Combustion and Flame 161 (2014) 222–239

Contents lists available at ScienceDirect

Combustion and Flame

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m b u s t fl a m e

Large-eddy simulation of spray combustion in a gas turbine combustor

W.P. Jones ⇑, A.J. Marquis, K. Vogiatzaki
Department of Mechanical Engineering, Imperial College London, Exhibition Road, London SW7 2AZ, UK

a r t i c l e i n f o a b s t r a c t

Article history: The paper describes the results of a comprehensive study of turbulent mixing, fuel spray dispersion and
Received 26 October 2012 evaporation and combustion in a gas-turbine combustor geometry (the DLR Generic Single Sector Com-
Received in revised form 16 July 2013 bustor) with the aid of Large Eddy Simulation (LES). An Eulerian description of the continuous phase is
Accepted 17 July 2013
adopted and is coupled with a Lagrangian formulation of the dispersed phase. The sub-grid scale (sgs)
Available online 19 August 2013
probability density function approach in conjunction with the stochastic fields solution method is used
to account for sgs turbulence-chemistry interactions. Stochastic models are used to represent the influ-
Keywords:
ence of sgs fluctuations on droplet dispersion and evaporation. Two different test cases are simulated
Large eddy simulation
Pdf-methods
involving reacting and non-reacting conditions. The simulations of the underlying flow field are satisfying
Stochastic field method in terms of mean statistics and the structure of the flame is captured accurately. Detailed spray simula-
Spray tions are also presented and compared with measurements where the fuel spray model is shown to
reproduce the measured Sauter Mean Diameter (SMD) and the velocity of the droplets accurately.
Ó 2013 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

1. Introduction jected into a swirling air flow that enhances the mixing between
the fuel and the air and thus the evaporation rate. The combustion
Numerical simulations complement traditional experimental performance and emissions are mainly influenced by the atomisa-
approaches and can provide a valuable aid to the design of com- tion of the liquid fuel, the dispersion and evaporation of the fuel
bustion devices with low emissions and more efficient operation. droplets and the mixing of fuel and air. Thus the accurate predic-
These simulations require accurate models to represent the inter- tion of the spray dynamics and combustion process is extremely
action of turbulence with chemical reactions, and, depending on important to determine flame stability over a wide range of oper-
the complexity of the combustion process additional models for ating conditions, to ensure safe and efficient utilisation of energy,
two-phase flows interactions, radiative heat transfer and soot for- and to better understand the mechanisms of pollutant formation.
mation may be required. To date practical combustion devices such Reynolds Averaged Simulations (RANS) are currently a major
as gas turbines are mostly designed based on the experimental tool for gas turbine combustion chamber designers, but over the
findings of expensive high quality tests. Simulations of similar last few years Large Eddy Simulation (LES) has undergone consid-
geometries could constitute a more economic alternative to exper- erable development and is starting to make a significant contribu-
iments but these are rather limited at the present time. The main tion to the design process. In LES a spatial filter is applied to the
reason for this is that the computational constraints often lead to equations of motion. The large energy-containing motions are then
simplifications such that the simulations are unable to describe computed directly while the effects of the small sub-filter scale
all of the complex and interacting physical phenomena taking motions are modelled. This makes LES an appropriate tool to cap-
place over a wide range of length and time scales. ture the complicated phenomena including unsteady effects pres-
One of the most prominent areas of modern research evolves ent in a practical combustor.
around liquid fuel combustion. Liquid fuels have a high volumetric LES has been applied to the simulation of a wide range of pre-
and mass energy density and are easy to store and transport. In a mixed and non-premixed combustion processes demonstrating
practical combustor liquid fuel is atomised into small droplets in the ability of the method to migrate from an academic to an indus-
order to increase the surface area of fuel exposed to the hot gases trial tool [1–3]. Crucial steps in this migration have been the devel-
and to facilitate rapid vaporisation and mixing with the surround- opment of numerical algorithms that are flexible enough to handle
ing air. In addition to this process, in many devices, the spray is in- complex configurations, yet accurate enough to simulate turbu-
lence and its interaction with the physical and chemical processes
taking place in a plurality of phases.
⇑ Corresponding author. In the present work a coupled Eulerian (for the gas-phase flow)
E-mail address: [email protected] (W.P. Jones). and Lagrangian (for the liquid-phase flow) formulation is used to

0010-2180/$ - see front matter Ó 2013 The Combustion Institute. Published by Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.combustflame.2013.07.016
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 223

Nomenclature

a particle acceleration Sh(d) deterministic part of the Sherwood number

B Spalding mass transfer number Sh(st) stochastic part of the Sherwood number
CD drag coefficient e
S ij filtered rate of strain tensor
Co model constant (Co = 1) T_ rate of change of droplet temperature caused by heat
Cp‘ liquid specific heat capacity transfer from the surrounding gas phase
Cs Smagorinsky parameter uj velocity of the gaseous phase
CV model constant (CV = 1) ep
u filtered gas velocity at the particle position
dt time interval v droplet velocity
dWt increment of the Wiener process vp velocity of the pth particle
FD drag force
Fg gravitational force Greek Symbols
fi momentum exchange between the carrier gas and the C0 total – molecular plus sgs – diffusion coefficient
dispersed flow D filter width
g gravitational acceleration gi [1, 1] dichotomic random vector
h specific enthalpy h droplet temperature
hfg latent heat of evaporation lsgs sgs viscosity
ksgs sgs kinetic energy nna stochastic fields for 1 6 n 6 N, 1 6 a 6 Ns
m_ rate of mass addition to the continuous phase per unit q density of the gaseous phase
volume through droplet evaporation q‘ density of the liquid phase
mp particle mass qp filtered gas density at the particle position
N_ rate of change of droplet number through droplet break- r Schmidt number
up and coalescence rij viscous stress of the gaseous phase
N number of species ssgs
ij
sgs stress
Ns number of scalars (Ns = N + 1) sp particle relaxation time
p pressure of the gaseous phase ssgs sgs mixing time scale
e sgs ðwÞ
P ensemble of N stochastic fields for each of the Ns scalars st sub-grid time scale which determines the rate of inter-
Pspr sgs spray pdf action between the particle and turbulence dynamics
R_ rate of change of the droplet radius through evaporation /a species a
r droplet radius x_ a species a reaction rate
Re Reynolds number based on the droplet diameter
Scg gas phase Schmidt number at the particle position
Sh Sherwood number

represent the spray dynamics and the interaction between the gas ture was measured with Laser-Induced Fluorescence (OH-LIF). A
and liquid phases flow [4–9]. The Eulerian–Lagrangian framework photograph of the set-up used in the experiments can be seen in
represents a natural approach for flows where a dispersed phase is Fig. 1. The computational domain used for the CFD calculations
present. It allows direct modelling of the actual processes that indi- consists of two parts: a cylindrical component (80 mm long) that
vidual droplets undergo (such as break-up, droplet dispersion, and surrounds the two radial air swirlers which are fed by pre-heated
wall interactions) in contrast to the more indirect modelling air from a plenum, and a rectangular combustor (300 mm long)
dependence of these processes on the volume fraction of the drop- with a converging duct at the exit as it is shown in Fig. 2. The fuel
lets or their number density distribution required by alternative (commercial aviation kerosene) used for the combusting test is
Euler–Euler approaches. The restriction however of the adopted supplied by two opposing fuel lines to an annular fuel gallery,
approach is the need for computationally efficient algorithms espe- and from there to a vertical slot through a circular array of 36 ori-
cially when the number of droplets is large. fices that surround the base of a prefilmer (see Fig. 3a). The actual
In this paper LES calculations are presented for two operating pipes, through which the fuel enters, are not included in the com-
conditions of a single sector combustor operating at a pressure of putations. Instead the liquid fuel is injected in the form of droplets
4 bar pressure for which DLR has performed detailed measure- via an annular ring (represented by 1000 discrete locations1 from
ments [10,11]; in particular an isothermal simulation (namely test which droplets are injected at random) at an axial location just
case E), and a reacting simulation (namely test case A). The focus of downstream of the fuel injector prefilmer lip. In addition, the com-
the work here is the assessment of the predictive capability of LES bustor walls have a series of effusion air cooling holes, however as
with sub-grid scale models for spray dispersion and evaporation this occurs downstream of the region of interest the effusion cooling
and subsequent combustion. The emphasis of the work is placed air was omitted from the present analysis. A structured, multi-block
on the effect of the unresolved velocity and temperature fields mesh of the DLR generic combustor was created using ICEM CFD
on the droplet statistics especially in the region close to the injec- v.11, and consisted in total of 2.2 million cells and 137 blocks [12].
tion point. The mesh is refined at the exit of the swirler and the size of the
smallest cells is 0.5 mm (see Figs. 2 and 3). The size of the largest
cells is 5 mm and are located in the middle section of the combustor.
2. Burner geometry The simulations were carried out using 16 cores of a Linux cluster
comprising Intel Xeon E5404 2.0 GHz CPUs.
The configuration under investigation is the DLR Generic Single
Sector Combustor (GENRIG): a detailed description can be found in 1
It would also have been possible to inject droplets at random locations around the
[10,11]. Velocities and droplet sizes of the evaporating sprays were annular ring. However given the uncertainty in spray boundary conditions it is
measured with Phase Doppler Anemometry (PDA) and tempera- doubtful that this would bring any significant changes.
224 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

