than those
which arise with unweighted averaging. More detailed discussions of the merits of both types of
averaging are given by Bilger (1976), Jones (1980),
and Libby and Williams (1980).
It should be noted that the solution of densityweighted equations yields density-weighted properties, and instruments may measure unweighted
values, density-weighted values or some other
value. An infinitely small thermocouple, for example, will detect an unweighted record of instantaneous temperature, but some degree of
density weighting is associated with thermocouples of finite size. The extent of this weighting
with 40-/~m-diam thermocouples is unlikely to exceed around 50C (see Attya and Whitelaw, 1981)
compared with differences between weighted and
unweighted values which can exceed 300C. In
contrast, sampling probes measure values very
close to density-weighted concentration, and it is
clear that the ability to calculate both unweighted
and density weighted is desirable. As will be shown,
this becomes possible if a probability-density function formulation is adopted.
If unweighted values of the dependent variables
are required, then the correlations involving fluctuating density must be evaluated: the correlation
p'ui' is needed to evaluate the unweighted mean
velocity, and similarly ~
is required to obtain
the unweighted mean value of ~,~. The former
quantity p'ui' must be modeled, i.e., related to
known quantities, whereas the correlation ~ can
be obtained from the combustion model-at least
for models where the appropriate probability den-

REACTING TURBULENT FLOWS


sity function is known. Alternatively, the quantity
can be determined via direct modeling assumption. However, most concentration measurements, including those with which comparison will
later be drawn, are obtained with sampling probes
which measure close approximations to densityweighted concentrations. Detailed consideration of
correlations of the type ~ is therefore not usually
required.
Considerable progress has been made in devising
"models" which allow the calculation of the Reynolds stress ,oui'u/ and turbulent flux pu/-~ r appearing in Eqs. (13) and (14) and is discussed in
detail in the fourth section. Like Eqs. (13) and
(14) the models are generally devised for high Reynolds-number flows and special care must be exercised in the interpretation of results in situations,
particularly laboratory flames, where the Reynolds
number of the isothermal flow upstream of the
flame may be insufficient in the presence of a temperature rise and associated changes in viscosity and
density to maintain a fully turbulent condition.

3. NUMERICAL SOLUTION METHODS


The uncertainties associated with finite-difference
assumptions and their application are clearly a subject of some controversy. Swithenbank et al.
(1979) and Drummond et al. (1979), for example,
present results for three-dimensional, combusting
flows and make no mention of possible uncertainties associated with numerical assumptions. In
contrast, McDonald (1979) and Jones and McGuirk (1979) raise doubts about these assumptions,
even for two-dimensional flows.
Numerical procedures represent differential
equations by approximate algebraic equations
which become less approximate with decreasing
distance between the node points used to link the
approximate equations. In principle, the number
of nodes can be increased until an asymptote to
the solution to the differential equations is
achieved. In practice, this procedure is limited by
computer storage and the cost of computer run
time. One obvious method of overcoming this difficulty is to use more accurate (or higher-order)
finite-difference approximations. Such methods

5
appear not to have progressed beyond the formative state, and an application of higher-order
schemes to a set of coupled nonlinear (elliptic)
conservation equations remains to be demonstrated.
The accuracy of calculations is dependent on
the numerical scheme and also on the flow and its
boundary conditions. The importance of the flow
is shown, for example, by contrasting the results
of Vlachos and Whitelaw (1979) obtained for the
laminar flow around an axisymmetric constriction
in a pipe with those of Green and Whitelaw (1980)
for an axisymmetric recirculation of a combustor.
In the pipe flow, calculations with 12 7 and
22 17 nodes showed a maximum difference of
around 20% of the maximum velocity value; the
velocity variations were monotonic with increase
in number of nodes; and the results with 22 17
nodes were argued to be correct, within error
bounds of a few percent. The same numerical
scheme was used by Green and Whitelaw and the
use of 22 16 nodes led to results which were far
from correct due mainly to the inability of this
mesh to resolve properly a small region of recirculation which greatly influenced the downstream
profile. Calculations with 28 X 30 nodes were
still inadequate until the positions of the mesh
lines were arranged so as to provide a proper resolution of the pressure field. Results obtained with
50 50 nodes were similar, i.e., maximum velocities within 4%, to those obtained with 28 X 30
nodes located to resolve the pressure field. In addition to problems associated with locating grid
nodes to ensure that small regions which exert a
large influence on the overall flow are adequately
resolved, errors can arise directly from the finitedifference approximations themselves and in the
case of "elliptic" flow problems, sometimes from
a failure to procure a sufficiently converged solution. Leonard (1979) has discussed some of these
problems and provided a reminder that lack of
skill in the application of numerical methods can
lead to serious errors.
The numerical schemes used, for example, by
Swithenbank et al., Drummond et al., Jones and
McGuirk, Vlachos and Whitelaw, and Green and
Whitelaw, make use of centered implicit differ-

6
encing and are in principal second-order accurate.
Unfortunately, however, with centered differencing of the convection terms unphysical oscillatory (in space) solutions almost always arise for
cell Peclet numbers P l Us I Axe,[~at greater than
two. For nonlinear problems these inevitably prevent convergence for elliptic flow problems and
instability occurs otherwise. Special procedures
must therefore be adopted to ensure that the
finite-difference equations retain the properties
of the partial differential equations from which
they were derived, e.g., "wiggles" do not develop
in the numerical solution. The procedure most
commonly adopted is to switch to donor cell differencing 1 for the convection terms for cell Peclet numbers greater than two. This is carried out
both locally and on a directional basis, e.g., if in
one finite-difference cell p lUI Ax//at is greater
than two but p l V I Ay//a t is less than two, then
the donor cell differencing is used for the x-direction but centered differencing retained for the
y-direction. In practice this switching is implemented by what amounts to adding a pseudoviscosity of magnitude p lf,~lAx,~/2 to the actual
viscosity (molecular plus turbulent) and as a consequence may result in a "false" or "numerical"
diffusion. Whether or not significant error is thereby introduced depends then on the magnitude of
the other terms in the equation: it is important if
the total, i.e., numerical plus physical diffusion in
the direction in which donor cell differencing is
used makes a significant contribution to the conservation equation.
Numerical diffusion errors are usually most
serious when donor cell differencing is used in situations where the velocity vector is aligned close to
the diagonals of a finite-difference grid. An estimate of its magnitude may be obtained from the
expression suggested by Runchal and Wolfstein
(1969),

W.P. JONES and J. H. WHITELAW


and the grid. In the case of the results of Green
and Whitelaw, for example, Peclet numbers up
to 500 were encountered and, with the 28 X 30
nodes referred to earlier, could not be avoided,
especially since special attention and a concentration of nodes were required to represent a small
but important region of the flow. The magnitude
of/anumetieal ranged from/at X 10- 3 to/a t in the
region of highest Peclet number. Fortunately, in
the regions of high numerical diffusion, the magnitude of the pressure and convective terms in the
momentum equations were much larger than the
diffusion terms.
The preceding paragraphs indicate that errors
can occur in two-dimensional flows due to finitedifference approximations and to the limited number of grid nodes which can be accommodated.
Usually, the errors can be arranged to be small,
but care is needed to ensure that this is so. With
calculations involving three independent variables,
the increased computer storage requirements
mean that results which are truly independent of
the number and distribution of grid nodes can not
always be achieved. The problem is further exacerbated with combusting flows, where a large
number of equations have to be solved. In some
flows, it is likely that the numerical solution
method will limit the accuracy of calculations.

4. TURBULENCE MODEL
The unknown velocity covariance of Eq. (13)-the
Reynolds stress tensor-may be expressed as the
dependent variable of the exact conservation equation:

P a t u,"u/' + -~u~ axk


- - ui"u/'

P'numerieal = 0.36p I UtlAx sin 20,


where I Ui [ is the magnitude of the velocity and
where 0 is the angle between the velocity vector
1 Donor cell differencing involves the use of backward or
forward space differences; which is to be used depends
on the direction of the velocity; e.g., see Roache (1976,
p. 73).

ax~ {
a ' + ut"
+ luj" op

ox,

axj - u j " ax~

ap' [

aa,

J - pu,"u "

-- P-us"uk " axk - -~eu"

(15)

REACTING TURBULENT FLOWS

A related equation can also be derived for the unweighted-averaged correlation, but, in both cases,
unknown correlations arise. This is a manifestation
of a recurrent difficulty prevalent in turbulencethe closure problem-to which a solution can be
attempted by relating higher-order correlations to
terms of a lower order and, thereby, closing the
equation set. In general the triple-velocity correlation, correlations involving fluctuating pressure,
and the viscous "dissipation" ei~ must all be approximated in terms of lower-order "known"
quantities.
Turbulence models based on direct closure of
the equations for the components of the stress
tensor have previously been proposed in the context of isothermal flows see, for example, Launder,
1975; Bradshaw, 1976; Lumley, 1978; and Bradshaw et al., 1981, and their extension to reacting
flows with associated large density variations has
been considered by Jones (1980). However, such
models are of considerable complexity, and for
recirculating flows there are substantial difficulties
in obtaining error free numerical solutions. As a
consequence, second-order stress closure models
must be considered to be largely untested in elliptic flows, and, therefore, at the present time
simpler eddy viscosity-type models, such as that
described below, are to be preferred. An alternative approach would be that of subgrid scale
modeling (see, for example, Reynolds, 1976; and
Schumann et al., 1980, but in view of both its
complexity and early state of development this is
also inappropriate at the present time.
The approach to the turbulence models commonly used for recirculating flows is a form of
the two-equation model introduced by Jones and
Launder (1972). This involves an assumed linear
relation between the Reynolds stress and rate of
strain:

