Beruflich Dokumente
Kultur Dokumente
t
+
u
i
x
i
= 0 (2.1)
momentum
u
j
t
+
u
j
u
i
x
i
=
ij
x
i
p
x
j
+g
j
for j = 1, 2, 3 (2.2)
species
Y
t
+
Y
u
i
x
i
=
J
i
x
i
+S
for = 1, 2, . . . , N
s
(2.3)
8
absolute enthalpy
h
t
+
hu
i
x
i
=
J
h
i
x
i
+
Dp
Dt
+
ij
u
j
x
i
+
Q
rad
. (2.4)
In these governing equations, a Roman index denotes a component of a three-dimensional
vector(e.g., i = 1, 2, 3), a Greek index denotes a chemical species (e.g., = 1, 2, . . . , N
s
),
and the usual summation convention applies over repeated Roman indices within a term.
Here u denotes velocity,
Y denotes the mass fractions of the N
s
chemical species, and h
the absolute enthalpy. Mixture mass density is , pressure is p, body force (per unit mass)
is g, and ,
J
,
J
h
, denote, respectively, the viscous stress tensor and the molecular uxes
of species and enthalpy. The chemical source term, S
, for species , S
= W
.
The volume rate of heating due to radiation is
Q
rad
.
Compared to the nonreacting case, this set of conservation equations includes
species transport equations with chemical source terms to address chemical transforma-
tions, and it employs absolute enthalpy instead of internal energy as the energy variable.
Chemical heat release produces high temperature in the ow eld and variable density
eects. The high temperature results in thermal radiation playing a dominant role in
heat transfer. These are a few complexities arising from chemical reactions that com-
pound the already dicult solutions of the Navier-Stokes equations. Detailed discussions
on these issues can be found in references [3, 9, 17].
All of these chemistry-related computations, including radiation properties for a
participating medium, are based on the information provided by a chemical reaction
mechanism of interest, which works with a thermodynamic database and a transport
property database for the chemical species involved. These databases usually organize
the thermodynamic and transport data in terms of polynomials as functions of temper-
ature: for example, CHEMKIN database [18], JANAF Tables [19], and NASA database
[20]. A reaction mechanism is a collection of elementary (or global) reactions necessary to
describe the chemical transformation and reaction rate coecients for each reaction; usu-
ally the latter are given in Arrhenius form. The rate coecients are derived mainly from
9
experimental measurements in shock tubes and ow reactors, from quantum-mechanical
calculations based on dierent theoretical approximations, or from rough estimates based
on simple collision theories. Dedicated software such as the CHEMKIN package [21] has
been developed to facilitate chemistry-related computations. Given a reaction mecha-
nism, a thermodynamic database and a transport database as inputs, the CHEMKIN
package can return information on chemical production rates, thermodynamic properties,
and molecular transport coecients.
The number of species and the number of elementary reactions in a mechanism
determine the level of complexity in chemistry calculations. In reality, a chemical mech-
anism for the transformation of hydrocarbon fuels can include as many as thousands
of elementary reactions involving hundreds of intermediate species. The combustion of
a simple fuel like methane in air requires 325 reactions and 53 species (GRI Mech 3.0
[22]) for a satisfactory chemical description. In the case of auto-ignition of a Diesel fuel
with the typical fuel cetane, several thousand elementary reactions are required to de-
scribe the overall process [3]. In practice, the number of species tracked in combustion
simulations impacts the computer memory usage and CPU time. To make the chem-
istry calculations tractable in three-dimensional time-dependent CFD simulations even
with the largest available supercomputers, only a moderate level of chemistry (tens of
species and hundreds of elementary reactions) generally can be allowed. Broadly used in
current combustion simulations are mechanisms reduced from their detailed versions, or
even simpler one-step global mechanisms that have only three generic species: reactant,
oxidizer, and product.
However, a detailed reaction mechanism is a prerequisite for a realistic and accu-
rate prediction of chemically reacting ows. Predictions that require detailed chemistry
include ame propagation speed, ame location and size, ame extinction and ignition
phenomena, and pollutant emissions. Many important issues that arise in practical com-
bustion applications and environmental concerns can be addressed numerically only with
detailed chemistry. The simulation of engine combustors requires detailed mechanisms
to address kinetically controlled phenomena [23] such as low-temperature auto-ignition,
10
and to address time-dependent phenomena such as combustion instability [24] and vari-
ability [25, 26]. The use of alternative fuels and fuel additives require detailed mecha-
nisms to address fuel composition issues [27]. The application of turbulent combustion
models in the chemical industry requires detailed mechanisms to study operating char-
acteristics of chemical reacting systems [3, 28]. Stringent pollutant emission regulations
require detailed mechanisms to predict trace pollutant species such as NO
x
, unburned
hydrocarbons, and soot [29]. With advances in computer technology and numerical al-
gorithms, detailed chemistry calculations will eventually become an intrinsic part of any
comprehensive, reliable, and credible combustion simulation tool.
2.2 Expenses in Chemistry Calculations
It is computationally demanding to solve the complete conservation equations
including detailed chemistry even for simple two-dimensional laminar ames [30]. For
more complex ows, especially turbulent ames, the inclusion of a detailed chemistry
calculation is even more challenging. For the foreseeable future, CPU time and com-
puter memory limitations will prohibit implementations of fully detailed descriptions of
chemistry into three-dimensional time-dependent CFD simulations of combustion appli-
cations [31]. This statement is based on three factors that make chemistry calculations
expensive in CPU time and demanding in memory requirement.
The rst factor is that the chemical system is usually described in terms of the
mass fractions of each chemical species as functions of space and time. Because of
the large dierences in species diusivities due to large dierences in their molecular
weight, and because of the large dierences in species chemical source terms due to
large dierences in their chemical bonds, each species requires one pde (equation 2.3) for
its conservation. Therefore, if there are N
s
species in the system under consideration,
usually N
s
1 species conservation equations (the sum of species mass fraction is unity)
are needed to determine the chemical state of the reacting system. For even a moderately
sized chemical mechanism, this number far exceeds the number of pdes required for mass,
momentum and energy conservation.
11
The dierences in species molecular diusivities mentioned above often can be
neglected in turbulent ows because of the relatively large eect of turbulent mixing.
However, the dierences in chemical source terms result in great complexity and intro-
duce fundamental diculties in the simulation of turbulent reacting ows.