droplet formation and the dispersed phase is also presumed dilute

so that collisions, coalescence and agglomeration are ignored [13].
It is recognised that very close to the fuel injector this latter
assumption is likely to be violated. In reality, the fuel issues into
the combustor in the form of a thin film, which then undergoes
breakup to form droplets. We do not attempt to model this break-
up process directly but rather instead apply an ansatz of the spray
inlet conditions so as to reproduce the measure profiles at
z = 7 mm. The local errors i.e. those arising in the immediate vicin-
ity of the injection point, associated with this process are greater
than those associated with the assumption of a dilute dispersed
phase.
In the LES calculations the density-weighted filtered Navier–
Fig. 1. Experimental geometry of the DLR generic combustor [10,11]. Stokes equations, with the contribution of the dispersed phase in-
cluded, can be written as:

3. Numerical details @q @qu ej _

þ ¼m ð1Þ
@t @xj
3.1. Mathematical model: Interaction of gaseous and dispersed phase
sgs
e i @qu
@qu ei u
ej  @r
@p  ij @ sij
In this section the pdf approach adopted for the two-phase sim- þ ¼ þ þ þ f i ð2Þ
@t @xj @xi @xj @xj
ulations in the LES context is presented. An Eulerian description of
the continuous phase is employed and coupled with a Lagrangian where rij, q, uj and p represent the viscous stress, density, velocity
approach for the dispersed phase. The two phases are coupled and pressure of the gaseous mixture respectively. The term m _ rep-
through the inclusion of the forces exerted on the droplets by the resents the rate of mass addition to the continuous phase per unit
continuous phase and vice versa and through the rate of phase volume through droplet evaporation and fi is the force per unit vol-
change. The characteristic dimensions of the dispersed phase are ume exerted on the continuous phase by the dispersed phase. The
presumed small compared to the length scales of the smallest re- over bars and tildes represent the spatially filtered and density
solved turbulent motions, enabling the droplets to be viewed as weighted filtered values with a filter width D respectively. The un-
point sources with respect to the continuous phase. Stochastic known sub-grid stresses are approximated using the Smagorinsky
models are introduced to account for the influence of sgs motions model [14], where the deviatoric part of the sub-grid stress is re-
lated to the filtered rate of strain tensor via ssgs e
ij ¼ lsgs S ij with the
on dispersion and evaporation. The effect of particles on turbulence
is poorly understood, with most studies being conducted in the  C s D2 k e
sgs viscosity given by lsgs ¼ q S ij k; ke
S ij k represents the Frobe-
context of Reynolds averaged approaches. In the present work nius norm. The filter width is taken as the cube root of the local grid
the possible modulating effects of sgs motions and sgs dissipation cell volume. The parameter Cs, which is the Smagorinsky parameter,
are ignored. No attempt is made at modelling film breakup and is determined as a function of space and time using the dynamic

Fig. 2. CFD geometry of the DLR generic combustor. The green box indicates the ‘window’ for which experimental data are available. (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3. DLR generic combustor grid at the area of the swirler: (a) Schematics of the burner [10,11]. (b) Slice through injector centre-line showing details of mesh through
swirlers. (c) 3-D picture of the double-swirlers.
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 225

model of [15]. If, as in the present case for simplicity, variations in tional Lagrangian rates of change are the cause of the indetermi-
thermodynamic pressure are neglected, then the equations describ- nacy. In order to solve the modelled form of Eq. (4) it is first
ing species mass fractions and enthalpy have a similar form. Hence, replaced with an equivalent Ito system, [19], of stochastic differen-
the equations for the Ns scalars required to describe combustion can tial equations, (sde’s) describing the trajectories of stochastic parti-
be written: cles in the phase space (V, R, H, N). These stochastic particles do not
! sgs
necessarily represent individual physical droplets but rather a
e a @q/
@q/ eauej @ l @ /e a @J j
þ ¼ þ _ a ð/Þ þ qx
þm _ a ð/Þ ð3Þ group of droplets with identical physical properties (size, velocity,
@t @xj @xj r @xj @xj temperature). These groups of particles serve to increase the accu-
racy of the spray statistics and reduce the computational require-
where a Lewis number of one is assumed so that r is the Prandtl or ments. The particles are inertial particles and follow the Stokes
Schmidt number as appropriate. The filtered source term, m _ a repre-
law stating that the drag force FD(t) exerted by the fluid on the par-
sents the rate at which /a is added to the continuous phase per unit ticles is proportional to the difference between the background
volume, through droplet evaporation and depends on the properties fluid velocity and the particle velocity and consequently the parti-
of both the dispersed and continuous phases. The former depen- cle trajectories follow the system of equations:
dence has been dropped in the interests of compactness. The phase
exchange source terms are obtained from the volume-averaged dxp ¼ v p dt ð5Þ
contributions of the dispersed phase, e.g. m _ ¼ 1 Pn m _
D3 p¼1 ðpÞ , where
the subscript (p) refers to the pth droplet. The presence of the inter- dv p
mp ¼ FD þ Fg ð6Þ
action source terms describes two-way coupling between the con- dt
tinuous and the dispersed phases. During the droplet motion, where the subscript p indicates the pth particle (droplet). The term
droplet-turbulence interactions occur: the droplets are dispersed FD represents the drag force and Fg the gravitational force. Basset
by the turbulence of the continuous phase and the turbulence of forces are ignored since the droplet density (kerosene) is much
the continuous phase is modulated by the presence of droplets. higher than the ambient density (air).
The characteristics of these interactions are captured by the The motion of a stochastic particle in a turbulent flow field can
momentum exchange between the carrier gas and the dispersed be viewed as a random process with position determined by a
flow through fi . However, in the present case the effects of the drop- deterministic part, evaluated in terms of filtered values and a sto-
lets on the turbulence of the continuous phase is small since the chastic component arising from the sgs turbulent motions of the
dispersed phase is dilute. gas phase. In this study only viscous drag and gravitational forces
The evaporation of droplets gives rise to sources of mass and /a are considered and a stochastic Markov model [17,20] is used to
for the continuous phase, (m _ and m
_ a ). When a liquid droplet is re- represent the influence of the unresolved carrier gas velocity fluc-
leased into a high ambient temperature heat transfer occurs. Ini- tuations experienced by a stochastic particle p over a time interval
tially most of the heat transferred to the droplets increases the dt which is added to the deterministic contribution:
droplet temperature (heating period). At some point the tempera-    1=2
ture of the droplet approaches a crucial point (the saturation tem- dv p ¼ s1 e p ðt Þ  v p dt þ q‘  q
U
p
gdt þ C o
ksgs
dWt ð7Þ
p
perature, which for kerosene is 549 K) the evaporation rate q‘ st
increases to its maximum value. where q‘ is the liquid density, vp is the velocity of the pth particle,
The chemical source terms appearing in Eq. (3), x _ a , are highly e p are the filtered gas density and velocity at the particle po-
q p and u
non-linear and cannot therefore be evaluated solely in terms of fil- sition, ksgs is the unresolved kinetic energy of the gas phase, Co is a
tered mean quantities. In the present work a statistical description model constant assigned the value unity, dWt represents the incre-
based on the pdf approach [16] is adopted and this will be dis- ment of the Wiener process and g is the gravitational acceleration.
cussed in Section 3.3. Finally, st is a sub-grid time scale which determines the rate of
interaction between the particle and turbulence dynamics, defined
3.2. Modelling of droplet dispersion and evaporation as [20]:
0 10:6
Following [17,18] the dispersed phase is described in terms of a sp A
set of macroscopic variables: the droplet velocity, v, the droplet ra- st ¼ sp @pDffiffiffiffiffi
ffi
ð8Þ
dius r, the droplet temperature h, and number, n. The required joint ksgs

pdf is Pspr (V, R, H, N, x, t), where (V, R, H, N) is the ‘phase’ space for qf C D
The particle relaxation time, sp is given by: s1
p ¼ 8
3 e
qp R j u p  v p j,
(v, r, h, n), which can be obtained, after suitable modelling, from:
where the drag coefficient CD is obtained from [21]:
    (  
 