--~
tt
"pU i Uj = ~ i J l

/
-Pk +tat ~Xk I

(16)

For the turbulent flux of scalar quantities a gradi-

ent diffusion model is used, viz.,

tat a~,~

~uy"~" -

(17)

ot axj
The turbulent (or eddy) viscosity is given by
tat = G ~ k 2 / e ,

where k(---- Ui"Ui"/2 ) and e are the turbulence kinetic energy and dissipation rate, respectively, the
values of which are obtained from the solution of
the transport equations:
~

Ok

Oxj
=.~

axj

. tat + .12

ax~

_ p._Ui,,Uj,,

axj

tat O~ O~

(18)

- U0x,
~Uj b

Oxj
tat + ta
ax~

axj
e [ _ - 7 7/ ,, a0, + ut a-~ a~

- C l k pui uj ~xj

-~2 axi Ox~

e2

(19)

-Cz-o k

The assumptions involved in these equations have


been extensively discussed by Jones and Launder
(1972) and by Jones and McGuirk (1980a) and
elsewhere. It is recommended that the constants
be assigned the following values:
Cu = 0.09,

C1 =1.44,

6'2 =1.92;

ok =1.0,

ae =1.30,

at =0.7.

Equations (16)-(19) can also be written in unweighted-average form and, as indicated by Bray

8
(1973) include additional terms which stem from
the fluctuating density and which must be modeled.
The use of Eqs. (16)-(19) involves the assumption
that the approximations which are normally used
for constant density flows and which include the
gradient diffusion model can simply be rewritten
in density-weighted form without any explicit account being taken of density fluctuations, the influence of these being assumed to be entirely described by the use of density-weighted averaging.
These assumptions have not been and are unlikely
to be fully tested in a direct sense, but the present
evidence will support their validity in many cases:
the countergradient diffusion observed by Moss
(1980) is an exception. In the majority of cases
the performance of the density-weighted k - e
model in combusting flows does not appear to be
significantly different to that of the constant density version appropriate to isothermal flows. The
merits of the two-equation model have been appraised for isothermal flows in several papers, for
example, Launder and Spalding (1974), Jones and
McGuirk (1980b), and Ribeiro and Whitelaw
(1980a), and appears to represent a good compromise.
The contrast between the flows of Jones and
McGuirk (1980b) and Ribeiro and Whitelaw
(1980a) is instructive. The trajectory of the jet in
crossflow is controlled largely by pressure forces,
and, in most of the flow, the assumed form of
turbulent diffusion is of secondary importance.
In the coaxial jets of Ribeiro and Whitelaw, however, turbulence assumptions are important, and
those associated with two-equation models and indeed existing Reynolds stress closures were shown
to be incorrect in several respects. That such discrepancies sometimes occur is not entirely surprising, for both Reynolds stress closures and twoequation models involve approximating unknown
terms purely in terms of local turbulence quantities. Implicitly contained within this procedure
is the assumption that departures of the turbulence structure from homogeneity are in some
sense small; an approximation which is undoubtedly
in conflict with observation for some flows. In
spite of this limitation and those associated with
closure of the turbulence energy dissipation rate
equation, which cannot be separately evaluated,

W.P. JONES and J. H. WHITELAW


the many comparisons with experiments which
have been reported suggest that at present the two.
equation model is the optimal choice for most
combusting flows and will limit obtainable accuracy
in a restricted number of circumstances which can
include, for example, near wakes behind bluff
bodies and some swirling flows. It should be remembered that all models necessarily have limits
to their range of applicability, and thus accurate
results cannot be expected under all circumstances; for example, and as discussed by Ribeiro
and Whitelaw (1980b), round jets issuing into
stagnant surrounds are among several turbulencecontrolled flows which are not well represented by
existing models.
In passing it should also be noted that the use
of gradient diffusion models clearly precludes the
possibility of representing so-called countergradient diffusion-that is, turbulent diffusion of,
say, species "up" its mean gradient. Libby and Bray
(1980a, b) have predicted such behavior to be important in premixed planar flames, and Moss
(1980) has observed counter gradient diffusion in
an unconfined premixed flame. At the present
time it is difficult to assess whether or not countergradient diffusion effects have significant influence in practical systems, though present indications are that they do not. Countergradient diffusion can be attributed to the preferential influence
of mean pressure gradient on low- and high-density
gas which is manifest through the appearance of
terms like

p',/'

and

in the exact Favre-averaged Reynolds stress and


turbulent scalar flux transport equations. In premixed flows this process can be described by the
Libby-Bray model. Alternatively, and potentially
more generally, it can be represented via a secondorder closure model, such as that outlined in Jones
(1980), in which transport equations are to be
solved for all the independent second-order mo~.
ments i n ci" U "amg ~u i u j " , u~ i q~,~" , a n.a p- -ur - - iw. nowever, for both approaches further testing and investigation is needed.

REACTING TURBULENT FLOWS


5. COMBUSTION MODELS
The combustion model in essence must provide a
method of evaluating the mean formation rate of
each species present, and in addition allow the calculation of mean fluid temperature and the density of the mixture. Under ideal circumstances we
should like to be able to handle reacting flows
comprising many components and in which reaction may occur via a large number of finite-rate
reaction steps. [For example, the simplified mechanism for the high temperature oxidization of
methane of Bowman (1970) involves 11 species
and 14 reaction steps.] In general this would necessitate the solution of conservation equations for
the mean value of each of the independent species,
which in turn requires the evaluation of the mean
formation rate of each species. Now the formation rates are inevitably highly nonlinear functions
of temperature and species concentrations, and
thus knowledge of the mean values of these latter
quantities is insufficient to allow the evaluation of
mean formation rates. In fact the determination of
mean formation rates represents a major difficulty
in the development of prediction methods for cornbusting flows. The importance of fluctuations in
turbulent flows is well known and must be represented in the evaluation of mean formation rates,
density, and temperature. Thus the reaction rate
for the one-step reaction of Section 1 can not be
accurately evaluated in terms of mean quantities,
and, for example, the equation

9
an additional potential problem, for the detailed
reaction mechanism whereby oxidation occurs is
known only for a few of the simpler hydrocarbons.
In practice, however, this may not present a significant problem. The complexities involved in
evaluating mean formation rates are such that, for
the foreseeable future, the use of only a small
number of finite-rate reaction mechanisms may be
the only practicable recourse.
Fortunately, however, the consideration of only
a small number of reactions does not appear to be
a serious drawback in most practical situations.The
reactions associated with heat release in the hightemperature oxidation of hydrocarbon fuels (and
hydrogen) usually have time scales very short compared with those characteristic of the transport
processes and the assumption of "fast" chemistry
thus provides a reasonable description under many
circumstances. Of course, it is not appropriate under all circumstances. In particular, the calculation
of the formation and emission of pollutants such
as carbon monoxide, unburnt fuel, and nitric
oxide all require consideration of finite-rate chemistry, as do ignition and extinction (blow-out)
phenomena. Nevertheless the "fast" chemistry assumption results in a considerable simplication as
far as modeling is concerned and does allow a reasonably accurate description of many features, including the fields of temperature and concentrations of major species in a wide range of continuous
flow combustion systems.
Diffusion or Unpremixed Flames

f A = ---pk,(T) C A C B

will lead to errors up to three orders of magnitude. The most convenient way to represent
the necessary scalar fluctuations is with a probability density function (p.d.f.). Other approaches
are possible and can be described in terms of p.d.f.
procedures. At the present time it appears that
only p.d.f, transport equation formulations offer
the possibility of handling large numbers of reacting
species, but even here in view of the multidimensionality of the approach, computer storage requirements and run times can be expected to be
very large, and their viability therefore remains to
be demonstrated. For hydrocarbon fuels there is

The "fast" chemistry assumption is invoked most


often in the context of situations where fuel and
oxidant enter the combustion system in separate
streams. Then, if the chemistry is assumed sufficiently fast for all reactions to go to completion
(or equilibrium) as soon as the reactants are mixed,
the thermochemical state of the resulting mixture
may be determined purely in terms of strictly conserved scalar variables. The need to evaluate mean
reaction rates is thereby removed. In principle
the equilibrium composition, temperature, and
fluid density of a gas mixture can be determined
if the elemental mass fractions of all elements present, the pressure, and enthalpy are known. How-