The second factor that makes chemistry calculations expensive is the nonlinearity
and complexity of the chemical source terms. A chemical reaction mechanism involving
N
s
species and L elementary reactions can be written as [9]
N
s
=1
l
X
N
s
=1
l
X
for l = 1, 2, . . . , L (2.5)
where
l
and
l
are the stoichiometric coecients for the
th
species X
and l
th
reaction. The net molar production rate, i.e., the chemical source term , for each
species can be written as
=
L
l=1
l
q
l
for = 1, 2, . . . , N
s
, (2.6)
where,
l
= (
l
), (2.7)
and,
q
l
= k
fl
N
s
=1
[X
l
k
rl
N
s
=1
[X
l
. (2.8)
Here [X
, k
fl
and k
rl
are, respectively,
the forward and reverse rate coecient for the l
th
reaction and they are related through
an equilibrium constant. The rate coecients are commonly expressed in Arrhenius form
as,
k
fl
= A
fl
T
b
fl
exp
E
A,fl
R
u
T
, (2.9)
where A
fl
is the pre-exponential factor, b
fl
is the temperature exponent, and E
A,fl
is the
activation energy. These parameters usually are determined by experiments. Inspection
of the rate expression shows that in order to compute the chemical source term of each
12
species, all the elementary reactions in the mechanisms have to be considered, in general.
This is tedious for a mechanism with hundreds or thousands of reactions.
The third expense in chemistry calculations results from the fact that there is a
large range of chemical time scales inherent in a reaction mechanism, typically ranging
from 10
10
to 1 second or more in combustion problems [32] . Mathematically, chemical
time scales can be dened by the inverses of the absolute values of the eigenvalues of
the Jacobian matrix dS
/dY
dt
= S
Y , T, p) for = 1, 2, . . . , N
s
. (2.10)
The ratio between the largest and smallest time scales characterizes the degree of sti-
ness of the ode set. The smallest time scales have to be resolved in the numerical
solution, even if one is interested only in the slow processes. Otherwise the numerical
solution tends to become unstable. The wide variation in time scales makes the above
odes very sti and severely increases the cost of solving the equations.
These three observations together illustrate why in a practical combustion calcu-
lation with a moderately detailed chemical mechanism, over 90 percent of the CPU time
may be spent on chemistry calculations. This poses a serious obstacle to combustion
simulations. For example, in a nite-volume simulation of an axisymmetric laminar jet
ame with a 122-species mechanism, the CPU time for the integration of equations (2.10)
is about 0.06 second per time step on a 2.8 GHz Intel Xeon processor. If we use a mod-
erately ne mesh of 5,000 cells, then 2,000 time steps would require 10
7
integrations,
which corresponds to seven days of CPU time for just the chemistry. This illustrates
that detailed chemistry calculations are impossible or extremely expensive to implement
in three-dimensional time-dependent CFD simulations of practical turbulent combustion
systems.
13
2.3 Strategies in Detailed Chemistry Calculations
Over the past decade, signicant progress has been made in the ability of compu-
tational models to model both chemical kinetics and uid mechanics in turbulent reacting
ows. Describing turbulent ames using full chemical kinetics and fully resolved uid
dynamics is still computationally infeasible due to time and memory constraints. Various
strategies have been developed to address this issue and these strategies can be classied
into three categories:
1. Simplication of the ow eld physics;
2. Reduction of chemistry;
3. Storage/retrieval schemes for using both detailed descriptions of chemistry and
uid dynamics.
2.3.1 Simplication of the Flow Field Physics
Simplications of the ow eld physics in turbulent ames have been made possi-
ble by the dramatic developments in experimental combustion diagnostics in recent years,
particularly in the use of lasers. These developments have led to an ability to make de-
tailed measurements within turbulent ames, from which it is possible to construct a
detailed picture of the structure of the ames [35]. Based on insightful experimental
observations, models that capture the salient features of combustion uid dynamics have
been proposed. One example is the Two-Stage Lagrangian (TSL) model [15] of turbulent
jet ames. The advantages of such models reside in the fact that they permit the use of
arbitrarily complex chemistry while retaining the main features of the ow elds. Dis-
advantages are that although the use of experimental correlations relax the constraints
imposed by computer time and memory, it precludes using these models as quantitative
simulation tools for detailed predictions of ame structure.
2.3.2 Reduction of Chemistry
The reduction of complex chemical systems has long been desired in combus-
tion simulations. Besides the driving force obtained from computational complexity in
14
combustion simulations, there are the large uncertainties in determining some elemen-
tary chemical kinetic rates, which may exceed three orders of magnitude [36]. Reduced
chemistry enables the simulations of the ames that could not be simulated otherwise,
and therefore enhances the understanding of these ames and provides more accurate
predictions or more educated guesses of the ame structures. By introducing approxi-
mations, however, reduced chemistry looses the ability to describe combustion chemistry
in a full, accurate, and general manner. Therefore they fail to address many subtle is-
sues in combustion applications and sometimes can only be applied to a limited range
of thermo-chemical conditions. Nevertheless, reduced chemistry is used predominately
in current combustion simulations.
Numerous approaches have been devised for the reduction of complex chemical
systems. Here, the reduction of complex chemical systems refers not only to the re-
duction of reaction mechanisms and number of the chemical species, but also to other
methods that lower the dimension, or number of degrees of freedom, in the description of
the time evolution of a chemical system. These other methods include the rate-controlled
constrained-equilibrium (RCCE) method [37, 38], the computational singular perturba-
tion (CSP) method [39, 40], and the intrinsic low-dimensional manifolds (ILDM) method
[41, 42, 43]. Compared to reaction mechanism reductions, which are applied through-
out the computational domain and time, these other methods can be viewed as reduced
mechanisms with time- and space-varying stoichiometric coecients, rate parameters,
and even number of reactions, although they usually do not provide reduced reaction
mechanisms in an explicit form. Therefore, mechanism reduction can be termed as
global chemistry reduction, and the other methods as local chemistry reduction.
All reduction methods aim at reducing the number of dierential equations re-
quired to solve the chemical system; this usually is the bottleneck in turbulent com-
bustion calculations. The techniques frequently employed include sensitivity analysis
(used to identify the rate-limiting reaction steps), reaction ow/ux analysis (used to
determine the characteristic reaction path), and eigenvalue/vector analysis (used to de-
termine the characteristic time scales and directions of the chemical reactions) [32]. In
15
reaction mechanism reductions, quasi-steady-state and/or partial equilibrium assump-
tions [9] commonly are employed to identify species in steady state and reactions in
partial equilibrium; these are then used to eliminate species and/or reactions that can
be represented by algebraic expressions. Reviews on mechanism reduction can be found
in [44, 45], and the general procedure of applying this approach is described in [46, 36, 31].
In local reduction methods, CSP and ILDM are based on approaches from dynamic sys-
tem theory, while RCCE is based on the maximum entropy principle of thermodynamics.
Both CSP and ILDM employ eigenvalue analysis to identify the fast time scales of the
chemical reacting systems, the dimensions of which can then be reduced. However, CSP
aims at decoupling the slow and fast time scales to remove the stiness of the odes (equa-
tion 2.10) of the chemical system, while ILDM aims at describing the chemical system
with only a small number of reaction progress variables. RCCE assumes that a non-
equilibrium reacting system will relax to its nal equilibrium state through a sequence
of rate-controlled constrained equilibrium states that can be determined by maximizing
the entropy subject to the instantaneous values of the constraints. Thus only the rate
equations for the constraints, the number of which is much smaller than the number of
species, must be integrated. The accuracy of RCCE can be systematically improved by
adding constraints imposed by increasingly faster reactions. All these reduction methods
take a detailed mechanism as input. CSP and ILDM can be automated. For mecha-
nism reduction and RCCE, a computer program can be written to realize automated
chemistry reduction [47, 38].