@Pspr  _ spr
 @ RP @ T_ Pspr @ N_ Pspr 24 2=3
1 þ Re6 : 0 < Re < 1000
Re
þ rv  aPspr þ þ þ ¼0 ð4Þ CD ¼ ð9Þ
@t @R @H @N 0:424 : Re P 1000
where a; R,_ and T_ represent the particle acceleration, the rate of
where Re is the Reynolds number based on the droplet diameter
change of the droplet radius through evaporation and the rate of
and the relative velocity of the droplet with respect to the gas
change of droplet temperature caused by heat transfer from the sur- phase. The sgs kinetic energy is obtained using equilibrium argu-
rounding gas phase respectively. Finally N_ is the rate of change of e e
ments from ksgs ¼ 2DC 2=3s S ij S ij . If it is assumed that the temperature
droplet number through droplet breakup and coalescence. These within each droplet is uniform and that equilibrium conditions pre-
rates of change terms can be written in the general form:
vail at the surface, [22,23] then the droplet temperature Tp and mass
  
Dwk  mp can be obtained from:
E W ¼ U where U ¼ v ; r; h and n
Dt  lg
dmp ¼ 2pr p Sh lnð1 þ BÞdt ð10Þ
Dwk Scg
which represents the expected value of conditioned upon W = U
Dt
anywhere in the filter volume.
NuCpg ðT g  T p Þ hfg dmp
Equation (4) is an exact unclosed hyperbolic partial differential dT p ¼ dt þ ð11Þ
equation for the joint pdf of the spray in which the filtered condi- 3Pr g Cp‘ sp Cp‘ mp
226 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

where hfg is the latent heat of evaporation, Cp‘ is the liquid specific quantities, appear in a closed form, whereas the molecular trans-
heat, Scg is the gas phase Schmidt number at the particle position, Sh port or micro-mixing must be modelled. Using the filtering opera-
is the Sherwood number, Nu is the Nusselt number, and B is the tion and following Gao and O’Brien [37] the equation describing
Spalding mass transfer number,. The time scale sp is defined as the evolution of the density weighted sub-grid (or more strictly
before. the density weighted filtered fine grained) pdf for the Ns scalars
In the current approach the droplets are considered to have can be derived:
sizes smaller than the filter size and thus their trajectories are af-
fected by both large (resolved) and small (unresolved) scales. If
e sgs ðwÞ @ qu
@qP e sgs ðwÞ qmðwÞ
~j P _
e ðwÞ þ
X
Ns
@ h i
þ  P qx_ a ðwÞ Pe sgs ðwÞ
Eqs. (7), (10) and (11) are evaluated in terms of the filtered quan- @t @xj qðwÞ sgs @w a
2   a¼1 3
tities only, the direct effects of sub-grid velocity fluctuations (small
X @ q ma ðwÞ  mðwÞwa
N s _ _
scales) would not be included. The relative velocity between the þ 4 e sgs ðwÞ5
P
droplets and the flow and the gas-phase temperature and vapour a¼1
@wa qðwÞ
mass fraction are filtered versions of the complete instantaneous @ h  i
fields. It would only be possible for the droplets to ‘feel’ the (indi- ¼ qui  qu~i j/ ¼ w Pe sgs ðwÞ
@xi " ! #
rect) effect of the sub-grid scales through the sub-grid models that XNs X Ns 
@2 l @/a @/b  e sgs ðwÞ
are used for the closure of the filtered momentum and continuity  / ¼ w P ð13Þ
equations. In order to account for the direct effects of small scales a¼1 b¼1
@wa @wb r @xi @xi 
on the droplets the stochastic model of [18] is adopted in the pres-
All of the terms on the left hand side of Eq. (13) appear in closed
ent work. This is substantially different from that of other studies
form and no modelling is required. In contrast the two terms on
of spray combustion (see for example [7,24]).
the right hand side of Eq. (13) describe the sub-grid transport and
The parameter governing convection, Sh, is rewritten as the sum
micro-mixing. These terms describe processes occurring at the
of the resolved deterministic contribution and a random part
small scales that are not resolved by the LES approach and are
(Sh = Sh(d) + Sh(st)). The deterministic contribution is obtained in a
therefore unknown and need to be modelled. In the present work
conventional manner from the correlations given in [25,26]. The
the dynamic Smagorinsky sub-grid viscosity model is used for
term Sh(st) accounts for the missing sgs turbulence effects, which
transport and the Linear Mean Square estimation closure (LMSE)
should vanish as LES approaches DNS and be consistent with the
[38] is applied for micro-mixing. The model was also proposed
Ranz-Marshall correlation (i.e. capture in a similar fashion unre-
independently in stirred reactor studies where it is known as Inter-
solved effects). Thus the stochastic contribution to Eq. (10) is rep-
action by Exchange with the Mean (IEM) [39]. The LMSE/IEM mix-
resented by:
ing model has been widely used in many sgs-pdf studies.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Including these models, Eq. (13) finally becomes:
u 1=2 t
u
ðstÞ 1=3 tqg ksgs 2r p jdW j 3=4
Sh dt ¼ C V Scg sp ð12Þ e sgs ðwÞ @ qu
@qP e sgs ðwÞ qmðwÞ
~j P _ XNs
@ h i
lg þ  e ðwÞ þ
P qx_ a ðwÞ Pe sgs ðwÞ
@t @xj qðwÞ sgs @w a
2   a¼1 3
where CV is a model constant assigned a value of unity [18]. The X @
N s q m _ ðwÞ  _
mðwÞw
a a
properties of the gaseous phase are evaluated using the one-third þ 4 e sgs ðwÞ5
P
rule [27–29]. a¼1
@wa qðwÞ
It is important to note that ksgs (used in both (Eqs. (7) and (12)) "  #
@ l lsgs @ Pe sgs ðwÞ
is modelled proportional to the sgs viscosity. A dynamic model is ¼ þ
used to determine Cs and a zero value of this implies no sub-grid
@xi r rsgs @xi
fluctuations. As a consequence in areas of the flow that are well re- q X N
@ h i
 e sgs ðwÞ
ðwa  /a ðx; tÞÞ P ð14Þ
solved the sgs turbulence kinetic energy will tend to zero. The ssgs a¼1 @wa
model is thus consistent with the findings of Pozorski and Apte
[30] who performed a systematic study of the direct effect of where rsgs has been assigned a value of 1.0 and x
_ a ðwÞ is the net spe-
sub-grid scale velocity fluctuations on particle motion in forced cies formation rate through chemical reaction. Following, e.g.
isotropic turbulence. It was shown that in poorly resolved regions, [40,41] the sub-grid mixing time scale is assumed given by:
where the sgs kinetic energy was more than 30% of the total the ef- l þ lsgs
fect on droplet motion was more pronounced. However, in well re- s1
sgs ¼ C d ð15Þ
solved areas, where the amount of energy in the sub-grid scales
q D2
was small its effect was not strong. This has the property that ssgs tends to zero as the filter width, D
The spray model, Eqs. (4)–(12) has previously been applied to tends to zero, which correctly drives the pdf towards a d-function
the simulation of dispersion in a water droplet laden mixing layer, as this limit is approached.
[17,31], to evaporating acetone and kerosene sprays, [18,32] and to The high dimensionality of Eq. (14) does not allow a solution by
an axisymmetric combustion chamber, [33,34]. conventional difference schemes because the cost increases expo-
nentially as the number of scalars is increases. The approach com-
monly used is that based on the use of Lagrangian stochastic
3.3. Combustion model
particles where an ensemble of particles is used to represent the
joint pdf [16,38,42]. In the current work an alternative approach
For the low-Mach number gaseous phase, multicomponent
is followed. Eq. (14) is solved using the Eulerian stochastic field
reacting mixture, the scalars of interest are the Ns scalar quantities
method. The P e sgs ðwÞ is represented by an ensemble of N stochastic
which include the N species mass fraction and the mixture specific
fields for each of the Ns scalars, namely nna ðx; tÞ for 1 6 n 6 N,
enthalpy h governed by Eq. (3) (Ns = N + 1). An evolution equation
1 6 a 6 Ns, viz:
for the one-point, one-time pdf for the set of Ns variables that
determine the local thermo-chemical state of a reacting system XN Z Y
Ns
e sgs ðw; x; tÞ ¼ 1
P qGðx  x0 ; DðxÞÞdx0  d wa  nm
can then be derived using established techniques, [16,35,36]. The a ðx; tÞ
N m¼1 X a¼1
main advantage of the pdf approach is that the chemical and phase
exchange source terms of the transport equations for the scalar ð16Þ
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 227