10
ever, in the case of turbulent flames a number of
simplifications are invariably made. With reasonable accuracy it can be assumed that all species
and heat have equal diffusion coefficients, i.e. the
turbulent Prandtl/Schmidt numbers are all equal,
and in many circumstances that the heat loss to
the surrounds is negligible compared with the heat
release. In this situation the instantaneous thermochemical state of the gas is determinable as a nonlinear function of a single strictly conserved (i.e.,
zero source) scalar variable. All the relevant strictly
conserved scalar variables are linearly related, so
the actual one to be used becomes a matter of
taste. Typical choices are the mixture fraction f,
defined here as the mass fraction of fuel both
burnt and unburnt, the elemental mass fraction of
any element present, and the enthalpy or any composite variable made up of these quantities. In the
present context the mixture fraction f will be
taken as the scalar variable to be used; all other
conserved quantities can then be obtained from
h = (h~u -- h a ) f + h a,
Ya = (Yatu -- Yaa)f + Y%,
where Ya is the mass fraction of element a and the
subscripts fu and a refer to the values appropriate
to the fuel and air streams, respectively. The instantaneous composition, temperature, and density
must now be related to the instantaneous value of
f. The simplest method of doing this is to assume
that the reaction proceeds via a single-step irreversible reaction of the form
fuel + oxidant ~ products.
The fast reaction assumption then implies that the
mixture (at any instant) consists of either fuel plus
products or oxidant plus products; i.e., fuel and
oxidant cannot coexist on an instantaneous basis.
A knowledge of f is then sufficient to determine
the (instantaneous) fuel, oxidant, and product
mass fractions, and with the adiabatic flow assumption the temperature may also be calculated.
In the past this simplified model has often been
used in conjunction with simplified thermodynamic
data, e.g., constant specific heats. This is an entirely unnecessary procedure, which brings negligible simplification, virtually no savings in computational effort, and can result in large errors in

W.P. JONES and J. H. WHITELAW


the calculated temperature. An alternative means
of relating the thermochemical state to the mixture fraction is via a chemical equilibrium assumption, i.e., reaction rates are sufficiently fast for the
mixture to be in a state of chemical equilibrium.
With this assumption it is not necessary to specify
a precise reaction mechanism, and the equilibrium
state including composition, temperature, and density can be obtained by minimizing the free energy.
A well-tested and reliable computer program based
on this technique is described by Gordon and McBride (1971). Insofar as major species and temperature are concerned both for the H2/air system
and lean hydrocarbon/air mixtures, there is little
to choose between the single-step irreversible reaction (with accurate thermodynamic data) and the
full equilibrium calculation. However, for hydrocarbon/air mixtures richer than stoichiometric
where large amounts of carbon monoxide may be
formed, differences between the two methods becomes significant: the single-step reaction model
which allows no carbon monoxide to be formed
may generate temperatures up to 200K higher
than the equilibrium temperatures.
Thus far we have described only the means
whereby the instantaneous thermochemical state
at any point may be related to the instantaneous
value of mixture fraction at that point. Because of
the strong nonlinearity of these relations it is necessary to take account of the fluctuations in mixture
fraction which arise in all turbulent diffusion
flames. The most convenient way of achieving this
is via the introduction of the probability density
function for mixture fraction,P(f, xi).
About the simplest and most popular approach
is to specify a two-parameter form of the probability density function in terms of the mean and
variance of f. If Favre averaging is used, then the
values of these quantities are obtained from the
solution of the equations

axj axj

axj

at Oxj

Oxi ot

OXj l

+ 2o, \a j

REACTING TURBULENT FLOWS


where CD has a value of 2.0.
The p.d.f, to be constructed is now a densityweighted function which allows evaluation of both
density-weighted and unweighted mean values.
Density-weighted mean values are given by

fO

11
posed and used. For example, Naguib and Lockwood (1975) proposed and used a clipped Gaussian
distribution:

P(.f, xi)
= L2 e r f c { ~ %

*ff)e(f, xi) af

and unweighted values by

+lerfc

c~ = P

[1 ~(f) P(f, xi) dr.


4

] 8(.t')

1 - - f ]/ 6 ( 1 - - f )
X/~o

[ H ( f ) - / ~ ( f - 1)]

o(f)

The density can be obtained from

off)

(Here stands for any quantity which may be


uniquely related to f, including temperature and
species mass fraction.)
The use of Favre averaging is not always recognized; often unweighted averaging is ostensibly
used, but the correlations involving fluctuating
density are ignored. However, the neglect of density fluctuations simply reduces the unweighted
averaged equations to a form identical to the
density-weighted averaged equations. The solution of these equations would then yield densityweighted values, except that the density is evaluated in a manner inconsistent with Favre averaging,
so that erroneous results are obtained.
Computations have been made using a number
of assumed forms for the probability density functions. For example, the rectangular wave variation
of f with time suggested by Spalding (1971) and
used by Gosman and Lockwood (1973) and Khalil et al. (1975) corresponds to two 6-functions located at f+ and f - :
e l f , x O = a~ ( f - f +) + (1

where H( ) is the Heaviside function and again the


parameters ro and Oo are obtained from the values
of f and ff'z. In this case explicit expressions for
fo and Oo cannot be obtained, and their values
must be obtained ileratively; for computational
purposes this is often done "once and for all"
and the results stored in tabular form. An alternative formulation of the clipped Gaussian p.d.f, was
adopted by Kent and Bilger (1977) whereby the
intermittency was utilized. They write for the p.d.f.
P(f, xi) = [1 - I ( x i ) ] 8 (f) + l(xi)e 1(f, xi).
For a jet diffusion flame an empirical correlation
was used to estimate the intermittancy l(xi): PI( )
is the probability function for the turbulent fluid
for which a clipped Gaussian distribution was
used.
The "clipping" of a Gaussian distribution so
that the probability is finite only in the allowable
range of f is arbitrary, and for this reason unappealing. A distribution free from this procedure is the
/3-probability density function utilized by Richardson et al. (1953), Rhodes, et al. (1974), Jones and
Priddin (1978), and Jones and McGuirk (1980). It
can be written

-a)8 ( f - f - ) ,

where a, f, and f - are determined from the


values of jr and f ' ~ . More realistic specifications of
the probability density functions have been pro-

P('f'xi)

p-l(1

- jOb-x

1
fo f " - l ( 1

, 0<f<

--JOb-1 df

1,

12

W.P. JONES and J. H. WHITELAW

where a and b can be determined explicitly from


the values ofjTand]7"2.
A comparison of calculations, performed with
density-weighted averaged equations and the
above three probability density specifications,
with the measurements of Kent and Bilger (1972)
is given in Jones (1980). In summary, the results
show the double delta function p.d.f, to be unsatisfactory in that it generated radial profiles of
unweighted temperature with two maxima and
near discontinuities in the weighted temperature
profiles. There was little to choose between the
"clipped" Gaussian and/3-p.d.f.s, and in both calculations the measured temperature and species
mass fraction profiles were accurately predicted.
The major species and temperature thus appear to
be relatively insensitive to the precise shape of
the p.d.f, once the mean and variance (its "width")
are determined, with the provision that it be constructed from continuous functions with perhaps
Dirac functions at the bounds.
Pollutant Formation in Diffusion Flames
Nitric oxide is normally present in trace quantities
and has negligible influence on the reactions associated with heat release, the thermodynamic
state of the gas, or the flow. It is generally accepted-at least under fuel-lean and near-stoichiometric conditions-that formation of thermalNO
is described by the extended Zeldovich mechanism:
O + Nz ~ N O + N

(i)

N+O 2~ NO+O,

(ii)

N + OH ~ NO + H.

(iii)

In most combustion systems the NO concentrations will be well below their equilibrium levels,
and if a steady-state approximation is then invoked for the N atom concentration, the instantaneous NO formation rate can be expressed as
MN O
SNo = 2pk~(O - CoCN2.
MoMN2
The rate constant kt ci) is a function of temperature alone, and if the oxygen atom concentrations are assumed to have equilibrium values, then

the instantaneous formation rate can be expressed


as a unique function of the mixture fraction. The
mean formation rate required in the mean NO mass
fraction equation can then be determined from

3No(x~.) =

SNo(f)Pff, xj)df.