2.3.3 Storage/Retrieval Scheme
The storage/retrieval schemes [48] comprise the third category of strategies in de-
tailed chemistry calculations. The basic idea of storage/retrieval schemes is that during
the course of solving the time evolution of the chemical system represented by equa-
tions (2.10), the information generated at each integration time step (such as the initial
conditions and the solutions), are stored in a specially organized table. When similar
initial conditions are encountered, instead of performing an expensive direct integration
of the odes, a retrieval or interpolation of the data in the table is performed with a CPU
16
time that is much less than that of the direction integration. For this approach to be
workable, table reuse must be suciently high. Intuitively, this should be the case when
the system approaches equilibrium or steady state. Mathematically, this has been shown
to be true for many situations in combustion simulations with the aid of the concepts
introduced in the ILDM method [41, 48, 32]. ILDM is a useful concept for understanding
the evolution of reacting systems, in addition to providing a method for chemistry reduc-
tion. Its conclusions regarding chemical composition space establish the foundation for
various storage/retrieval methods. Together with the understanding obtained via ILDM
for reacting ows, the storage/retrieval scheme constitutes an eective solution to the
problem of using detailed chemistry in CFD simulations.
Several methods that have been developed recently fall into the storage/retrieval
category. These methods can be grouped into methods based on tabulation, such
as structured look-up table (LUT) [49], in situ adaptive tabulation (ISAT) [48], and
database on-line for function approximation (DOLFA) [50]; methods based on neural
networks [51, 52]; and methods based on orthogonal polynomials [53, 54, 55]. General
criteria by which storage and retrieval methods can be judged are the CPU time re-
quired to create the store; the memory required for the store; inaccuracy in the retrieval
(or interpolation error); the CPU time required for the retrieval/interpolation; and the
degree to which the method is generally applicable and automated [48].
Among these storage/retrieval methods, one of the most successful and promis-
ing is in situ adaptive tabulation. ISAT achieves an ecient solution of the reaction
equations through the dynamic creation of a look-up table based on direct integration
results and an accuracy controlled retrieval based on eigenvector analysis of the reacting
system, thus allowing for the implementation of detailed chemistry in turbulent reacting
ow calculations. Successful application of ISAT to a detailed chemical mechanism of
160 species and 1540 reactions has been reported [56]. ISAT will be discussed further in
section 2.4.
17
2.4 In Situ Adaptive Tabulation (ISAT)
The primary concern of ISAT for reacting ow calculations is the ecient solution
of the reaction equations (2.10). The essential features of ISAT that set it apart from
other storage/retrieval methods are on-the-y (or dynamic) tabulation of reaction data
as needed (in situ), unstructured and adaptive tabulation of data of unknown topology,
and explicit control of retrieval/interpolation error. It is these features that make ISAT
suitable to facilitate implementation of detailed chemistry, or chemistry acceleration, in
turbulent reacting ows. The basic idea and operations of ISAT are described in the
paper by Pope [48].
At any spatial location and time in a reacting ow, the thermo-chemical state, or
the composition, of the gas mixture can be characterized by the mass fractions Y
( =
1, 2, . . . , Ns) of N
s
species, the enthalpy h or temperature T, and the pressure P. These
are called the composition variables and can be written in vector form as
=
1
,
2
, . . . ,
N
s
+2
. (2.11)
There are dependences among the components of
; for example, the mass fractions sum
to unity. Therefore, if the number of independent components is assumed to be D, then
the thermo-chemical state of uid is completely determined by
=
1
,
2
, . . . ,
D
. (2.12)
One set of values of
can be considered as one point, or a vector, in the D-
dimensional composition space. All physically possible values of
dene the realizable
region of the composition space, which is a convex polytope in D-dimensional space. The
accessed region of the composition space is dened as the set of all compositions that
occur in a particular ame calculation; it is much smaller than the realizable region.
To illustrate this, consider a location where the temperature and pressure are 1,800 K
and 1 atm, respectively. In principle, the mass fraction of fuel species can be any value
between zero and unity; however, under these conditions, its value is almost certain to
18
be zero. Therefore, the accessed region is a small subset of the realizable region, and
is an intrinsic low-dimensional manifold (ILDM). This observation forms the foundation
on which various storage/retrieval methods are based.
ISAT addresses the ecient solution of the reaction equations (2.10) over a time
step starting from the initial condition
(t
0
) =
0
, where temperature and pressure are
taken to be constant without loss of generality. The composition
at time t
0
+ t is a
unique function of
0
; this is called the reaction mapping and is denoted by
R(
0
). The
conventional approach to determine the mapping is to integrate the reaction equation
numerically using a sti ode solver such as DVODE [57]. That approach is referred to
as direct integration, or DI. For a detailed mechanism, DI is computationally expensive
for the reasons that have been discussed in section 2.2. ISAT alleviates this problem
by building up a table during the course of the reacting ow simulation. Each table
entry corresponds to a solution evolved from a certain initial condition. As the table,
or as the intrinsic low-dimensional manifold, becomes populated, computation eciency
is realized by solving the reaction equation though retrieving and/or interpolating from
the existing solutions instead of performing expensive DI.
The coupling of ISAT with a reacting ow simulation code is illustrated in Fig-
ure 2.1.
The ISAT algorithm works as follows. Every time the ow code needs a solution,
or a mapping, of the reaction equation, a query is sent to the ISAT module. At the
beginning, the ISAT table is empty, so a DI is performed and the resulted mapping is
forwarded to the ow code and is stored in the table together with its initial condition.
The store corresponds to a generation of a table entry, or a record, and it is termed an
add. During an add, computed and stored as well is the mapping gradient
A(
0
), which
is the Jacobian of the mapping
R(
0
) and dened as
A
ij
=
dR
i
(
0
)
d
j
. (2.13)
The mapping gradient contains information about the sensitivity of
R to the variations
in
and this information is used for two purposes. First, it establishes a neighborhood
19
Reacting
Flow
Simulation
Code
ISAT
Module
Reaction
Mapping
Calculation
t,
tol
0
A( )
0
R( )
q
R( )
0
R( )
Fig. 2.1. Interconnection between simulation code and ISAT algorithm
centered on
0
, where a linear interpolation suces to provide for any point (initial
condition) in the neighborhood the reaction mapping to an accuracy specied by the ow
code. This neighborhood is called the region of accuracy (ROA). Second, it provides the
coecients necessary to perform the linear interpolation.