Table 1 ner process, different for each field but independent of the spatial
Operating conditions for the simulated cases. location x. The above equation preserves the boundedness of the
Operating Inlet air pressure Preheated air Burner scalar as the gradient of the field vanishes when the scalars go to
conditions (bar) temperature (K) AFR extrema and therefore the stochastic contribution tends to zero.
A 4 550 20 Each field satisfies the mass conservation and boundedness proper-
E 4 295 – ties of the modelled pdf equation: the species mass fraction will re-
main positive and sum to unity. It is important to stress that the
stochastic fields given by Eq. (17) do not represent any particular
Table 2 realisation of the real field, but rather form an equivalent stochastic
Mass flow boundary conditions for the simulated cases. system (both sets have the same one-point pdf) that is smooth on
the scale of the filter width. All one point moments can be obtained
Operating Burner feed air Film cooling air Fuel M.F.R.
conditions M.F.R. (g/s) M.F.R. (g/s) (g/s) from averaging over the stochastic fields, e.g.:
A 59.5 17.0 2.98 X
N
E 81.3 23.2 – ea ¼ 1
/ nn ð18Þ
N 1 a

3.3.1. Chemical mechanism

A detailed description of kerosene combustion involves a very
large number of chemical species and reaction steps. From a com-
putational standpoint an essential aspect of combustion modelling
using the pdf methodology (and most others approaches) is a
reduction in the number species for which transport equations
have to be solved to manageable proportions. For this reason a glo-
bal mechanism for kerosene combustion has been devised, based
on the reduced mechanism of [47] for the combustion of alkanes.
Kerosene is represented by C12H23 and reaction involves the fol-
lowing four global reaction steps:

C 12 H23 þ 6O2 * 12CO þ 11:5H2 ðiÞ

C 12 H23 þ 12H2 O * 12CO þ 23:5H2 ðiiÞ
H2 þ 12 O2  H2 O ðiiiÞ
CO þ H2 O  CO2 þ H2 ðiv Þ

with rate constants2:

 
30; 000 0:75 pffiffiffi
r_ i ¼ 4:4  1011 exp q nC12 H23 n1:25
O2
RT
 
30; 000
r_ ii ¼ 3:00  108 exp qnC12 H23 nH2O
RT
 
2:50  1016 40; 000 0:75 pffiffiffi 2:25
r_ iii ¼ exp q nH2 nO2 =nH2O
T RT
 
20; 000
r_ iv ¼ 2:75  109 exp qnCO nH2O Þ
Fig. 4. Schematic of the geometrical approach in order to define the injection angles RT
H, H0 . The plot on the top of the figure demonstrates the experimental data of mean
droplet velocity at z = 7 mm The performance of the scheme in laminar premixed flames is
discussed in [48]. In the present context it is to be noted that for
Two formulations of the method can be devised depending on a stoichiometric kerosene-air mixture at a pressure of 4 bar and
whether an Ito or Stratonovich interpretation of the stochastic inte- an initial temperature of 550 K the mechanism predicts a burning
gral is adopted; for a description of these alternatives see [43,44]. In velocity of around 145 cm/s compared with a value of 170 cm/s gi-
the present work the Ito formulation is adopted in which case the ven by the detailed skeletal mechanism [49,50]. Although the value
stochastic fields are continuous and differentiable in space and con- of 145 cm/s is probably a little on the low side the few measure-
tinuous though not differentiable in time. The method has been suc- ments of kerosene-air flame speed available are insufficient to jus-
cessfully applied in various problems in the context of RANS [45,46] tify change.
as well as LES [33,34,40,41]. In the present study the influence of
the sgs fluctuations of the dispersed phase is neglected and the sto-
4. Simulation specifics
chastic fields evolve according to:
sffiffiffiffiffiffiffiffi The in-house block-structured, parallel, boundary conforming
@nna @ n
0 @na 2C0 @nna
n
q dna ¼ q u~i dt þ C dt þ q  dW ni coordinate LES code, BOFFIN-LES, [51] has been used for the calcu-
@xi @xi @xi q @xj lations presented in this paper. LES was performed for two of the
q  n e  
_ na ðnn Þdt þ m e a Þ  mð

e a Þnn dt operating conditions at which DLR-AT carried out measurements
 n  / a dt þ qx _ að/ _ /
2ssgs a a
(see Table 1): an isothermal simulation at condition E, and reacting
ð17Þ simulations at condition A. A summary of the mass flow boundary
conditions used in these CFD simulations is given in Table 2.
where C0 represents the total – molecular plus sub-grid scale – dif-
n kg mol
fusion coefficient and dW i represents increments of a vector Wie- 2
The units of the reaction rates are: kg s
.
228 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

Fig. 5. Radial profiles of axial, radial and tangential mean (left) and rms (right) of velocity at z = 5 mm for isothermal flow field. Squares represent the experiments [10,11] and
solid lines the LES simulations.

Fig. 6. Radial profiles of axial, radial and tangential mean (left) and rms (right) of velocity at z = 10 mm for isothermal flow field. Squares represent the experiments [10,11]
and solid lines the LES simulations.

The LES code is based on a finite-volume approach using an im- considered the turbulence in the immediate vicinity of a solid sur-
plicit low-Mach number formulation with a two-step approximate face and, indeed, the wall shear stress exert a negligible influence
factorisation pressure correction technique used to ensure mass on the overall flow structure. Non-reflecting conditions are applied
conservation. Spatial derivatives are approximated with second- at the outflow boundary. Further details of method can be found in
order central differences and a Crank–Nicolson scheme is used e.g. [48].
for temporal discretisation. The convective part of the Ito equations The Weiner process
pffiffiffiffiffi
ffi of Eq. (17) is represented by time step
of the stochastic field equations is discretised with a Total Varia- increments gi Dt where gi is a [1, 1] dichotomic random vector
tion Diminish (TVD)-scheme, [52], applied in a linearised implicit [19,54]. Random numbers g2i1 for 1 6 i 6 N/2 are selected from
manner. The simulations are performed with a filter width equal a dichotomic distribution and the remaining numbers are deter-
to the cube root of the local grid cell volume and a dynamic version mined from g2i = g2i1 for 1 6 i 6 N/2. Providing an even number,
of the Smagorinsky model is used for the sgs stresses. Wall-func- N of fields is selected the procedure ensures that the mean and
tions, based on the semi-logarithm law of the wall [53] are applied variances of the random vector are zero and unity regardless of
at all solid boundaries. In combustion chamber flows of the type the number of fields. The Ito process is discretised using the
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 229

Fig. 7. Snapshots of the spray injection at three times.