Whereas there is some evidence that the above


procedure may be adequate at typical gas turbine
combustor operating pressures and inlet temperatures (e.g., Jones and Priddin, 1978), it does not
produce accurate results at atmospheric conditions.
For example, when applied to H2/air flames it
does not reproduce the "rich shift" of the maximum NO concentration observed by Bilger and
Beck (1975) and gives NO levels substantially different from those measured. In view of the very
strong dependence of NO formation rate on temperature and hence mixture fraction, the p.d.f.
cannot be entirely eliminated as the source of this
discrepancy, but a more popular view is that the
error is associated with the equilibrium oxygen
atom assumption. The approach of free radicals,
such as oxygen atoms, to equilibrium is via relatively
slow three-body recombination reactions. Typically, oxygen atom concentrations about an order
of magnitude greater than equilibrium values can
arise; associated with this will be some deviation
from the equilibrium temperature. A description
of such features is embodied in the two-variable
formalism adopted by Janicka and Kollmann
(1980) for the hydrogen-air flame. The bimolecular reactions of the H2/air system were equilibrated
and, following the work of Dixon-Lewis (1975), a
single combined reaction progress variable r (which
is related to a composite mass fraction variable) was
introduced to describe the three-body reactions.
The thermochemical state of the gas can then be
determined in terms of two variables so that the
joint p.d.f, for mixture fraction and reaction
progress variable r is needed. Janicka and Kollmann assumed f and r to be statistically independent and constructed a p.d.f, in terms of the means
and variances of f and r using a beta function for
mixture fraction and three Dirac delta functions
for the reaction progress variable. Up to about 80
nozzle diameters the resulting NO predictions were
in good agreement with measurements, as were the

REACTING TURBULENT FLOWS


major species and temperatures, though the latter
were little different from those which can be obtained with the fast chemistry-full equilibrium
model. The extension of the Janicka-Kollmann
formalism to hydrocarbon-air systems can be
achieved in a relatively straightforward manner by
defining a reaction progress variable in terms of
the total number of moles/unit mass: only the
three-body reactions contribute to the net source,
i.e., production rate, of this quantity. It is likely,
however, that the specification of a more realistic
form for the joint p.d.f, would be beneficial and
lead to more reliable results.
An alternative method of describing finite rate
effects in diffusion flames is the perturbation
method formulated by Bilger (1980). In this approach mass fractions are separated into their
equilibrium values dependent on the mixture fraction plus a perturbation (or departure) from this
value. Exact equations are then derived for the perturbed mass fractions. However, the averaged forms
of these equations contain unknown terms comprising the mean formation rate of the mass fraction perturbation and a mixing term associated
with the fine scale fluctuations and related to the
scalar "dissipation" rate. At this stage it is not obvious that anything has been gained, and the usefulness of the approach depends both on the provision of an adequate model for the mixing term
and the formation rates of the perturbations being
well-conditioned, i.e. approximately linear, functions of the independent variables. Bilger has argued that this is indeed the case for photochemical
smog reactions and for the two variable formulation of the H2/air system adopted by Janicka and
Kollmann, and in the latter situation that it may
be possible to calculate nitric oxide concentrations without resort to a joint p.d.f. However,
both situations involve only trace species and
small departures from equilibrium temperatures.
It is as yet unclear whether the approach will yield
useful results for appreciable departures of major
species, such as CO and unburnt hydrocarbons
from equilibrium.
Premixed Flames
Premixed turbulent flames, in contrast to diffusion
flames, require the evaluation of mean reaction rate.

13
For this reason, and perhaps also because perfectly
premixed flames occur less frequently in practical
systems, they appear to have received less attention than have diffusion flames. So far most work
on premixed flames has assumed that combustion
can be characterized by a global single step reaction
of the type
fuel+ oxidant ~ product
with, say, instantaneous product formation rate
given by an expression of the form

Sp, =AC~Co~ exp(-Ta/T),


where both A and TA are (known) constants. It is
the evaluation of the mean (time averaged) value
of this formation rate which presents problems. If
fluctuations are neglected and the rate evaluated in
terms of the mean values of temperature and mass
fraction, the result can be in error by typically one
order of magnitude and will exhibit a strong dependence on temperature, pressure, and mixture
strength. In contrast, experimental results for premixed turbulent flames are only weakly dependent
on temperature, pressure, and mixture strength.
This fact led Spalding (1971) to propose the eddybreakup model, which, for the mean formation
rate of product, gives
C

and which is based on the idea that the mean reaction rate is determined solely by the rate of scale
reduction via a process of turbulence vortex stretching. Thus the model takes no explicit account of
chemical kinetics and relates to combustion which
is entirely mixed controlled. In this situation it has
been shown to be in good accord with the available
evidence for premixed flames, though it must be
noted that, in the original formulation and many
subsequent applications, the fuel mass fraction
variance was obtained from a modeled transport
equation in which a potentially large term involving
the fuel burnup rate was unjustifiably ignored. The
model has also been used occasionally for nonpremixed flames: it is inappropriate in this situation
and Pope (1977) has shown that it does not then
necessarily provide unique solutions.

14

W.P. JONES and J. H. WHITELAW

A firmer basis for the eddy-breakup model and


its limitations is provided by the Bray,Moss (1977)
model (see also Bray, 1980) for premixed flames.
For reaction via a single global step and adiabatic
flow a single reaction progress variable, defined as
the ratio of product mass fraction to its fully burnt
value, is introduced and is sufficient to determine
the (instantaneous) composition, temperature, and
density of the mixture. The probability density
function for the reaction progress variable r is then
assumed to have the form

P(r, xj) = ot(xj~ (r) + ~(xi)5(1 -- r)


+ [H(r)--H(r-- 1)] 7(xj)F(r, xy).
In the limit of large DamkShler and Reynolds numbers, i.e., "y ,< 1, the mean product formation rate is
determined to be

Sp" = Ca

r-) - CR e

'

where CR is a constant dependent on the continuous part F( ) of the p.d.f. The burning mode part
of the p.d.f.F( ) can be determined from a laminar
flamelet description though results are found to be
relatively insensitive to the shape chose (see, e.g.,
Bray 1978). For small and intermediate Damkohler
numbers this is unlikely to be the case and the
shape of the burning mode part of the p.d.f, will
be of greater importance, though its form can still
be determined from a laminar flamelet model. The
model has also been extended to cover more complex reaction mechanisms, though the assumption
invoked here is that the reaction steps proceed sequentially. Champion et al. (1978) have adopted
this latter approach to calculate a one-dimensional
premixed propane/air flame. Constant specific
heats were assumed and density-weighted averaging
used in conjunction with a combustion mechanism
involving two sequential stages. In the first stagethe delay zone-propane was taken to combine
with oxygen to form hydrogen and carbon monoxide via a single global reaction, the rate expression used being that of Edelman and Fortune
(1969). The second stage of the reaction involving
the oxidation of hydrogen and carbon mon-

oxide-the combustion zone-was the assumed to


commence only after the complete disappearance
of propane. In the combustion zone to simplify
the reaction mechanism a partial equilibrium assumption for the fast bimolecular reactions was
invoked and the ratio of oxygen to hydrogen atom
radical concentrations was assumed constant. This
latter assumption was necessitated by the desire to
describe the state of the reacting mixture in terms
of a single reaction progress variable, namely, the
temperature. For the delay zone, calculations were
presented using two models: the method suggested
by Borghi (1974) whereby the averaged species
mass formation rate is evaluated by an averaged
Taylor series expansion truncated at second order;
and a model based on an assumed p.d.f, for the
reaction progress variable. The p.d.f, used comprised a Dirac delta function at the equilibrium
temperature plus a continuous contribution for
which both rectangular and triangular forms were
utilized. As perhaps might be anticipated-the delay zone is a region of small heat release and therefore small temperature fluctuation-there were only
small differences between the two methods, and
the results appear also to be relatively insensitive
to the p.d.f, shape. The combustion zone is a region with potentially large fluctuations in temperature, and the truncated series expansion technique is not applicable. The presumed p.d.f, formalism is not so restricted, and the results obtained
with this model show a large and plausible influence of temperature fluctuations on the mean temperature in the combustion zone. However, no
comparison with experiment is presented presumably because of the nonavailability of suitable
measurements.
The Bray-Moss model for premixed combustion has been extended by Libby and Bray (1980a,
b), who utilize the concept of laminar flamelets in
order to derive models for turbulent transport and
dissipative processes in one-dimensional planar
flames. This is achieved through the introduction
of a joint p.d.f. P(u, r) for the reaction progress
variable and the component of velocity normal to
the averaged position of the turbulent flame front.
The fast reaction assumption is invoked and the
joint p.d.f, can then be separated into two conditioned p.d,f.s corresponding to reactant and

REACTING TURBULENT FLOWS


product nodes, the probability of events in the
burning mode being negligible. The model leads to
a description of premixed combustion in terms of
four quantities, namely, the means and variances
of the two conditioned velocities. Rapid distortion
theory is used to eliminate one of these quantities,
and the complete model can then be specified in
terms of three quantities, the mean velocity Uand
the fluxes -flu"c" ano- -pu
- " 2 , the values of which are
to be obtained from the solutions of their respective transport equations. The closure of these equations is effected by a laminar flamelet model for
the dissipative terms appearing therein and, in our
view, the rather implausible neglect of quantities
involving fluctuating pressure-the pressure redistribution and pressure scrambling terms (see, e.g.,
Lumley, 1978). An interesting and attractive feature of the approach is that recourse to gradient
diffusion arguments is avoided throughout. However, the extension of the model to multidimensional flows appears not to be straightforward, for
in this situation the joint p.d.f, of the velocity
vector and reaction progress variable is required,
its dimensionality being thereby substantially increased. The model has been applied to the calculation of planar flames by Bray et al. (1981) and
the results compared with the measurements of
Moss (1980) in an open burner premixed flame. In
the light of Moss's experiments the model was revised, with the rapid distortion aspect being replaced by an assumption relating the difference
between the conditioned variances to the difference in the conditioned mean velocities: the various empirical constants appearing in the model
were then assigned values on the basis of the measurements-though not with any rigorous optimization. Although the flame studied by Moss was not
exactly planar and there exists, therefore, some
uncertainty in interpretation of the data for comparison with the theory, the computations and
measurements display an extremely encouraging
and plausible agreement. The results also serve to
illustrate very clearly the importance of mean pressure gradient-turbulence interactions in the case of
planar premixed flames. The generalization of this
approach to multidimensional flows is likely to require considerable effort since the complexity of
the model will increase very significantly. In addi-