For subsequent queries by the ow code with an initial condition
q
, ISAT at-
tempts to nd a ROA that contains the query point
q
. Three situations can occur: if
such a ROA is found, then linear interpolation is performed and the resulting mapping
is returned. This outcome is termed a retrieve. If such a ROA is not found, then a
direct integration is performed. Based on the mapping
R(
q
) from DI and the mapping
gradient, some ROAs are examined to determine if the linear approximation in a certain
ROA for
R(
q
) is suciently accurate. If it is, then that ROA is grown to include the
query point
q
and this outcome is termed a grow. If not, the a new ROA is added and
the outcome is an add.
The CPU time savings of ISAT comes from mainly the economy of the retrieval/
interpolation operations compared to the direct integration. One important aspect is
how fast a ROA can be found that contains the query point
q
. The organization of the
20
record in the table plays an important role in this regard. ISAT adopted a binary tree
[58] structure for its record organization. Each leaf of the tree is a record, which consists
of the tabulation point
0
, the reaction mapping
R(
0
), the mapping gradient
A(
0
),
and information on the shape of region of accuracy (consisting of a unitary matrix and
a vector). Each node of the tree is a cutting plane, which is dened by a vector and a
scalar such that as the tree is traversed from the top for a query composition
q
, the
information given by the vector and the scalar directs the search to either the left branch
or the right branch of the tree. In this manner, the tree is traversed quickly to nd a
leaf that can be used to approximate the query point
q
.
The CPU time savings of ISAT is counterbalanced by the cost of add and grow
operations, since both operations include direct integration in addition to other manip-
ulations. Specically, the fraction of total queries that result in adds and grows is a
key parameter that determines the gain of ISAT in computational eciency over direct
integration. This fraction depends both on the problem at hand and the algorithm used
to determine the shape and the extent of a ROA added to the table. In general, ISAT
should be more ecient for state-steady cases than for transient cases, for premixed
problems than for nonpremixed problems, and for homogeneous reactants than for non-
homogeneous reactants. There are parameters that control the total number of adds and
grows, the number of trees used in the table, the error tolerance for linear interpolation,
and the scaling and transformation of the tabulation points and their mappings.
ISAT was developed originally for Lagrangian Monte Carlo PDF-based turbulent
combustion modeling, and most applications to date have been in that context. Most
applications have been limited to homogeneous systems or to statistically stationary
congurations with small to moderate-sized mechanism (fewer than 50 species) [48, 59].
In such cases, speedups compared to direct integration of up to a factor of 1,000 have
been reported. In other cases [60, 61], speedups of 10 to 100 have been found. ISAT
also can be used for grid-based CFD applications, such as nite-volume calculations of
turbulent reacting ow. A dierence with respect to a particle PDF method is that the
number of computational cells used in grid-based CFD calculations is smaller than the
number of particles used in PDF-based calculations by about two orders of magnitude.
21
ISAT can also be used for large mechanisms containing more than 100 species. The
diculty is that the table size becomes quite large and memory usage becomes an issue
(memory scales as square of the number of species). Special measures have to be taken
to work around this issue for large mechanisms. Therefore, the benets of using ISAT for
grid-based CFD calculations and for large mechanisms are expected to be lower than for
PDF-based calculations and for medium-sized mechanisms. Still, examples with various
degrees of success have been reported [62, 56].
2.5 Turbulence/Chemistry Interactions
Most combustion applications take place in a turbulent environment. Turbulent
ow and chemical kinetics are among of the most challenging problems in non-linear
physics. The strong nonlinear interactions between turbulence and chemistry make tur-
bulent combustion even harder to understand. Turbulence/chemistry interactions have
been a central subject in the research of turbulent reacting ows.
Turbulence/chemistry interactions (TCI) arise from the fact that mixing processes
in turbulent ow are not fast compared with the rates of chemical reactions. The time
scales of chemical reactions can range from 10
10
s to more than 1 s [32], while the
time scale of turbulent mixing typically is no smaller than 10
3
s or 10
4
s [48]. TCIs
arise also from the fact that large spatial and temporal variations in species composition
and temperature occur in turbulent combustion. In nonpremixed combustion, turbulent
mixing creates pockets of fuel-rich and fuel-lean mixture, while in premixed combustion,
turbulent mixing creates pockets of cold unreacted and hot reacted mixture. The mean
chemical reaction rate can not be evaluated directly from the averaged values of species
composition and temperature [63].
The eect of turbulence on chemical reactions takes place through the large-scale
stirring motions of turbulence. By stretching and curvature, the large-scale motions of
turbulence enhance greatly the molecular diusion rates of chemical species and heat,
and therefore enhance greatly chemical reactions, which occur at molecular scales and
must be balanced by molecular diusion. The dierential diusion of chemical species
also can become important. In addition, the turbulent motions produce large variations
22
in species composition and temperature, which cause the mean reaction rates to be
strongly coupled to molecular diusion at the smallest scales of the turbulence.
The eect of chemical reaction on turbulence takes place through the large heat
release from chemical reactions. On one hand, the large heat release produces large
density variations, which inuence greatly the solution of Navier-Stokes equations.The
resultant large density gradients produce a source term in the vorticity transport equa-
tion and therefore enhance the intensity of turbulence. On the other hand, the heat
release produces local dilatation, which acts as a sink term in the vorticity equation and
therefore reduces the intensity of turbulence. Furthermore, the heat release produces
high temperature regions and therefore high viscosity regions, which enhance the dif-
fusion of the vorticity in the turbulent ow eld. The net eect of heat release on the
turbulent intensity depends on the specic conditions in which chemical reactions occur
in turbulent ows.
Great success has been achieved in applying the conservation equations (2.12.4)
and detailed chemistry to laminar ames. This demonstrates that our knowledge of
chemical kinetics and molecular transport processes is sucient for accurate predictions
from rst principles. Turbulent reacting ows are governed by the same conservation
equations and the same chemical kinetics (a chemical mechanism for turbulent ames
should not be dierent from one for laminar ames, since chemical reactions occur at
molecular level [64]). However, turbulent reacting ows are characterized by a broad
spectrum of length and time scales and a complete numerical simulation of such ows
must resolve the smallest and the largest of all these scales. Direct numerical simulation
(DNS) is an established technique for that purpose.
DNS involves solving the conservation equations directly with extremely ne
meshes to fully resolve all relevant length and time scales and to provide a sequence
of realizations of the ow that contains a vast amount of details. DNS has been ap-
plied to simple turbulent ows with simple reactions and to the study of fundamental
processes involved in turbulent ames, e.g. [65, 66]. It is a powerful research tool from
which much can be learned about the physics of turbulent combustion and from which
23
turbulent combustion models can be developed and calibrated. However, even in nonre-
acting ows, DNS can provide useful information only for simple geometry ows at low to
moderate Reynolds numbers. In reacting ows, the composition elds introduce length
and time scales that may be much smaller than those of the velocity elds. The species
conservation equations with highly nonlinear reaction source terms are more dicult
to solve than the incompressible Navier-Stokes equations. Therefore for the foreseeable
future, DNS will not be feasible for accurate predictions of practical turbulent ames.