Euler–Maruyama scheme [54], which is a variant of the commonly An initial estimate of injection angle was obtained by geometrical
used Euler scheme. For the particle transport equation similar considerations, see Fig. 4, and a series of computations were per-
schemes are used. The Euler–Maruyama scheme is used for dis- formed for the range of angles 120–170°. A value of 160° was found
cretisation and the
ffi increment of the Weiner process, DW is repre-
pffiffiffiffiffi to result in good agreement with the measured profile at 7 mm
sented by ni Dt where ni is a random vector sampled from a downstream. The droplets are presumed to ‘bounce’ on impact
standard Gaussian distribution, independent for each time step with a solid surface, although very few droplets do so in the pres-
and each coordinate direction. ent study. The time-step used was 0.1 ls and after the start-up a
For the non-reactive case a time-step of 0.2 ls was chosen. The number close to one million droplets were present in the domain
geometry under consideration is 0.38 m long and given an average at each time step. The algorithm for the combined Eulerian–
bulk velocity of 40 m/s one flow-through time corresponds to Lagrangian model is as follows:
0.0095 s. Thus initially 60,000 time-steps, corresponding to around
1.3 flow-through times, were performed in order to flush-out the  The equations for the continuous phase are first solved to deter-
initial conditions and allow the flow field to develop. Statistics mine the gas-phase properties everywhere in the domain.
were then gathered over a further 4.2 flow-through times,  The gas-phase properties are interpolated from the Eulerian
210,000 time-steps. The inlet velocity for the reactive case is higher grid to the particle position using a trilinear interpolation
(around 60 m/s) which results in mean flow through time of scheme.
0.0063 s. For the reactive case 120,000 time steps (0.012 s) for  The particle properties are updated by integrating the stochastic
the corresponding isothermal case without the spray was run ini- differential equations in time.
tially to allow the underlying flow field to develop, followed by an-  Statistical averages are calculated for the dispersed phase using
other 180,000 time steps with injection of spray and reaction an ensemble average from the set of the particles found in the
(0.018 s) to ‘flush out’ the initial conditions. Statistics were then appropriate Eulerian cell.
collected over a further 300,000 time steps (0.03 s).  The source terms of the flow field equations due to the dis-
Reacting simulations were performed at operating condition A persed phase are updated; to be used by the continuous-phase
using the coupled Eulerian (continuous phase) and Lagrangian (li- solver.
quid phase) formulation that was described in the previous sec-  The process is repeated.
tion. The fuel used is a commercial aviation kerosene (Jet-A) and
is represented as C12H23 with properties given by Rachner, [55]. 5. Results and discussion
The combustion process is described by the global 4-step, 7-spe-
cies reaction mechanism [47] described above. The liquid fuel flow 5.1. Isothermal simulations: Case E
is modelled using particles that are injected in a ring of discrete
locations at an axial location just downstream of the fuel injector Attention is first directed towards an evaluation of the ability of
prefilmer lip. A range of particle size groups are injected, using a the LES formulation to reproduce the velocity field characteristics
Rosin–Rammler pdf. The droplet temperature is 295 K and the of the continuous gas phase. Air at temperature of 295 K is injected
injection velocity is 50 m/s, a value estimated from the measure- into the plenum and acquires a swirling motion as it passes
ments available at 7 mm downstream of the injection ring. A trial through the vanes of the double swirler. Figures 5 and 6 show a
and error procedure was used to determine the injection angle. comparison of the measured, [10,11] and simulated time averaged

Fig. 8. (a) Snapshot of temperature from the flamelet simulation after 0.006 s. (b) Snapshot of temperature from the pdf simulation after 0.042 s.
230 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

Fig. 9. (a) Mean axial velocity for the non-reactive case E. (b) Mean axial velocity for reactive case A without ignition. (c) as (b) but with droplet injection. (d) Mean axial
velocity for the reactive case A.

and rms profiles of the axial, radial and azimuthal velocity compo- 5.2. Reactive simulations: Case A
nents at planes normal to the combustor axis at two locations
(5 mm and 10 mm) downstream of the fuel injector. It is evident The reacting flow simulations were started by first simulating a
the simulations of both mean and rms of the three velocity compo- non-reacting flow with liquid kerosene injected from the injection
nents compare well with the measured data. The size of the recir- ring. The evolution of the injection process is illustrated in Fig. 7.
culation zone and the magnitude of all three components is The fuel mass flow is set to 2.98 g/s corresponding to an AFR of
accurately reproduced. It is observed that the mean velocity down- 20. Once spray particles are injected into the chamber and, as the
stream of the injector is closely axisymmetric whilst the rms pro- air injected in the inlet is already preheated (550 K), the droplets
files indicate a significant amount of unsteadiness. The start evaporating creating a combustible mixture. The air–fuel mix-
simulations show some asymmetry for z = 10 mm which however ture is ignited using an essentially unstrained ‘mixed is burnt’ flam-
can be explained form the rather short time of the simulation due elet model in place of the sgs-pdf model. A snapshot of temperature
to computational restrictions. arising from the flamelet simulations can be seen in Fig. 8a.

Fig. 10. Comparison of OH-PLIF measured mean temperature field (left) with LES with the pdf combustion model with one (middle) and eight (right) fields. The white line
indicates the swirler exit plane.
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 231

Fig. 11. Slices at different axial locations of instantaneous contours for LES calculations with one field. Figures at the top show instantaneous temperature fields. Black lines
demonstrate the stoichiometric mixture fraction. Figures in the middle show instantaneous velocity field and figures at the bottom the droplet distribution coloured
according to their size. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 12. Slices at different axial locations of instantaneous contours for LES calculations with eight fields. Figures at the top show instantaneous temperature fields. Black lines
demonstrate the stoichiometric mixture fraction. Figures in the middle show instantaneous velocity field and figures at the bottom the droplet distribution coloured
according to their size. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

The burning flamelet solution was then used to initiate the sgs- of the highest temperatures (>2100 K) are more localised with the
pdf equation/stochastic fields method. Figure 8b shows a snapshot flamelet whereas the temperature is more uniformly distributed
of the temperature field arising with this method. The solutions with the sgs-pdf method. This provides an indication of the impor-
with the two models are quite different. It is evident that the areas tance of the combustion model in the simulation of the instanta-
232 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

Fig. 13. Radial profiles of mean axial and radial velocity (left) and temperature (right) at different axial locations for the reactive case A with LES. Dashed lines represent 1
field predictions and solid lines 8 fields.

Fig. 14. Radial profiles of mean axial and radial velocity (left) and temperature (right) at different axial locations for the reactive case A with LES. Dashed lines represent 1
field predictions and solid lines 8 fields.

neous (and consequently the mean) flame structure. In the follow- a more detailed analysis two extra computations, not correspond-
ing sections the influence of the sgs-pdf model on the simulations ing to any measured case, have been included. Figure 9b shows re-
of the flow field and how this affects the spray statistics is investi- sults for the same input velocity and temperature as the reactive
gated. Following [40] eight stochastic fields have been used to case A without the liquid fuel spray and thus no combustion oc-
characterise the influence of the sub-grid fluctuations; in [40] eight curs. Figure 9c corresponds to the same conditions but with the
and sixteen fields were found to give very similar results. The influ- spray included; no external ignition process is applied and thus
ence of the number of stochastic fields has also been investigated no combustion occurs. It is evident in all four cases the mean axial
in [56] where similar conclusions are drawn. Results were also ob- velocity field displays a symmetrical cone-angle as it exits the fuel
tained using a single field; a special case where the effects of sub- injector together with a strong recirculation zone (area coloured in
grid fluctuations on combustion are neglected and the sgs-pdf col- blue) in the centre of the cone. For the reactive test case A the recir-
lapses to delta functions at the filtered values of the scalars. culation zone is stronger and extends further downstream com-
Figure 9 shows contour plots of the mean axial velocity of the pared to the other three cases which is consistent with the lower
non-reactive case E (Fig. 9a) and the reactive case A (Fig. 9d). For densities arising due to combustion. In addition, it can be seen that
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 233

exerted on the continuous phase by the droplets, term f i of Eq. (2),

are small.
Figure 10 compares a picture of the DLR-AT measured temper-
ature just downstream of the fuel injector [10,11] with LES simula-
tions using one and eight stochastic fields. The experimental
temperatures were calculated from absolute OH densities which
were in turn measured using simultaneous PLIF and absorption.
In lean flames and under the assumption of OH being in chemical
equilibrium, temperatures can be inferred from OH concentrations.
There is some uncertainty at the determination of local OH concen-
tration (less than 30%) mainly because of the uncertainty of the
fluorescence quantum yield for the unknown local gas composi-
tion. The OH super-equilibrium concentration within the flame
front leads to an overestimation of the temperature of about
100 K but the relaxation has been reported to take place within less
than 50 ls at 10 bar [10,11]. As can be seen the V-shape of the
flame is captured reasonably well by the simulations. The temper-
ature field for both simulations seems to be somewhat high in the
Fig. 15. Radial profiles axial velocity rms at different axial locations for the reactive vicinity of the centreline. The distinct ‘lobes’ of high temperature in
case A with LES. Dashed lines represent 1 field predictions and solid lines 8 fields. the experiment are also reproduced although the lift off height is
not accurately reproduced as the lobes from the simulations ap-
pear to be located further upstream than is measured. A stable
pocket of hot gases in the recirculation zone located around the
centreline close to the injection point is also reproduced similar
to the experiments. It appears that qualitatively the inclusion of
the sgs combustion model (by increasing the number of fields from
one to eight) does not influence the structure of the flame appre-
ciably. However, a more detailed analysis, to follow, involving a
comparison of the radial profiles reveals otherwise.
Figures 11 and 12 shows ‘slices’ of the instantaneous tempera-
ture, the instantaneous velocities and droplet diameter at down-
stream locations that correspond to three positions in the area
indicated in Fig. 10. Figure 11 corresponds to the results with eight
fields whilst Fig. 12 are those for one field. It can be seen that the
higher areas of temperature are located in the outer part of the
recirculation bubble. Inside the recirculation bubble the tempera-
tures are high especially at the most downstream location. The
instantaneous structure of the flame does not seem to change sig-
nificantly when the number of fields is increased from one to eight
except possibly at z = 0.045 where the recirculation zone appears
less strong when eight fields are used. Also the temperature distri-
Fig. 16. Radial profiles temperature rms at different axial locations for the reactive bution is wider at this location.
case A with LES. Dashed lines represent 1 field predictions and solid lines 8 fields.
Figures 13–16 show the simulated radial profiles of mean and
rms of temperature and axial and radial velocity at different axial
the presence of droplets does not change the flow field signifi- locations arising from one field and eight stochastic fields. The first
cantly, (Fig. 9b and Fig. 9c), the spray is dilute and thus the forces three axial locations correspond to the area close to the injector