15
tion, Reynolds stress production is negligible in
the examples considered so far, with consequent
emphasis on pressure forces.
Models Applicable to Premixed and Diffusion
Flames

The development of general methods for calculating the properties of both premixed and diffusion flames in which finite rate reactions (both
fast and slow) arise is obviously desirable. About
the simplest model which can be considered is the
moment closure method developed by Borghi
(1974). The mean reaction rate was obtained by
decomposing the temperature, density, and mass
fractions appearing therein into mean and fluctuating components, expanding the term via a Taylor series expansion, and then averaging. The resulting infinite series involving moments of all
orders was then truncated through the neglect of
moments higher than second order. For a simple
second-order reaction of the form
S = ApCaC# exp ( - T A / T )

with conventional decomposition averaging the


result is
= A~ exp ( - - T A / T ) I C ~ G

+2
+ a4\

T2 /

1'

+aat, r +
+

p'T'
1 + al - -

pT

-P
+'"'

where

al = TAI~',

az= TAI~(TA--I~,

TAas = T Ca,

and

\2rl

TA
a4 = T ~a

The expansion is valid provided I T' I/T < 1. In itself this is not particularly restrictive. More serious,
however, is the fact that the series converges very

16
slowly unless TA/T is small; in practice this is a
severe restriction since, TA/T > l for most flame
reactions. Thus the neglect of correlations greater
than second order is in general unsatisfactory,
whereas their retention renders the approach intractable.
An alternative approach is to presume a form
for the joint probability density function for the
appropriate number of scalars in terms of their
means, variances, and covariances. The values of
these latter quantities then have to be obtained
from solution of transport equations for which
closure assumptions are required. The "typical
eddy" model of Donaldson (1975) can be viewed
as an attempt to construct a joint p.d.f, from Dirac
delta functions at fixed locations in composition
space. There exists some arbirariness in specifying
the locations of the delta functions, and it is likely
this will be reflected in predictions: mean reaction
rates are likely to be sensitive to the precise "typical eddy" treatment particularly where the chemistry is fast. A possible method of overcoming this,
but with some considerable increase in complexity,
is the construction of joint p.d.f.s from continuous
functions, for example, exponential functions, as
suggested by Jones (1980). However, the presumed
shape joint p.d.f, formalism suffers from the disadvantage that the number of moment equations
increases very rapidly with the number of independent species present. For example, a system
which can be described instantaneously in terms of
N species requires the solution of N(3 + At)/2 transport equations for the moments. The approach is
thus only practicable for simple kinetic schemes
involving only a small number (say up to three) of
independent scalars.

Direct Calculation of p.d.f.s


Rather than specify or construct a p.d.f., a potentiaUy more powerful technique is to obtain
the joint p.d.f, from its transport equation. The
most convenient technique for deriving such equations is that employed by Lundgren (1967)
whereby a fine grained density is utilized in combination with the (instantaneous) conservation
equations to derive the following exact form of

W.P. JONES and J. H. WHITELAW


the single point joint p.d.f, transport equation:

0:=1
-

(~<u/'b>)

axj
(20)
0:=1 18=1 ~ 0 : ~ / 3

19"~X.i

~Xj /

where
N

P=H

8(~a -- 0:) is the fine grained pdf and

ot=l

and P(xi, t, ~0:) with a = 1, "-, N is the density


weighted single-point joint p.d.f, for a set of N
scalar quantities. For a more detailed derivation
and a discussion of the properties of p.d.f, transport equations the reader is referred to O'Brien
(1980).
All the terms on the left side of Eq. (20) are
closed and thus do not require approximation;
included in these are the terms describing chemical reaction, this being the major attraction and
advantage of p.d.f, transport equations for calculating reacting flows. However, this advantage
is not gained without penalty. The dimensionality
of the problem is substantially increased, for Eq.
(20) has an independent variables xi, t, and 4)0:;
a = 1, 2, 3, "', N, where N is the number of independent scalars. In addition, the terms on the
right side of Eq. (20), for which closure assumptions are required, are not well understood, and
their modelling appears to be more difficult than
is the case for, say, second-order moment closures.
This is particularly true for the final term in Eq.
(20), which represents molecular mixing (in ~
space), and it is easily shown (e.g., Pope, 1976);
and Kolmann and Janicka, 1980) that a satisfactory
modeling purely in terms of local (in ~,~ space)

REACTING TURBULENT FLOWS


quantities is not possible; an integral formulation is required. Nevertheless, and in spite of these
difficulties, the approach does offer substantial
advantages for reacting flows, and some progress
has already been made in both devising closure
approximations and solution techniques.
As mentioned above, the major modeling difficulty arises from the molecular mixing term, and
for this a number of proposals have been made.
The simplest of these is the quasi-Gaussian approximation of Dopazo and O'Brien (1974). This
can be expected to work well for near-Gaussian
p.d.f.s but in general has severe limitations. For example, in a homogeneous turbulence field two
initially unmixed species (corresponding to a p.d.f.
consisting of two delta functions, one on either
side of the mean) will remain unmixed during the
subsequent evolution, the p.d.f, retains its initial
shape, and the two delta functions simply "move"
toward the mean. In order to overcome these deficiencies a number of alternatives have been proposed. For a single scalar p.d.f. Pope (1976), using
the mathematical constraints which the mixing
term must satisfy for realizability as a guide, suggested the approximation

o \\axi/

f l g(", 6)P(") de",

where C is a constant and g( ) is a continuous function which was specified on a purely intuitive basis.
Pope (1976) obtained qualitatively plausible results for a reacting homogeneous flow, though no
comparisons with experiment have been reported.
Numerical difficulties in obtaining finite-difference
solutions using the model have been reported by
Janicka et al. (1978). The only other models proposed for the mixing term involve applications and
extensions of the model of Curl (1963) for dispersed phase mixing to turbulent flow. Applied

17
to the mixing term Curl's model can be written

a6,~a6~

=G ~

e(4,~ -

o a~ a~.

2~

~')

~(6,~ + ,,~')

de' -

e(4,~)],

where dO = d1 doe "'" drbN and CO is an empirical


constant. While the above model satisfies all the
necessary realizability conditions, it does have two
significant deficiencies. First, if the model is used
to describe the effects of mixing on a p.d.f, consisting initially of two Dirac delta functions, then
in the subsequent evolution and in the absence of
other influences an increasing number of delta
functions develop and it is only after infinite time
that a smooth distribution evolves (see, e.g. Janicka
et al., 1979). Such behavior is obviously a qualitatively incorrect representation of molecular diffusion. Since a two-delta function p.d.f, corresponds
to what would be found in the initial region of a
mixing layer, this deficiency is likely to have severe
consequences. In practice the effects are likely to
be partly hidden by "smearing" of the numerical
solution and mitigated by other influences such as
transport (in physical space) and reaction. Second,
it is to be expected that the p.d.f, of an unbounded
strictly conserved scalar quantity, in the absence
of mean strain and mean scalar gradients, will relax to a Gaussian form. There is no formal proof
that this will happen, but it is the most likely resuit. This behavior is not reproduced by Curl's
model, but the consequences of its failure to do so
are difficult to assess.
A model which overcomes the former of the
above deficiencies though not the latter is the generalization of Curl's model suggested by Janicka
et al. (1979) whereby for a single scalar bounded
between 0 and 1 the approximation for the com-

18

W.P. JONES and J. H. WHITELAW

plete mixing term may be written

) aa'

P(4, ,',3

de' -e(4.)],

where the function P(~, @', ") is a transition probability. A form was suggested for P ( ) , though recently KoUmann and Janicka (1981) have reported
that the model is relatively insensitive to the precise form chosen provided that C~, is adjusted to the
same rate of decay of the variance of under
homogeneous conditions. In a further development
of the model Kollmann and Janicka decomposed
the joint p.d.f, into an atomic part, represented by
delta functions, plus a continuous part and derived
transport equations for each part. The separation
of the p.d.f, allows the intermittency to be determined and thus turbulent and nonturbulent fluid
to be identified. The mixing model was then applied to the turbulent fluid only for the continuous part of the p.d.f, though there are additional
contributions to molecular mixing involving the
atomic parts for which further assumptions were
invoked. For more details the reader is referred to
the original paper.
For the term representing turbulent transport
of probability only two closure assumptions appear
to have been suggested. For the case of a single
scalar Dopazo (1975) writes the turbulent transport term in the form