Even if a DNS solution of a practical turbulent ame could be obtained, the large
amount of detail in time and space would be overwhelming and of little practical interest.
In practice, it is the mean quantities, such as the mean fuel consumption rate and mean
pollutant formation rate, that typically are desired. This requires limiting the dynamic
range of length and time scales in the problem and that is accomplished by applying
various averaging techniques to the conservation equations. Favre averaging (density-
weighted averaging) is usually employed because of large uctuations in density due to
chemical heat release. Averaging reduces greatly the number of degrees of freedom in the
problem, but it introduces unclosed terms that need to be modeled. The Favre-averaged
conservation equation are written as follows [16]:
mass
)
t
+
) u
i
x
i
= 0 (2.14)
momentum
) u
j
t
+
) u
j
u
i
x
i
=
)
j
u
i
x
i
+
ij
)
x
i
p)
x
j
+g
j
) for j = 1, 2, 3 (2.15)
species
)
t
+
)
u
i
x
i
=
)
i
x
i
i
)
x
i
+)
for = 1, 2, . . . , N
s
(2.16)
24
enthalpy
)
h
t
+
)
h u
i
x
i
=
)
i
x
i
J
h
i
)
x
i
+
Dp)
Dt
+
ij
u
j
x
i
) +
Q
rad
). (2.17)
Here the angle bracket <> and the tilde operators denote the conventional Reynolds
averaging and Favre averaging, respectively. The superscript
denotes uctuations from
the Favre-averaged mean.
The eects of turbulence on chemistry calculations are embodied in the unclosed
terms: turbulent uxes of species ()
i
) and energy ()
i
), and the mean species
chemical reaction rates (
i
=
t
Sc
t
x
i
(2.18)
enthalpy
)
i
=
t
Pr
t
h
x
i
. (2.19)
Here
t
is the turbulent viscosity estimated from a turbulence model, and Sc
t
and
Pr
t
are the turbulent Schmidt number for the species and turbulent Prandtl number,
respectively.
Theoretical [68] and experimental work [69] have shown that turbulent diusiv-
ity for reacting scalars is quite dierent from that for nonreacting cases and further
investigations are required. Nevertheless, these hypotheses are employed ubiquitously
in practice. Alternative treatments for the turbulent uxes that provide more accurate
results include second-order transport models and PDF methods [3].
The determination of the mean reaction rates is the central problem of practical
interest in turbulent combustion simulations. The various approaches for modeling the
mean reaction rates are generically referred to as turbulent combustion models.
25
Turbulent combustion models can be divided into two major categories according
to the relative time scales for chemistry compared to other physical processes: fast-
or-slow-chemistry models and nite-rate-chemistry models. In each category, turbulent
combustion models are further divided into models for premixed, nonpremixed, or par-
tially premixed reactants. Although quite dierent in appearance, the various models
essentially attempt to represent the same physical processes. Common links among these
models have been explored, for example, by Veynante and Vervisch [63]. The wide vari-
ety of combustion model reects the diculties that arise in averaging the reaction rate
terms in the species equations.
Fast-or-slow-chemistry combustion models essentially ignore the interactions be-
tween turbulence and chemistry by assuming that the time scales associated with chem-
ical reactions are very short or very long in comparison with time scales for turbulent
mixing. That is, mixing or chemistry is the rate-determining process. In the fast- or
slow-chemistry limit, models are developed based on dierent concepts. First, a simple
and widely used approach is to express the mean reaction rate in the same functional
form as that of the unaveraged reaction rate. Because the latter usually are expressed
in Arrhenius form, this is sometimes referred to as the Arrhenius model [3]. Second,
based on the concept of a turbulence energy cascade in nonreacting ows, eddy-breakup
(EBU) models [70] and variants, such as the eddy-dissipation model [71], link the mean
reaction rate to the rate of turbulent mixing and hence to a turbulence time scale. Third,
based on the concept of laminar amelets [72], a variety of amelet models have been
formulated for premixed, nonpremixed, and partially premixed combustion regimes [68].
Classical amelet models that fall into this category include the conserved scalar equilib-
rium models (CSEM) [68] for nonpremixed ames, the Bray-Moss-Libby (BML) model
[73, 74] and the coherent ame model (CFM) [75, 76] for premixed ames. These fast-or-
slow-chemistry models sometimes provide plausible predictions, but they cannot capture
the nite-reaction-rate eects that are often the focus of interest.
Finite-rate-chemistry combustion models treat chemistry more realistically by
considering the full range of chemical time scales and by dealing explicitly with the in-
teractions between turbulence and chemistry. However, special mathematical approaches
26
have to be employed to handle the strong nonlinearities in determining the mean reac-
tion rates. Statistical approaches are most suitable to this task. These include models
based on PDF methods and models based on conditional moment closure (CMC) meth-
ods. The PDF-based models can be further divided into transported PDF models, where
a modeled PDF transport equation is solved, usually by means of a Lagrangian Monte
Carlo particle-based method [77, 78, 14], and presumed PDF models, where the shape or
form of the PDF is parameterized and modeled equations are solved for the parameters
[3, 17]. Here we shall use the term PDF method to refer to a transported PDF method.
The CMC method [79, 80] considers conditional averages and higher moments of quan-
tities such as species mass fractions and enthalpy, conditional on mixture fraction (for
nonpremixed combustion) or reaction progress variable (for premixed combustion). The
concept of CMC and its applications have been reviewed in a recent paper by Klimenko
and Bilger [81].
Laminar amelet concepts also can be applied to the modeling of nite-rate chem-
istry. Flamelet models based on a scalar G-equation for premixed combustion and
based on mixture fraction for nonpremixed ames have been developed and widely used
[68]. Another approach to account for nite-rate chemistry is the linear eddy model
(LEM), which was rst formulated for nonreacting ows and later was extended to react-
ing ows [82, 83]. LEM is a method of simulating molecular mixing on a one-dimensional
domain embedded in a turbulent ow and therefore is capable of capturing molecular
eects in turbulent reacting ows.
Finite-rate-chemistry combustion models have been a key research subject in the
simulation of turbulent combustion, and the above overview oers only a brief catego-
rization of the most popular models currently in use. More detailed descriptions and
formulations can be found in many dedicated books and monographs, such as Libby and
Williams [17], Peters [68], and Haworth [3].
The above discussion is based on invoking a statistical averaging technique to re-
duce the dynamic range of scales in turbulent reacting ows. This approach is commonly
referred to as Reynolds-averaged Navier-Stokes (RANS) modeling. With appropriate clo-
sure models, the RANS approach allows only for the determination of mean quantities,
27
which may dier largely from the instantaneous values. Strong unsteady mixing eects
are observed in turbulent ames and the knowledge of steady statistical means may not
always be sucient to describe the turbulent combustion phenomena of interest. An
alternative approach to reducing the dynamic range of scales while accounting explicitly
for the unsteady eects is a spatial ltering technique. This is the approach that is taken
in large eddy simulation (LES).