Fig. 17. 3-D plot of the instantaneous velocity (arrows), droplet distribution (black dots) and temperature iso-contour for T = 2100 K.
234 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

Fig. 18. Slice of the velocity field (arrows) and droplet diameter contour. The yellow line indicates the iso-contour of T = 1500 K and the blue line the iso-contour of zero
velocity in order to locate the recirculation zone. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 19. Mean axial velocity field and spray diameter, temperature and axial velocity contours. The red line limits the area with flow temperature higher than 1500 K. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 20. Spray pdf at different axial locations: (a) reactive case and (b) non-reactive case.
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 235

Fig. 21. Spray pdf at different axial and radial locations. Squares represent the experiments [10,11], and lines the LES simulations with 8 fields.

Fig. 22. Radial profiles of particle velocities at different axial locations. Squares represent the experiments [10,11], the dashed lines the LES simulations with 1 field and the
solid lines the LES simulations with 8 fields.

just before the flow enters the combustion zone (in Fig. 10 the area a radial position of roughly 20 mm and in a smaller penetration
below the ‘lobes’) and the other three correspond to axial locations of the flow into the combustion region (in the black circle of
inside the reaction zone, as is indicated by the magnitude and Fig. 10).
shape of mean temperature profiles. Attention is now turned to an investigation of the spray statis-
The profiles of the mean and rms velocities do not change as the tics and how these are affected by the combusting flow field. Fig-
number of fields is increased from one to eight, however some ures 17–19 provide a general overview of the gaseous and liquid
small differences (that, as will be demonstrated below, are impor- phases. More specifically, Fig. 17 shows a 3-D snapshot of the flow
tant when the spray statistics are considered) are evident in the field and droplets present (black dots) with the red line represent-
mean and rms temperatures. The increase in the number of fields ing gas temperatures greater than 1500 K while Fig. 18 is a slice
from one to eight implies that instead of the delta function corre- through the computational domain with the two coloured V-
sponding to the single field a ‘broader’ shaped pdf is expected to shaped branches indicating the position of the droplets A vector
arise with eight fields. With eight stochastic fields the number of representation has been used in Fig. 17 to represent the continuous
samples is equal to the number of fields multiplied by the number phase flow field.
of time steps thus allowing the pdf to be computed much more In Figure 18 where a slice of the combustor and droplet diame-
accurately. This broadening of the pdf is manifested in the radial ter contours are shown, along with the stagnation contour for the
profiles of temperature as smaller gradients in the region around velocity, two recirculation zones are evident. The first recirculation
236 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

flow balancing their tangential velocity. Also the droplet tempera-

tures are higher at the location where larger droplets are present
(at the centre of the branches). These droplets also appear to have
higher velocities. It is interesting to note that droplets are still
present even at gas temperatures higher than 1500 K (the area is
indicated by the red line in the figures) which implies a complex
flame structure with the coexistence of premixed combustion
(air with vapour created from the vaporisation process at
z < 15 mm) that has already occurred and non-premixed combus-
tion from the liquid droplets that are still present at this location.
Looking again at Figs. 11 and 12a diffusion flame is expected to
be located close to the areas with the stoichiometric mixture frac-
tion. However there are pockets of high temperature inside the
zone of the stoichiometric mixture fraction that probably corre-
spond to premixed flames. These findings are consistent with a re-
cent work by Luo et al. [57]. The droplets that enter the high
temperature zone are the bigger droplets that take longer to evap-
Fig. 23. Radial profiles of SMD at different axial locations. Squares represent the orate fully.
experiments [10,11], the dashed lines the LES simulations with 1 field and the solid Figure 20 shows the evolution of the droplet size pdf at different
lines the LES simulations with 8 fields. axial locations for both the reactive and the equivalent non-reac-
tive cases in order to better understand how combustion alters
the rate of evaporation and thus the distribution of the droplets.
zone arises from the presence of the double swirler and is of par-
In both cases the initial pdf (z = 0) corresponds to a Rosin–Rammler
ticular interest because it is located in between the ‘V-shaped’
pdf with a mean droplet size of 6lm. For the combusting case,
spray stream and is expected to affect the mixing of the fuel vapour
Fig. 20(a) it can be seen that initially (for z < 20 mm) the proportion
with the preheated air. The second recirculation zone is located
of large droplets is maintained close to the initial distribution con-
close to the combustor wall on the sides of the lobes and is ex-
sistent with the smallest droplets disappearing almost at the same
pected to stabilise the opening angle of the spray. Unfortunately,
rate as the large ones reduce in size. Further downstream, as the
experimental measurements of the mean and rms of the gas phase
droplets enter the reaction zone and their temperatures increases
velocity components are not available in order to assess the accu-
rapidly, the large droplets continue heating up but the temperature
racy of the LES simulations. However, given that the grid is fine en-
of the smaller droplets reaches boiling point at an increased rate.
ough to reproduce accurately the non-reactive test case E (see
The flattering of the pdf indicates that the large droplets reduce
Figs. 5 and 6) and that the flame structure is well reproduced at
in size slower than the rate of disappearance of the small ones.
a qualitative level (see Fig. 10) it is to be expected that the flow
In other words the number of small droplets, initially larger after
field statistics are well reproduced.
entering the combustion zone reduces rapidly and at some point
Figure 19a–c show a zoom-in of the region close to the swirler
balances out the number of the large droplets that evaporate
exit and display contours of the spray diameter, droplet tempera-
slower.
ture and mean axial velocity respectively thus providing a qualita-
Figure 20(b) shows the evolution of the droplet size pdf for a
tive representation of the spray dispersion. Regarding droplet
case equivalent to reactive case A in which no ignition process is
dispersion it can be seen in Fig. 19 that the small droplets disperse
applied and thus the underlying flow field temperature remains
more rapidly and are thus mostly present at the outer part of the
relatively constant. The behaviour of the droplet-size pdf is signif-
spray cone whilst the large droplets remain in the centre. This is
icantly different. Although there is no combustion, the initial tem-
expected from a physical point of view, the higher inertia of the
perature of the droplets is 295 K, significantly lower from the
larger droplets leads to higher radial velocities through the forward
temperature of the underlying flow field (550 K) causing the drop-
lets to heat up and evaporate, at a rate different to that of the pre-
vious case. The pdf starts flattening out almost immediately
downstream, indicating that the small droplets disappear quicker
than the rate that the large ones reduce in size since the latter
do not immediately reach the boiling point. It should be noted here
that the comparison of the two cases is only qualitative as the flow
fields are different and this influences the dispersion of the drop-
lets (see Fig. 9). This is likely to affect to a certain extent the move-
ment of the droplets closer to the swirler exit where the
recirculation is created and thus the behaviour of the droplet-size
pdf.
Figure 21 shows a comparison of measured and simulated
spray pdfs at different axial and radial locations. Squares repre-
sent the experiments [10,11], and lines the LES simulations with
8 stochastic fields. The computed pdfs were reconstructed from
samples collected from the simulations over one flow through
time and their lack of smoothness arises because of small finite
volumes of the flow solver. Nevertheless it can be seen that the
pdf is reproduced to a reasonable accuracy by the simulations.
Fig. 24. Radial profiles of the droplets’ temperature at different axial locations.
The measured and simulated radial profiles of the three compo-
Dashed lines represent the LES simulations with 1 field and solid lines the LES nents of droplet velocity are compared in Fig. 22 at three axial loca-
simulations with 8 fields. tions close to the injector. Squares represent the experiments
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 237