<u/'~ = E(u~ 1 4)P(~)


and then introduces the assumption that the conditional p.d.f, is Gaussian. The result then is
u/'"

iSThe model can be expected to work satisfactorily


when the velocity and scalar fields are near-Gaussian, but in other cases its adequacy is difficult to
assess. The only other model which seems to have

been proposed is a generalization of the well-known


eddy viscosity-gradient diffusion model. In terms
of the k-e turbulence model this can be expressed
as

at

OXj

where at is a constant turbulent Prandtl/Schmidt


number.
Because of the complex nature of p.d.f, transport equation models, solutions have been obtained only for a few simple flows. Janicka et al.
(1978), using their extended version of Curl's
model and gradient diffusion, obtained finite-difference solutions for the p.d.f, of a single strictly
conserved scalar for the H2/air flame. As might be
expected from measurements in inert flows, bimodal p.d.f, shapes were calculated and mean concentration and temperature profiles well predicted,
though in the latter case these were not noticeably
better than could be obtained with the presumed
p.d.f, method. More recently Kollmann and Janicka
( 1981) have compared finite-difference solutions of
a single passive (i.e., strictly conserved) scalar in a
round jet and in plane and axisymmetric mixing
layer flows. For the round jet the measurements
and predictions of the p.d.f, were in excellent accord. For the mixing layers good results were also
obtained through some discrepancies here existed.
In part, it is possible that these are associated with
measurement errors near the outer edge of the
plane flow and with the transitional nature of the
axisymmetric mixing layer.
In the case of joint probability density functions of, say, two or more scalar variables it is
dear, because of computer storage requirements,
that finite-difference solutions are not feasible.
To overcome this Pope (1981) devised a Monte
Carlo method for solving multidimensional p.d.f.
transport equations. Only modest computer storage
requirements are claimed for the method, and
computational time, in contrast with finite-difference solutions, increases linearly with the (scalar)
dimensionality of the p.d.f. Using Curl's model
and gradient diffusion, Pope has applied his solution method to the calculation of the p.d.f, of a

REACTING TURBULENT FLOWS


passive scalar in a mixing layer, where the results
were in good agreement with the measurements
of Batt (1977). Computations have also been made
by Pope (1980)of a one.dimensional (in physical
space) premixed turbulent propane/air flame. In
this case the p.d.f, had ostensibly three scalar dimensions-the mass fractions of Call 8, CO and
NO. However, the reaction mechanism used implied a very rapid disappearance of Call8, which
consequently reduced the dimensionality of the
p.d.f, to two over most of the flow, and NO has
virtually no influence on the Call 8 and CO reactions or temperature and could have been computed separately. Nevertheless, and notwithstanding the uncertainties introduced by the one (space)
dimension approximation, the results were in
plausible agreement with the measurements of
Robinson (1974). The Monte Carlo solution
method thus offers considerable promise for
solving joint p.d.f, equations: its success or failure
depends to a certain extent on the type of approximations invoked to close the p.d.f, equation
particularly for the "mixing" term. For example,
it is not easy to see how the method could be applied if the "mixing" approximation of Pope
(1976) were used.
The model of Curl has also found use in the
coalescence-dispersion reactor model developed
by Pratt (1979). Such modular approaches are outside the scope of the present review and are therefore not considered in detail. However So et al.
(1981) have recently combined the coalescencedispersion model with a finite-difference flow field
calculation. The method of combining the two
models and the solution procedure adopted is very
similar to a Monte Carlo solution of a p.d.f, transport equation, though in this case some of the assumptions appear to be (unrealistically) tied to the
purely computational details of a particular finitedifference code. The model has been applied with
reasonable success to a range of inert thin shear
layer flows.
Finally we should mention the so-called
ESKIMO model being developed by Spalding and
coworkers (see Spalding, 1978; andMa et al. 1980).
While this is not strictly speaking a p.d.f, transport
equation method in that no direct calculation of a
p.d.f, is attempted, the model contains sufficient in-

19
formation to allow the construction of the joint
p.d.f. (from solutions) if this is required. The complexity of the model and computational effort required to obtain solutions also appear comparable
with the p.d.f, transport equation methods. In the
ESKIMO model the analysis is separated into two
parts: a biographic part in which the details of reaction and molecular diffusion within "folds" are
treated in an essentially one-dimensional manner:
and a demographic part which involves the specification and description of the "fold" distribution.
The model appears to have been formulated on
largely intuitive grounds, and no use is made of the
exact conservation equation in the derivation. It is
therefore difficult to assess the implications and
any deficiencies of the various assumptions that
are invoked in the model. The mathematical
formulation of the demographic part is at present
under development and poses a number of as yet
unanswered problems.
6. CONCLUDING REMARKS

It is clear from the preceding sections that a general statement to quantify the extent to which
calculation methods can represent turbulent
reacting flow is not meaningful. The models associated with numerical solution techniques, turbulence, heat transfer, and combustion all contribute
uncertainty in proportions and with magnitudes
which depend upon the details of flow. The following paragraphs provide some useful reminders,
mainly in the context of gas turbine combustortype geometries, of present status and future requirements.
The geometry of Fig. 1 corresponds to the reacting flow experiments of Owen (1976). The flow
is axisymmetric and the fuel (methane) and air enter the combustor separately through a central jet
and a surrounding annulus. No swirl component of
velocity is present in either of the incoming central
fuel jet or the annulus air flow. The calculated results were obtained with 53 32-grid nodes and
the computer program of Jones (1975). The form
of the k-e turbulence model quoted in Eqs. (18)
and (19) was used, and further details may be
found in Jones and McGuirk (1980a). The axisymmetric nature of the flow allowed detailed numeri-

20

W.P. JONES and J. H. WHITELAW


1'0-

120

"6-

16'0

- -

4 -

"-~

I P~ed'Ct'o,sl
12"0
JO.O~ ,

~60

--2_.,.
---

0:5

I"0

I, 5

2' 0

x/2r o
Fig. 1. Measurements and predictions of the axial mean velocity contours.

cal tests, and the solution is believed to be independent of the number and position of nodes. It
may be said that the calculated and measured resuits are in general agreement, but it is clear that
quantitative differences exist. In particular, the
end of the recirculation region is calculated to be
downstream of the measurement, and the subsequent velocity rise is calculated to proceed at too
slow a rate, this latter behavior being typical of
flows with central regions of recirculation and
downstream recovery (for example, bluff body
flows). Lest it be thought that these flow field

errors are a consequence of reaction-variable density influences it should be noted that discrepancies
of a similar nature and magnitude are found in
constant density flows.
The result emphasizes the need for improvements in turbulence models. Existing higherorder models would probably not provide a significant improvement. It should be remembered, however, that large regions of the flow field in practical
combustors are pressure rather than turbulentdiffusion dominated, so that, even though improvements to the turbulence model are essential, the

1'0"

.8.--I

1200

----'----1600

'6'4."2-'2-./~.

--~000

Meosuremenfs I

---6-

_;:7-I

1:0

2:0

3:0
x/2r o

Fig. 2. Measurements and predictions of temperature contours.

/~.0

5:0

REACTING TURBULENT FLOWS


results of Fig. 1 may over emphasize the problem.
In contrast to practical systems, the flow pattern
in this axisymmetric combustor is more or less
completely determined by turbulent transport.
The temperature contours of Fig. 2 were obtained in the same configuration as that of Fig. 1,
and the calculated results are again believed to be
free of numerical errors. The fuel was methane gas
injected in the core flow and the combustion
model used was the chemical equilibrium with
/3-p.d.f. for mixture fraction model. The results are
strongly affected by the limitations of the turbulent transport model, particularly that for the diffusion of mixture fraction. The discrepancies between the measured and predicted temperatures
can be partially attributed to errors in the velocity
field, as discussed above, but mainly to insufficient
predicted radial turbulent diffusion of mixture
fraction, with the result that the mixing of fuel
and air proceeds at too slow a rate. This behavior
is not alleviated by the simple (but unjustified)
expedient of reducing the turbulent Prandtl/
Schmidt number to a value even as low as 0.2. An
additional consequence of the incorrectly calculated fuel/air mixing is that the CO and CO 2 profiles are also in error. However, insofar as it is possible to estimate, if the predicted mixture fraction
field could be brought into agreement with measurements, then so also would be the CO and CO 2
levels, though there may be some errors in fuel-rich
regions.
Figure 3 presents measurements and calculations reported by Green (1981) for the flow of
water in a five-hole arrangement intended to represent some of the features of a combustor. The
flow enters radially inward through four round
holes equally spaced around the circumference and
one diameter downstream of the head of the device and axially through a single hole located on
the center line, the flow through this latter hole
being deflected radially outward by a circular disk,
as is shown. These three-dimensional calculations
made useof23 11 8 nodes and cannot be said
to be entirely grid independent, although the
limited amount of testing carried out by Green
suggests that large errors do not exist. The two
sets of results are in good agreement and better
than those of Fig. 1 due to the reduction in im-

21
portance of the turbulence model in this flow.
Velocity results of this quality are expensive of
time and money and require considerable attention to detail. It is clear, however, that they can
be achieved and that the deficiencies of simpler
geometries may not be of such major importance
as the greater pressure control of real combustor
flows. Major difficulties do, however, exist in correctly representing small regions of flow which
have a large influence on their surroundings, and
the incorrect representation of the diffusion of
scalar quantities, although less important than
in Fig. 2, can still lead to large errors. In addition,
it is important to remember that adequate numerical testing is always difficult and sometimes impossible for three-dimensional flows.
No results are presented for three-dimensional
flows with combustion. Very few such calculations
have been made, though the results of Jones and
Priddin (1978) and Mongia et al. (1979) represent
early attempts to calculate practical gas turbine
combustion chamber flows. The major reason for
this situation appears to be because of the difficulty of handling the complicated and arbitrary
geometrical shapes of practical systems, the problem being that the boundaries do not in general lie
along coordinate lines. There exist a number of
possible methods of representing arbitrary shapes,
the most general of which is the use of a general
(numerically generated) boundary-fitted curvilinear
coordinate system such as described in some detail
by Thompson (1980). While the technique is potentially powerful, it has to be said that the form
of conservation equations written in general coordinates can be exceedingly complicated. Alternative methods include finite element methods,
the use of a curvilinear orthogonal coordinate system-though its generation in three dimensions appears difficult-and cartesian coordinates with
piecewise linear interpolation of the boundary
surface between mesh lines and a special finitedifference formulation for node points adjacent
to the boundary. This last method, though conceptually straightforward, can be tedious to implement in practice. While many workers are actively pursuing the development of methods such
as those outlined above, for arbitrary geometries
none appear to have yet reached the stage where