In large eddy simulation [67, 68, 84, 85], the larger energy-containing scales are
resolved explicitly while the eects of unresolved smaller scales are modeled. The distinc-
tion between the resolved large scales and the modeled small scales usually is determined
by the grid resolution that is aordable in the computation domain. The small scales
are the subgrid scales and the models sometimes are referred to as subgrid models.
Mathematically, the large and small scales are separated by ltering the instantaneous
governing equations. The resulting ltered equations contain unclosed terms that need
to be modeled. The modeled equations are solved numerically to simulate the unsteady
behavior of the large-scale motions. Compared to RANS, LES provides information on
the large resolved scales, which is valuable in many practical applications such as com-
bustion stabilities in a gas turbine combustor. As in RANS, the interactions between
turbulence and chemistry occur at unresolved scales of computation. Therefore, the
basic tools and formalism of RANS-based turbulent combustion modeling can carried
directly to LES. Most of the RANS combustion models discussed above can be modied
and adapted to LES subgrid-scale modeling [63].
2.6 Summary
Issues that arise in calculating detailed chemistry in turbulent reacting ows have
been discussed in this chapter. Detailed chemistry is critical for realistic and accurate
predictions of turbulent combustion. However, the characteristics of detailed chem-
istry make the calculations quite dicult. This is mainly due to the complexity of the
chemistry, e.g., large number of species and elementary reactions required, and the broad
28
range of time scales involved, from 10
10
s to more than 1 s. Dierent strategies to over-
come these diculties have been developed, such as simplication of ow eld descrip-
tion, chemistry reduction, and storage/retrieval schemes. One of the storage/retrieval
schemes, the in situ adaptive tabulation, has been described in detail. This is one of the
most promising strategies for implementing detailed chemistry in turbulent combustion
calculations and is the approach that has been adopted for this study. The nature of
turbulence-chemistry interactions also has been discussed. Direct numerical simulation
can provide a complete description of a turbulent reacting ow, but is not feasible in
the foreseeable future for practical combustion systems because of computational power
limitations and the huge amount of data that is generated. Averaging and ltering tech-
niques are employed to reduce the range of scales to be resolved in the ows to make
the simulation tractable. LES is capable of capturing the eects of unsteady mixing by
explicitly resolving large energy-containing scales and modeling the more homogeneous
small scales. RANS models can only provide mean quantities but are sucient for many
practical applications. Averaging and ltering introduce unclosed terms that need to
be modeled in the governing equations. The various combustion models that have been
developed can be divided into fast-or-slow chemistry and nite-rate-chemistry models.
Fast-or-slow chemistry models are used widely in combustion simulations due to their
simplicity. Finite-rate-chemistry models provide more realistic results but need statisti-
cal tools to accommodate the random nature of turbulent ames. The PDF method with
a Monte Carlo solution scheme is considered to be one of the most promising approaches
for the modeling of nite-rate turbulent combustion.
29
Chapter 3
Soot Calculations
3.1 Introduction
Most combustion processes that involve hydrocarbons produce soot. Under ideal
conditions, combustion of hydrocarbons leads mainly to carbon dioxide and water. Under
practical conditions, in locally fuel-rich regions, the combustion or pyrolysis of hydrocar-
bons generates intermediate species and radicals that lead eventually to the appearance
of soot particles.
Soot consists mainly of carbon. Other elements such as hydrogen and oxygen
are usually present in small amounts. For example, soot emitted from long-residence-
time turbulent nonpremixed ames, including toluene, benzene, acetylene, propylene,
and propane ames burning in air, have the following elemental mole ratio ranges: C:H
of 8.318.3, C:O of 58109, and C:N of 292976 [86]. Soot density is less than that of
carbon black and usually in the range of 17001800 kg/m
3
, depending on the porosity
of soot [87]. Soot particles are generally small, ranging in size from 5 nm to 80 nm, but
may be up to several micron in extreme cases [88]. While mostly spherical in shape, soot
particles may also appear in agglomerated chunks and even long agglomerated laments
[86]. Experiments in diusion ames of hydrocarbon fuels have shown that the soot
volume fraction generally lies in the range of 10
6
10
8
[88]. These physical properties
of soot particles aect their optical properties, which in turn aect the accuracy of
determining soot quantities and radiation in both experiments and simulations.
Soot formation is a complicated phenomenon that involves highly coupled chem-
ical and physical processes. It is remarkable that hydrocarbon fuel molecules containing
only a few carbon atoms transform into soot particles containing millions of carbon
atoms. The study of soot processes in combustion systems has drawn great attention.
Progress has been achieved in understanding the essential features of the chemistry and
30
physics. However, the understanding is incomplete and many questions persist and de-
bates continue regarding the details of soot nucleation, growth, and oxidation. This
reects the diculty of the problem that soot formation poses for combustion systems.
The prediction of soot formation is of interest for the following four reasons:
The formation of soot stems from incomplete combustion, which reduces the com-
bustion eciency.
Soot is a major pollutant. It contains trace elements that have hazardous eects
on human health. The particles themselves also are an issue.
Industrial applications, such as furnace and heat generators, require formation
of soot to enhance the heat transfer via radiation. However, the soot has to be
oxidized before these devices release the exhaust into the environment.
Soot is an important industrial product refered to as carbon black. It nds wide
application, such as ller in tires or other materials, toner in copiers, and black
pigment in color printings.
3.2 Soot Formation and Oxidation: Current Understanding
The study of soot formation and oxidation includes experimental, theoretical, and
computational eorts. Most information obtained so far regarding soot formation steps
comes from shock tubes, and laminar premixed and nonpremixed ames. The prevailing
view of soot formation is that the process is a highly coupled chemical and physical
phenomenon and can be divided into four principlal sub-processes: soot particle inception
or nucleation, surface growth, particle coagulation, and particle oxidation. These sub-
processes present in a wide variety of combustion systems ranging from small laboratory
ames to large industrial devices. They can occur simultaneously or to varying degrees
for individual processes depending on the specic combustion environments. Frequently
debated are the details of each sub-process. Good reviews of these issues can be found
in [89, 90, 91]. In the following, a general description of the sooting process is given.
31
3.2.1 Soot Particle Inception
Soot particle inception, or nucleation, is the formation of the smallest solid soot
particles from the gas-phase hydrocarbon molecules of relatively low molecular weight.
These smallest particles have sizes of the order of 1 nm and masses of about 1000 atomic
mass units. Soot particle inception acts as the link between the main gas-phase combus-
tion zone chemistry and soot particle dynamics, and it determines the number of nascent
soot particles in ames.