Fig. 25. Scatter plots of mean droplet velocity against diameter at (a) z = 7 mm and at (b) z = 30 mm for one and eight fields. Pseudo-colour represents mean droplet
temperature. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

[10,11], black lines the simulations with one field and the red lines diameters (SMD) at six axial locations are presented. The experi-
simulations with eight stochastic fields. To aid the comparison the mental data shows that only a very small decrease of SMD is ob-
simulated for gas phase velocities (dashed lines) are also included. served with increasing downstream distance. Also relatively
The width of the simulated recirculation zone of the gas phase and small radial variations are present, except in the region of the recir-
that of the dispersed phase are similar and both simulations (with culation zone where the size of the droplets reduces considerably.
one and eight fields) reproduce quite accurately the experimental This size distribution is typical of a hollow cone atomiser where
data. However at the downstream locations the agreement with the smaller size droplets are entrained in the core and the larger
the measured profiles is improved when eight stochastic fields droplets travel to the edge of the spray. As the droplets travel
are used, i.e. the influence of sgs fluctuations are included. An inter- downstream they evaporate and the radial distribution tends to
pretation of these results is aided when they are viewed together become more uniform. This behaviour is captured reasonably accu-
with Fig. 23 in which the radial profiles of the droplets Sauter mean rately both quantitatively and qualitatively by the simulations. The
238 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239

inclusion of the sgs fluctuation model, by increasing the number of fully coupled with a Lagrangian description of the dispersed phase.
fields from one to eight stochastic fields improves the results in the Comparisons of the LES results with measurements for the non-
areas around the edges of the recirculation zone (r = 10 mm). This reacting case demonstrate that the essential features of the flow
can be explained by the temperature profiles presented in Fig. 13. are reproduced to a good level of accuracy. In the combusting case
The simulated gas phase temperatures at these locations are the results indicate that the simulated flame structure is similar
slightly higher if eight fields are included (see Figs. 13,14). In terms with and without the inclusion of the sgs combustion model. How-
of mean flow field statistics a difference of an order of 20 K may not ever the simulated spray statistics differ considerably with the two
be considered significant but an equivalent increase in the droplet cases and the results serve to demonstrate the importance of the
temperature (see Fig. 24) will considerably change the droplet size sgs combustion model. The model for the dispersed phase - the fuel
because of its effect on the droplet evaporation rate. This is most spray - used in this study reproduces the SMD and the velocity of
pronounced at locations some way downstream of the injector the droplets very accurately. Some discrepancies between the
where the droplets have already started to heat up due to the pre- experimental data and simulations are evident in the recirculation
heated flow field. The higher gas field temperature results also af- zone and these are mostly attributable to the uncertainty in the in-
fect the droplet velocity. Since the droplet size is captured more let spray boundary conditions and thus further experimental
accurately the dispersion, dependent on droplet diameter, is also investigation is desirable. A second issue that needs to be investi-
more accurately reproduced. As expected the smaller droplets fol- gated in future work is the suitability of a Lagrangian formulation
low closer the gas phase flow field streamlines, as can be seen in for the dispersed phase close to the injector. The method used is
Figs. 22 and 23. In the recirculation zone there is disagreement be- mostly valid for a dilute spray with droplet sizes sufficiently small
cause according to the simulations no droplets are present in this so that they can be considered as point sources of mass, momen-
zone, whereas the measurements indicate the presence of droplets. tum, species and energy. However, in areas close to the injection
An explanation for this is difficult to provide. The smaller droplets point the liquid phase is not always present in the form of droplets
are expected to follow the gas flow and thus enter the recirculation but rather as liquid blobs or ligaments of size comparable to the
zone as can be seen from the measured SMDs and corresponding scales of the motion of the continuous phase.
velocities at 7 mm and 10 mm. However, if these small droplets ac-
quire a negative velocity and enter the recirculation zone where Acknowledgments
the temperatures are higher than 1500 K then they are likely to
evaporate rapidly. It may be, however, that these discrepancies This work received funding from the European Community
are related to the high temperatures in the centreline observed through the Project TIMECOP-AE (Project No. AST5-CT-2006-
in Fig. 13. 030828). It reflects only the authors’ views and the Community
In order to investigate in more detail the relationship between is not liable for any use that may be made of the information con-
the three variables, temperature, velocity and diameter that char- tained therein. The authors are also grateful to Dr. C. Goddard for
acterise droplet dispersion and evaporation the results at two providing the grid used in the present calculations and Dr. S. Nav-
locations are considered. The first is close to the swirler exit arro-Martinez for valuable discussions.
(z = 7 mm) and the second is in the fully burning region at
z = 30 mm. At the first location close to the injector where the References
recirculation zone is developed the sub-grid contribution is
important. The second position is more indicative of the perfor- [1] K. Mahesh, G. Constantinescu, S. Apte, G. Iaccarino, F. Ham, P. Moin, Large-eddy
mance of the sgs combustion model. As the radial profiles of simulation of reacting turbulent flows in complex geometries, J. Appl. Mech. 73
(3) (2006) 374–381.
the mean and rms of velocities and temperatures do not show [2] G. Boudier, L.Y.M. Gicquel, T.J. Poinsot, D. Bissières, C. Bérat, Comparison of LES,
any significant differences a more detailed examination of droplet RANS and experiments in an aeronautical gas turbine combustion chamber,
statistics is provided. Proc. Combust. Inst. 31(2) (2007) 3075–3082.
[3] G. Boudier, L.Y.M. Gicquel, T.J. Poinsot, Effects of mesh resolution on large eddy
Figures 25(a) and 25(b) show scatter plots of mean droplet simulation of reacting flows in complex geometry combustors, Combust.
velocity against droplet diameter at a downstream location of Flame 155 (1–2) (2008) 196–214.
z = 7 mm and z = 30 mm for one and eight stochastic fields, i.e. [4] M. Boileau, S. Pascaud, E. Riber, B. Cuenot, L. Gicquel, T. Poinsot, M. Cazalens,
Investigation of two-fluid methods for large eddy simulation of spray
with and without the sgs combustion model; pseudo-colour rep-
combustion in gas turbines, Flow Turbul. Combust. 80 (2008) 291–321.
resents mean droplet temperature. The two results show very [5] Ying-wen Yan, Jian-xing Zhao, Jing-zhou Zhang, Yong Liu, Large-eddy
little difference at the first location, as expected because there simulation of two-phase spray combustion for gas turbine combustors, Appl.
is no combustion in this region. However there appears to be Therm. Eng. 28 (11–12) (2008) 1365–1374.
[6] T. Lederlin, H. Pitsch, Large-Eddy Simulation of an Evaporating and Reacting
clear correlation between diameter and velocity suggesting that Spray, Tech. Rep. Center for Turbulence Research, Annual Research Briefs,
the random term of Eq. (7) does not play a significant role, 2008.
especially for the smaller diameters, in this region. In the case [7] S.V. Apte, K. Mahesh, P. Moin, Large-eddy simulation of evaporating spray in a
coaxial combustor, Proc. Combust. Inst. 32 II (2009) 2247–2256.
of the second location the differences between the one and eight [8] M. Sanjose, J. Senoner, F. Jaegle, B. Cuenot, S. Moreau, T. Poinsot, Fuel injection
field solutions are larger with the droplet diameters and veloci- model for Euler–Euler and Euler–Lagrange Large–Eddy simulations of an
ties being essentially uncorrelated with the sgs combustion evaporating spray inside an aeronautical combustor, Int. J. Multiph. Flow 37
(2011) 514–529.
model. An interesting branching phenomenon is also evident at [9] G. Hannebique, P. Sierra, E. Riber, B. Cuenot, Large eddy simulation of reactive
the top left figure which implies that droplets with the same two-phase flow in an aeronautical multipoint burner Flow, Turbul. Combust.
diameter have different axial velocities depending their 90 (2013) 449–469.
[10] U. Meier, J. Heinze, S. Freitag, C. Hassa, Spray and flame structure of a generic
temperature. injector at aeroengine conditions, in: Proceedings of ASME Turbo Expo 2011:
Power for Land, Sea and Air: GT2011, vol. GT2011-45282, Vancouver, Canada,
2011.
[11] S. Freitag, U. Meier, J. Heinze, T. Behrendt, C. Hassa, Measurement of initial
6. Conclusions
conditions of a kerosene spray from a generic aeroengine injector at elevated
pressure, in: Proc. 23rd Annual Conference on Liquid Atomization and Spray
In the present work LES is applied to the simulation of the two- Systems, (ILASS-Europe 2010), Brno, Czech Republic. ISBN-978-80-7399-997-1
phase flow in the GENRIG combustor, which is an aeronautical- (137).
[12] C. Goddard, S. Sadig, S. Stow, M. Zedda, Toward Innovative Methods for
type swirl stabilised spray burner, fuelled by Jet-A liquid kerosene. Combustion Prediction in Aero-Engines, Tech. rep., EU Project TIMECOP-AE
An Eulerian description of the continuous phase is adopted and is D4.1.8.
 W.P. Jones et al. / Combustion and Flame 161 (2014) 222–239 239