22

W.P. JONES and J. H. WHITELAW

"d
'3O

o=

"1

t./'3
C"q

o
8

>o

.o
o
e~

o=

REACTING TURBULENT FLOWS


they can be used with ease for practical combustion systems.
Liquid fuels have not been considered so far
but are used in many practical systems. For these,
one of two simple approaches can be adopted.
Either a droplet-tracking model such as that used
by for example, E1 Banhawy and Whitelaw (1980)
or a model based on solution of the droplet size
p.d.f, equation as used by, for example, Jones and
Priddin (1978) can be adopted. A comparison of
these two models for gas turbine-type air inlet
temperatures and spray droplet sizes (Jones and
McGuirk, 1981) indicates that both models produce results of comparable accuracy and require
very similar computational run times, with one of
the few major differences being that the droplettracking model is very sensitive to the initial
droplet trajectory whereas the p.d.f, model is not.
Indications are that either of these two simple
models, while both have some difficiencies, particulafly near the fuel injection point, will probably
suffice and will not limit the accuracy of calculations at the present time.
Although a general statement of the type referred to at the beginning of this section cannot usefully be given, some guidelines for the choice of
combustion models can be provided. In continuous flow systems with fuel and air injected separately, it is probable that a p.d.f, in terms of a
single strictly conserved variable will provide an
adequate description of the heat release, temperature, and concentrations of major species. This
may also be true of premixed systems if the single
variable is taken as a reaction progress variable, but
to date no computations have been made to support the possibility. For blow-out and relight of
gas turbine combustors, for internal combustion
engines, and the calculation of pollutant formation and emission levels, fast-reaction assumptions are inappropriate and the consideration of
finite-rate chemistry is required, as described in
Section 5; this requires p.d.f.s with at least two
scalar dimensions, and if many reactions have to
be considered, the solution of p.d.f, transport
equations or their equivalent will be necessary;
where the p.d.f, has two scalar dimensions, as
might be used in the calculation of carbon mon-

23
oxide emission levels, a presumed form of p.d.f.
may represent a possible approach.

NOMENCLATURE
mass fraction of species a
ca
C~,,Cl, C2, CD constants in turbulence model
conditional expected value of u
E(u I)
f
gi

h
ji(~)

k
Ms
N
p
P( )
Pr

R
ra
Sa( )
Sc

t
T
Ui

ui
xi

5( )
~i~
e
/~
p
ot, ak, oe
7ij

4~
4)"
( )

given
mixture fraction; mass fraction of
fuel, burnt and unburnt
gravitational acceleration vector
specific enthalpy (including enthalpies of formation)
flux of scalar q~ arising from molecular transport
turbulence kinetic energy
molecular weight of species a
number of chemical species
static pressure
probability density function
molecular Prandtl number
universal gas constant
net formation rate of c~ per unit
mass
source term in scalar conservation
equation
molecular Schmidt number
time
temperature
velocity vector
fluctuating component of velocity
vector
coordinate in direction i
Dirac delta function
Kronecker delta
turbulence kinetic energy dissipation rate
molecular viscosity
fluid density
turbulent Prandtl/Schmidt number
for q~, k, and e
viscous stress tensor
generalized scalar quantity
fluctuating component of scalar
quantity
ensemble or time average quantity

24

W . P . J O N E S and J. H. W H I T E L A W

Overbars and Superscripts


scalar c o n f i g u r a t i o n space
c o n v e n t i o n a l l y averaged q u a n t i t y
density-weighted average q u a n t i t y

Subscripts
t
a

turbulent
species

REFERENCES
AGARD (1980). Combustor Modelling. Conference Proceedings No. 275.
A. M. Attya and J. H. Whitelaw (1981). Velocity, Temperature and Species Concentration in Unconfined
Kerosene Spray Flames, A.S.M.E. Paper 81-WA/HTD.
R. G. Batt (1977). Turbulent Mixing of Passive and
Chemically Reacting Species in a Low Speed Shear
Layer, J. FluidMech. 82:53.
R. W. Bilger (1976). Turbulent Jet Diffusion Flames,
Progr. Energy Comb. ScL 1:87.
R. W. Bilger (1980). Perturbation Analysis of Turbulent,
Non-Premixed Combustion, Comb. ScL Tech. 22:251.
R. W. Bilger and R. E. Beck (1975). Further Experiments
in Turbulent Jet Diffusion Flames, X V Symposium
(Int) on Combustion, p. 541.
R. Borghi (1974). Chemical Calculations in Turbulent
Flows-Application to CO Containing Turbojet
Flame, Adv. Geophysics 18B: 349.
C. T. Bowman (1970). An Experimental and Analytical
Investigation of the High Temperature Oxidation
Mechanisms of Hydrocarbon Fuels. Combustion,
Science and Technology, 2:161.
F. V. Bracco (1976). Reacting Turbulent Flows, Special
Issue of Comb. Sci. Tech. 13.
P. Bradshaw (1976). Topics in Applied Physics, Vol. 12,
Turbulence, Springer-Verlag, New York.
P. Bradshaw, T. Cebeci, and J. H. Whitelaw (1981). Calculation Methods for Turbulent Flows, Academic
Press, New York.
K. N. C. Bray (1973). Equations of Turbulent Combustion I. Fundamental Equations of Reacting Turbulent
Flow, Univ. Southampton AASU Rept. No. 330.
K. N. C. Bray (1978). The Interaction between Turbulence and Combustion, XVI1 Symposium (InO on
Combustion, p. 223.
K. N. C. Bray (1980). Turbulent flows with premixed
reactants, in Turbulent Reacting Flows. Topics in
Applied Physics (P. A. Libby and F. A. Williams,
Eds.), Springer-Verlag, New York.
K. N. C. Bray and J. B. Moss (1977). A Unified Statistical Model of the Premixed Turbulent Flame, Acta
Astronaut. 4: 291.

K. N. C. Bray, P. A.. Libby, G. Masuya, and Moss (1981).


Turbulent Production in Premixed Turbulent Flames,
Combustion, Science and Technology 25:127.
M. Champion, K. N. C. Bray, and J. B. Moss (1978), The
Turbulent Combustion of a Propane-Air Mixture,
Acta Astronautica 5:1063.
R. Curl (1963). Disposed Phase Mixing: 1. Theory and
Effects in Simple Reactions, A.1. Ch.E. Journal 9:175.
G. Dixon-Lewis (1970). Flame Structure and Flame Reaction Kinetics, Prec. R. Soc. London A 317:227.
G. Dixon-Lewis, J. B. Greenberg, and F. A. Goldsworthy
(1975). Reactions in The Recombination Region of
Hydrogen and Lean Hydrocarbon Flame, XVth Symposium (InU on Combustion, p. 717.
C. du P. Donald (1975). On the Modelling of the Scalar
Correlations Necessary to Construct a Second-Order
Closure Description of Turbulent Reacting Flows, in
Turbulent Mixing in Non-Reactive and Reactive Flows
(S. N. B. Murthy, Ed.), Plenum Press, New York.
C. Dopazo (1975). Probability Density Function Approach for a Turbulent Axisymmetric-Heated Jet,
Phys. Fluids 18:397.
C. Dopazo and E. E. O'Brien (1974). An Approach to
the Autoignition of a Turbulent Mixture, Acta Astron.
1:1239.
J. P. Drummond, R. C. Rogers, and J. S. Evans (1980).
Combustor Modelling for Scramjet Engines, AGARD
CP 275.
R. B. Edelman and O. F. Fortune (1969). A Quasi-Global
Chemical Kinetic Model for Finite Rate Combustion
of Hydrocarbon Fuels with Application to Turbulent
Burning and Mixing in Hypersonic Engines and Nozzles, AIAA paper No. 69, 86.
Y. El Banhawy and J. H. Whitelaw (1980). Calculation of
Flow Properties of a Confined Kerosene Spray Flame,
A I A A J. 18:1503.
A. Favre (1965). Equations des gaz turbulents compressibles, J. Mechanique 4:361.
S. Gordon and B. J. McBride (1971). Computer Program
for Calculation of Complex Equilibrium Composition,
Rocket Performance, Incident and Reflected Shocks
and Chapman-Jouguet Detonations, NASA SP-273.
A. D. Gosman and F. C. Lockwood (1973). Prediction of
the Influence of Turbulent Fluctuations on Flow and
Heat Transfer in Furnaces, Prec. Heat Mass Transfer
Seminar, Int. Summer School, Trogir. Yugoslavia 4:
215.
A. Green (1981). Isothermal Models of Combustion
Chamber Flows, Ph.D. Thesis, University of London.
A. Green and J. H. Whitelaw (1980). Measurements and
Calculations of the Isothermal Flow in Axisymmetric
Models of Combustor Geometries, J. Mech. Eng. Sci.
22:119.
J. Janicka and W. Kollmann (1979). Prediction model for
the p.d.f, of Turbulent Temperature Fluctuations in a
Heated Round Jet, Prec. 2nd Syrup. Turb. Shear
Flows, Imperial College 1:7.