There are three major proposals regarding the inception pathway that dier
mainly in the key gaseous precursors that are assumed: polyacetylenes, ionic species,
or polycyclic aromatic hydrocarbons (PAHs). The PAH hypothesis is the most generally
accepted based on conclusions from numerous experimental and modeling studies.
The PAH pathway involves the formation and growth of aromatic species. The
starting point is the formation of the rst aromatic ring (benzene) from small aliphat-
ics. These small aliphatics are the building blocks of soot formation; they can be
even-carbon-atom species such as acetylene (C
2
H
2
), or odd-carbon-atom species such as
propargyl radical (C
3
H
3
), or both. Following the formation of rst ring aromatics, they
grow to form large PAHs through the so-called H-abstraction-C
2
H
2
-addition (HACA)
mechanism. HACA involves the abstraction of a hydrogen atom from the reacting aro-
matic molecules by a gaseous hydrogen atom, followed by the addition of a gaseous
acetylene molecule to the radical site so formed. Some of the large PAHs have stable
structures at high temperature. These species are called islands of stability that allow
more and more building blocks to be added via HACA. At a certain size, some PAH
species begin to stick to each other during collision, while individual PAHs keep increas-
ing in size at the same time via HACA. This combined growth by molecular chemical
reactions and physical collisions leads to the appearance of solid particles.
Parallel to aromatics growth is aromatics oxidation. The primary mechanism
seems to be the oxidation of aromatics by oxygen molecules. The oxidation by hydroxyl
radical is unimportant at this early stage of soot formation. However, the role of O
2
in
PAHs oxidation is twofold. On one hand, it suppress carbon mass from further growth
by oxidation. On the other hand, it promotes soot formation by building up the radical
32
pool, especially H atoms for HACA. Soot appearance in the ame is determined by the
competition between soot formation and oxidation even at this early stage of the process.
3.2.2 Soot Surface Growth
Surface growth is the addition of gas-phase materials to the already formed parti-
cles. It inherently involves heterogeneous reactions and that complicates the modeling of
soot formation. Soot surface growth is primarily responsible for the mass accumulated
in soot particles.
Although still uncertain, it is generally accepted that the principal gas-phase
species that reacts at the particle surface is acetylene, although PAHs may also play a
role, and that this deposition process follows a rst-order rate law. This understanding
is widely used in the surface growth models of soot formation.
It is postulated that the heterogeneous reactions taking place on a soot particle
surface are analogous to those of large PAHs (chemical similarity hypothesis). In other
words, the surface growth is assumed to be governed by the HACA mechanism. Many
studies show that the surface growth occurs in environments of acetylene abundance,
thereby supporting this model of surface reactions.
The factors that determine the growth rate include available surface area and the
number of active sites on the surface. Experimental observations indicate that the surface
growth rate decreases with time or with the extent of particle growth. The slowdown of
surface reaction relates to the decrease of active sites and the decrease of the available
surface area for HACA reactions. This phenomenon is termed surface aging.
3.2.3 Soot Particle Coagulation
After the initial soot particles are formed, the physical process of collisions be-
tween these particles leads to the formation of larger soot particles. This is called soot
particle coagulation, and occurs simultaneously with the surface growth process. Unlike
surface growth that changes the total mass of soot particles in ames, coagulation only
changes the evolution of the soot-particle-size distribution.
33
Particle coagulation usually is classied into two distinct processes: coalescent
collision and agglomeration. In coalescent collision, two particles come together and
merge to form a single larger particle, while in agglomeration, two particles stick to each
other to form a chain-like structure but the identity of individual particles is maintained.
The two processes may not be separate in time as often presumed and agglomeration
may begin at the onset of soot nucleation.
Before and after coalescent collisions, particles usually assume spherical shapes.
This process can be described by concepts and mathematical equations from aerosol
dynamics [92]. Three collision regimes can be distinguished according to the ratio of
mean free path to particle size, the Knudsen number, K
n
. These are: the free-molecule
regime for particles of small size; the continuum regime for particles of large size; and the
transition regime for particles in between. The evolution of the particle size distribution
is then determined by the formulations for these three regimes.
During agglomeration, soot particles form large, branched chain-like structures
and are called soot aggregates. Numerous transmission electron microscopy (TEM)
measurements indicate that soot aggregates consist of nearly spherical primary particles
having relatively uniform diameters. Soot aggregates compound the diculties in predic-
tion of soot quantity and in determination of soot optical properties. Recent evaluations
of those TEM measurements show that soot aggregates exhibit mass fractal behavior
[93, 94], a term used to indicate irregular repeated structures. The fractal dimension
seems to be conned to a rather narrow range of 1.71.8 for most soot aggregates formed
in ames. The determination of soot aggregate optical properties and the evolution of
size distributions are based upon this mass fractal approximation.
3.2.4 Soot Particle Oxidation
Soot particle oxidation counterbalances soot surface growth by changing the mass
of solid soot particles back into gas-phase species. It is the only process that reduces the
total amount of soot present in ames and hence the only mechanism that removes soot
emission from the exhaust.
34
Oxidation of soot primarily occurs as a result of surface reactions with oxygen
molecule and OH radical to form CO and CO
2
. For oxidation by O
2
, the expression
of the oxidation rate has been studied by several groups [95, 96, 97]. Although large
uncertainties remain, the expression given by Nagle and Strickland-Constable [96] and
its various improvements (see refs. in [90]) nd widespread use.
The modeling of OH oxidation is less rigorous than that of O
2
since the OH
radicals usually exist at super-equilibrium concentrations in ames. The oxidation rate
by OH is expressed in terms of a temperature-independent reaction probability. The
data given by Neoh et al. [98] from laminar premixed ames and Roth et al. [99] from
shock tubes are usually used for the modeling of OH attack on soot.
3.3 Soot Formation and Oxidation: Modeling
Accurate prediction of soot formation and oxidation can save time and cost in
studies aimed at identifying the parameters that control the soot yield in practical sys-
tems. However, soot formation and oxidation entail highly coupled nonlinear processes of
both chemical and physical nature. Small uncertainties in one process may result in large
variations in other processes and the current description of each sub-process cannot reach
a satisfactory accuracy. Diculties and uncertainties in understanding sooting processes
impede progress in soot predictions. Nonetheless, models of dierent sophistication have
been put forward and applied to many practical systems. These models capture dierent
degrees of the physics of the sooting process and can be divided into three categories
[90]: (1) phenomenological models based on purely empirical correlations, (2) detailed
models based on rst principles, (3) semi-empirical/semi-detailed models.
In the early stage of soot modeling eorts, soot predictions relied solely on em-
pirical correlations. Most modeling eorts focused on gas turbines and diesel engines,
which suer from unacceptable emission of smoke under certain conditions. The empir-
ical correlations attempted to collapse into one expression a variety of eects on soot
yield, such as temperature, pressure, fuel composition, residence time, and even ow
mixing. This resulted in a large number of empirical modeling parameters so that good
agreement with measurements could be ensured. Although economic in computation,
35
this modeling approach has the problem that the validity of the correlations is quite
uncertain for situations that vary only slightly from the conditions where the modeling
parameters are calibrated. Kennedy [90] oers a detailed description of these empirical
models.