[13] S. Elghobashi, On predicting particle-laden turbulent flows, Appl. Sci. Res. 52 [37] F. Gao, E.E. O’Brien, A large eddy simulation scheme for turbulent reacting
(4) (1994) 309–329. flows, Phys. Fluids A 5 (1993) 1282–1284.
[14] J. Smagorinsky, General circulation experiments with the primitive equations, [38] C. Dopazo, E.E. O’Brien, Functional formulation of nonisothermal turbulent
J. Mon. Weather Rev. 91 (1963) 99–164. reactive flows, Phys. Fluids 17 (1968).
[15] U. Piomelli, J. Liu, Large-eddy simulation of rotating channel flows using a [39] J. Villermaux, L. Falk, A generalized mixing model for initial contacting of
localized dynamic model, Phys. Fluids 7 (4) (1995) 839–848. reactive fluids, Chem. Eng. Sci. 49 (24B) (1994) 5127–5140.
[16] S.B. Pope, PDF methods for turbulent reactive flows, Prog. Energy Combust. Sci. [40] W.P. Jones, S. Navarro-Martinez, Large eddy simulation of autoignition with a
11 (2) (1985) 119–192. subgrid probability density function method, Combust. Flame 150 (3) (2007)
[17] M. Bini, W.P. Jones, Large-eddy simulation of particle-laden turbulent flows, J. 170–187.
Fluid Mech. 614 (2008) 207–252. [41] W.P. Jones, V.N. Prasad, Large eddy simulation of the Sandia flame series (d–f)
[18] M. Bini, W.P. Jones, Large eddy simulation of an evaporating acetone spray, Int. using the Eulerian stochastic field method, Combust. Flame 157 (9) (2010)
J. Heat Fluid Flow 30 (3) (2009) 471–480. 1621–1636.
[19] C. Gardiner, Handbook of Stochastic Methods, Springer-Verlag, New York, [42] C. Dopazo, Probability density function approach for a turbulent axisymmetric
1984. heated jet: Centerline evolution, Phys. Fluids 18 (1975) 397–404.
[20] M. Bini, W.P. Jones, Particle acceleration in turbulent flows: a class of nonlinear [43] L. Valiño, A field monte carlo formulation for calculating the probability
stochastic models for intermittency, Phys. Fluids 19 (2007) 035104. density function of a single scalar in a turbulent flow, Flow Turbul. Combust.
[21] M. Yuen, L.W. Chen, On drag of evaporating droplets, Combust. Sci. Technol. 14 60 (1998) 157–172.
(1976) 147–154. [44] V. Sabel’nikov, O. Soulard, Rapidly decorrelating velocity-field model as a tool
[22] W. Sirignano, Fuel droplet vaporization and spray combustion theory, Prog. for solving one-point Fokker–Planck equations for probability density
Energy Combust. Sci 9 (4) (1983) 291–322. functions of turbulent reactive scalars, Phys. Rev. E (Stat. Nonlinear Soft
[23] G.M. Faeth, Evaporation and combustion of sprays., Prog. Energy Combust. Sci. Matt. Phys.) 72 (1) (2005). 16301-1.
9 (1–2) (1983) 1–76. [45] A. Garmory, R.E. Britter, E. Mastorakos, Simulation of the evolution of aircraft
[24] V. Sankaran, S. Menon, Les of spray combustion in swirling flows, J. Turbul. 3 exhaust plumes including detailed chemistry and segregation, J. Geophys. Res.
(2002) N11. 113(8) (2008) D08303–D08303.
[25] W. Ranz, W. Marshall, Evaporation from drops: Part I, Chem. Eng. Prog. 48 (3) [46] A. Garmory, E. Mastorakos, Aerosol nucleation and growth in a turbulent jet
(1952) 141–146. using the stochastic fields method, Chem. Eng. Sci. 63 (16) (2008) 4078–4089.
[26] W. Ranz, W. Marshall, Evaporation from drops: Part II, Chem. Eng. Prog. 48 (3) [47] W.P. Jones, R.P. Lindstedt, Global reaction schemes for hydrocarbon
(1952) 173–180. combustion, Combust. Flame 73 (3) (1988) 233–249.
[27] E.M. Sparrow, J.L. Gregg, The variable fluid property problem in free [48] W.P. Jones, A. Tyliszczak, Large eddy simulation of spark ignition in a gas
convection, Trans. Am. Soc. Mech. Eng. 80 (1958) 879–886. turbine combustor, Flow Turbul. Combust. 85 (3–4) (2010) 711–734.
[28] G.L. Hubbard, V.F. Denny, A. Mills, Droplet evaporation: effects of transients [49] J. Luche, Elaboration of Reduced Kinetic Models of Combustion: Application to
and variable properties, Int. J. Heat Mass Transfer 18 (1975) 1003–1008. a Kerosene Mechanism, Ph.D. thesis, LCSR Orléans, 2003.
[29] G.M. FAETH, R.S. LAZAR, Fuel droplet burning rates in a combustion gas [50] B. Franzelli, E. Riber, M. Sanjose, T. Poinsot, A two-step chemical scheme for
environment, AIAA J. 9 (11) (1971) 2165–2167. kerosene-air premixed flames, Combust. Flame 157 (2010) 1364–1373.
[30] J. Pozorski, S.V. Apte, Filtered particle tracking in isotropic turbulence and [51] W.P. Jones, F. di Mare, A.J. Marquis, LES-BOFFIN: Users Guide, Technical
stochastic modeling of subgrid-scale dispersion, Int. J. Multiphase Flow 35 (2) Memorandum, Imperial College, London, 2002.
(2009) 118–128. [52] B.J. Van Leer, Towards the ultimate conservative difference scheme: II
[31] W.P. Jones, S. Lyra, A.J. Marquis, Large eddy simulation of a droplet laden Monotonicity and conservation combined in a second order scheme,
turbulent mixing layer, Int. J. Heat Fluid Flow 31 (2010) 93–100. Comput. Phys. 14 (1974) 361–370.
[32] W.P. Jones, S. Lyra, A.J. Marquis, Large eddy simulation of evaporating kerosene [53] U. Piomelli, E. Balaras, Wall layer models for large eddy simulation, Annu. Rev.
and acetone sprays, Int. J. Heat Mass Trans. 53 (2010) 2491–2505. Fluid Mech. 34 (2002) 349–374.
[33] W.P. Jones, S. Lyra, S. Navarro-Martinez, Large eddy simulation of a swirl [54] P.E. Kloeden, E. Platen, Numerical Solution of Stochastic Differential Equations,
stabilized spray flame, Proc. Combust. Inst. 33 (2011) 2153–2160. Springer-Verlag, New York, 1992.
[34] W.P. Jones, S. Lyra, S. Navarro-Martinez, Large eddy simulation of swirling [55] M. Rachner, Die stoffeigenschaften von kerosin jet a-1, Tech. rep., Institut fur
kerosene spray flames, Combust. Flame 159 (2011) 1539–1561. Antriebstechnik, DLR, Koln, 1998.
[35] E.E. O’Brien, The probability density function (pdf) approach to reacting [56] I.A. Dodoulas, S. Navarro-Martinez, Large eddy simulation of premixed
turbulent flows, in: P. Libby, F. Williams (Eds.), Turbulent Reacting Flows, turbulent flames using the probability density function approach, Flow
vol.44, Springer-Verlag, 1980, pp. 185–218. Appeared under ‘Topics in Applied Turbul. Combust. 90 (2013) 645–678, http://dx.doi.org/10.1007/s10494-013-
Physics’. 9446-z.
[36] W. Kollmann, Pdf transport equations for chemically reacting flows, in: R. [57] K. Luo, H. Pitsch, M. Pai, O. Desjardins, Direct numerical simulations and
Borghi, S. Murthy (Eds.), Turbulent Reactive Flows, Springer-Verlag, 1989, pp. analysis of three-dimensional n-heptane spray flames in a model swirl
715–730. combustor, Proc. Combust. Inst. 33(2) (2011) 2143–2152.