REACTING TURBULENT FLOWS


J. Janicka and W. Kollmann (1980). A Prediction Method
for Turbulent Diffusion Flames Including NO Formation, AGARD Proc. No. 275.
J. Janicka, W. Kolbe, and W. Kollmann (1978). The Solution of a p.d.f. Transport Equation for Turbulent
Diffusion Flames, Proc. 26th Heat Transfer and Fluid
Mechanics Institute, 296.
J. Janicka, W. Kolbe, and W. Kollmann (1979). Closure
of the Transport Equation for the Probability Density
Function of Turbulent Scalar Fields, J. Non.Equilibrium Thermo. 4:47.
W. P. Jones (1975). PACE: A Computer Program for Solving Three-Dimensional Flow Problems, Rolls Royce
Limited Report. Not for general circulation.
W. P. Jones (1980). Models for Turbulent Flows with
Variable Density, VKI Lecture Series 1979-2, published by Hemisphere Pub. Corp. inPrediction Methods
for Turbulent Flows (W. Kollmann, Ed.).
W. P. Jones and B. E. Launder (1972). The Prediction of
Laminarisation with a Two Equation Turbulence
Model, Int. J. Heat Mass Transfer 15 :301.
W. P. Jones and J. McGuirk (1980a). Mathematical
Modelling of Gas-Turbine Combustion Chambers,
AGARD CP 275.
W. P. Jones and J. McGuirk (1980b). Computation of a
Round Turbulent Jet Discharging into a Confined
Cross Flow, Turbulent Shear Flows 2:233.
W. P. Jones and J. McGuirk (1981). A Comparison of
Two Droplet Models for Gas Turbine Combustion
Chamber Fiows, Proc. 5th ISABE, Bangalore, lndia.
W. P. Jones and C. Priddin (1978). Prediction of the
Flow Field and Local Gas Composition in Gas Turbine Combustors, X V l l t h Symposium (lnt.) on
Combustion, p. 399.
J. H. Kent and R. W. Bilger (1972). Turbulent Diffusion Flames, University of Sydney, Report 7N F-37.
See also 14th Symposium in Combustion, 1973,
p. 615.
J. H. Kent and R. W. Bilger (1976). The Prediction of
Turbulent Diffusion Flames Fields and Nitric Oxide
Formation, X VISymposium (Int.) on Combustion.
E. E. Khalil, D. B. Spalding, and J. H. Whitelaw (1975).
The Calculation of Local Flow Properties in Two-Dimensional Furnaces,Int. J. Heat Mass Transfer 18:775.
W. Kollmann and J. Janicka (1981). A Transport Equation for the p.d.f, of a Passive Scalar in Turbulent
Shear Flows (to be published).
B. E. Launder (1975). Progress in the Modelling of Turbulent Transport, Lecture Series 76, yon Karman Institute for Fluid Dynamics, Belgium.
B. E. Launder and D. B. Spalding (1974). The Numerical
Computation of Turbulent Flows, Cornputer Methods
in Appl. Mechs. and Engng. 3:269.
B. P. Leonard (1979). A Survey of Finite Differences of
Opinion on Numerical Modelling of the Incomprehensible Defective Confusion Equation, ASME, AMD
Vol. 34.

25
P. A. Libby and K.N.C. Bray (1980a). Counter Gradient
Diffusion in Premixed Turbulent Flames, A1AA Paper
No. 80-0013.
P. A. Libby and K. N. C. Bray (1980b). Implications of
the Laminar Flamelet Model in Premixed Turbulent
Combustion, CombusL Flame 39:33.
P. A. Libby and F. Williams (1980). Turbulent Reacting
Flows, Topics in Applied Physics, Vol. 4, SpringerVerlag, New York.
J. L. Lumley (1978). Computational Modelling of Turbulent Flows, in Advances in Applied Mechanics No. 18,
Academic Press, New York.
T. S. Lundgren (1967). Distribution Function in the Statistical Theory of Turbulence, Phys. Fluids 10: 969.
A. S. C. Ma, M. A. Nosier, and D. B. Spalding (1980). Application of the ESKIMO Theory of Turbulent Combustion, AIAA Paper No. 80-0014.
H. McDonald (1979). Combustion Modelling in Two and
Three Dimensions-Some Numerical Considerations,
Prog. Energy Comb. ScL 5:97.
H. Mongia, R. Reynolds, E. Coleman, and T. Bruce (1979).
USARTC-TR-78-55B Combustor Design Criteria Validation, Volume 2-Development Testing of Two FullScale Annular Gas Turbine Combustors.
J. B. Moss (1980). Simultaneous Measurements of Concentration and Velocity in an Open Premixed Turbulent Flame, Combustion Science and Technology 22:
115.
E. E. O'Brien (1980). The Probability Density Function
(pdf) Approach to Reacting Turbulent Flows, in Turbulent Reacting Flows, Topics in Applied Physics.
(P. A. Libby and F. A. Williams, Eds.), Springer-Verlag, New York.
K. Owen (1976). Measurements and Observations in Turbulent Recirculating Flows, A I A A Journal 14.
S. B. Pope (1976). The Probability Approach to the
Modelling of Turbulent Reacting Flows, Combust.
Flame 27: 299.
S. B. Pope (1977). The Implications of the Probability
Equations for Turbulent Combusting Flows, Combust. Flame 29:235.
S. B. Pope (1980). Monte Carlo Calculations of Premixed
Turbulent Flames, XVlI1th Symposium on Combustion.
S. B. Pope (1981). A Monte Carlo Method for the p.d.f.
Equations of Turbulent Reactive Flow, Combustion
Science and Technology 25:159.
D. T. Pratt (1979). Coalescence Dispersion Modelling
of High Intensity Combustion, J. Energy 3:177.
W. C. Reynolds (1976). Computation of Turbulent Flows,
Ann. Rev. Fluid Mech. 8:183.
R. P. Rhodes, P. T. Harsha, and C. E. Peters (1974). Turbulent Kinetic Energy and Analysis of Hydrogen-Air
Diffusion Flames, Acta Astronautica 1:443.
M. M. Ribeiro and J. H. Whitelaw (1980a). Coaxial Jets
with and without Swirl, J. Fluid Mech. 96:769.

26
M. M. Ribeiro and J. H. Whitelaw (1980b). The Structure
of Turbulent Jets, Prec. Roy. Soc. A 370:281.
J. M. Richardson, H. C. Howard, and R. W. Smith (1953).
The Relation between Sampling-Tube Measurements
and Concentration Fluctuations in a Turbulent Gas
Jet, I V Symposium (Int.) on Combustion, p. 814.
P. J. Roache (1976). Computation Fluid Dynamics,
Hermosa Publishers, Albuquerque, NM.
G. F. Robinson (1974). Pollutant Formation in Turbulent
Flames, Ph.D. thesis, Northwestern University.
A. K. Runchal and M. Wolfshtein (1969). Numerical Integration Procedure for the Steady-State Navier-Stokes
Equations, J. Mech. Eng. Science 11:445.
U. Schumann, G. Grotzbach, and L. Kleiser (1980). Direct Numerical Simulation of Turbulence, VKI Lecture Series 1979-2, Published by Hemisphere Pub.
Corp. in Prediction Methods ]'or Turbulent Flow
(W. Kollmann, Ed.).
R. M. C. So,D.T. Pratt, and M. D. Hooven (1981). Modelling of Scalar Transport in Turbulent Shear Flows (submitted to Chem. Eng. Sc).

W.P. JONES and J. H. WHITELAW


D. B. Spalding (197'1). Concentration Fluctuations in a
Round Turbulent Free Jet, Chem. Eng. Sci. 26:95.
D. B. Spalding (1978). The Influence of Laminar Transport and Chemical Kinetics on the Time-Mean Reaction Rate in Turbulent Flow, XVII Symposium (Int.)
on Combustion, p. 431.
J. Swithenbank, A. Turan, P. G. Felton, and D. B. Spalding (1980). Fundamental Modelling of Mixing, Evaporation and Kinetics in Gas Turbine Combustors, AGARD
CP 275.
J. F. Thompson (1980). Numerical Solution of Flow Problems Using Body-Fitted Co-Ordinate Systems, in Computational Fluid Dynamics, (W. Kollmann, Ed.),
Hemisphere Publ. Corp., New York.
N. S. Vlachos and J. H. Whitelaw (1979). Low Reynolds
Number Flow in the Vicinity of Axisymmetric Constrictions, J. Mech. Eng. Sci. 21:73.

Received 24 August 1981, revised 2 7 November 1981

You might also like