Semi-empirical/semi-detailed soot models represent the next level of soot model-
ing. In this case, some aspects of the physics and chemistry of the soot formation and
oxidation process are incorporated explicitly. Unlike empirical models, where the soot
phenomenon is treated globally, semi-empirical models begin to incorporate the details
of sooting sub-processes, such as particle inception, surface growth, and particle oxida-
tion. Rate expressions are developed for each sub-processes, and the rate coecients
for each sub-process are usually adjusted to match experimental data of soot quantities,
or are determined from rate measurements under noname conditions. Semi-empirical
models usually include two dierential equations to predict soot yield, and the quantities
solved for are usually particle number density (with no size distinction) and a quantity
of soot mass, such as soot volume fraction or soot mass fraction. Although still crude
in nature, semi-empirical models help to further understand the sooting process, and
modeling progress reects the convergence of this understanding to the current state.
Kennedy [90] oers a thorough review on semi-empirical models.
Detailed soot models represent the highest level of soot modeling. The common
aim of all such models is the applicability over a wide range of conditions. Unlike
semi-empirical models, where a single rate expression characterizes each sub-process,
detailed models employ detailed chemical kinetic and physical models to describe each
sub-process that occurs in the gas phase, solid phase, and on the surface of soot particles.
Soot-particle inception is described by detailed chemical kinetics that involve tens of
species and hundreds of reactions and cover the formation and growth of PAHs up to
four-ring aromatics or even higher by using a linear lumping technique. The transition
from the gas-phase to solid soot particle usually is assumed to take place at the collision
of pyrene and larger aromatics. Surface growth and particle oxidation are treated with
detailed descriptions that usually are based on chemical analogies to gas-phase aromatic
chemistry. And the coagulation of soot particles is modeled using methods from aerosol
36
particle dynamics. The combination of these detailed descriptions of sooting process
expands greatly the range of conditions to which these models can be applied. One
problem, however, in these models is that many elementary reactions involved come
from the chemistry analogy and some of reaction rates are purely estimations. This
impairs the usefulness of detailed models, although progress is being made continuously.
Frenklach [91] and Kennedy [90] oer good reviews on these models.
In semi-empirical models, the total particle number density is computed and no
knowledge of the particle size distribution can be obtained. In detailed models, the use
of methods from aerosol dynamics is feasible and this allows the determination of the
soot particle size distributions in a ame. This is an advantage because the soot size
distribution aects the optical properties of a soot cloud and the radiative heat loss from
a ame. The methods used most frequently for the soot particle dynamics include the
discrete-sectional method [100, 101, 102] and the method of moments [103, 104, 105,
106, 107, 108]. In the discrete-sectional method, a certain number of bins in the size
spectrum of the entire particle ensemble are taken to characterize the size distribution,
and the particle properties are averaged within each bin. In the method of moments,
the particle dynamics processes are formulated in terms of moments of the particle size
distribution function (PSDF). The method of moments has the merit of computational
economy even with its full accounting of the particle size distribution. The method of
moments is the approach that is adopted in this study. We discuss the method in detail
in the next section.
3.4 Method of Moments
The method of moments (MOM) is a numerical approach to the solution of a
particle dynamics system. Particle dynamics describe the time evolution of a particle
ensemble undergoing simultaneously several particle processes, such as nucleation, co-
agulation, and surface growth. MOM is mathematically rigorous while computationally
economical. It was rst introduced by Dobbins and Mulholland [109] in 1984 and since
then the method has undergone further development to account for various aspects of
particle dynamics by Frenklach [103, 104, 105, 106, 107, 110] as well as extensive testing
37
in reactive ow simulations of practical systems (eg. [110, 111, 112, 113, 114]). In the
following, we introduce the basic idea and formulation of MOM.
In a polydisperse particle system where particles have many dierent sizes, the
most important physical characteristic of the system is its particle size distribution.
The distribution can take two forms: a discrete size distribution or a continuous size
distribution. For a soot particle ensemble in a ame, the size distribution is discrete and
we can assume that the particles of size i have volume of v
i
= iv
1
and mass of m
i
= im
1
,
where v
1
and m
1
are the volume and mass of the smallest soot particle that is formed
from the collision of two PAHs (e.g., pyrene). If we consider only the coagulation process,
the time evolution of the soot population can be described by Smoluchowshis master
equations [92] as
dN
1
dt
=
j=1
1,j
N
1
N
j
, (3.1)
dN
i
dt
=
1
2
i1
j=1
j,ij
N
j
N
ij
j=1
i,j
N
i
N
j
for i = 2, . . . , . (3.2)
Here N
i
is the particle number density (number of particles per unit volume) of size
class i, and is the collision coecient. N
i
depends on particle size, spatial location,
and time, and its functional form is the particle size distribution function (PSDF). The
collision coecient is a function of colliding particle sizes and of temperature. The
negative terms on the right-hand side of the equations represent the deduction of particles
of size i due to their collisions with other sized particles to form larger particles, and
the positive terms represent formation of size i particle due to the collisions of smaller
particles. The dependence of on the size of colliding particles inhibits a closed-form
solution of equations (3.1) and (3.2).
It is instructive to consider the various moments of the PSDF, since they have
great physical importance. We dene the r
th
moment of the PSDF as
M
r
=
i=1
m
r
i
N
i
, (3.3)
38
where m
i
is the particle mass of size class i. The zeroth moment of PSDF is the total
number density of soot particles in a ame at a given position and time, the rst moment
is the total mass density of soot particles, and so on. In principle, the knowledge of all
the moments (r = 0, 1, . . . , ) is equivalent to the knowledge of the PSDF itself. In most
practical applications, however, the properties of interest are fully determined by just
the rst few moments. Therefore, instead of solving an innite number of dierential
equations for the moments of PSDF, we need to consider only a small number of equations
for the lowest-order moments. This is the source of the numerical economy of the method
of moments.
To obtain the moment equations, we simply multiply equations (3.1) and (3.2) by
m
r
i
and then sum over the size classes:
dM
0
dt
= G
0
=
1
2
i=1
j=1
ij
N
i
Nj , (3.4)
dM
1
dt
= G
1
= 0 , (3.5)
dM
r
dt
= G
r
=
1
2
r1
k=1
r!
k!(r k)!
_
_
i=1
j=1
m
k
i
m
rk
j
ij
N
i
Nj
_
_
for r = 2, 3, . . . . (3.6)
The summations on the right-hand sides of the equations (abbreviated as G
r
) are the co-
agulation source terms; here coalescent collisions are assumed. The number of equations
is suciently taken to be six (r = 0, 1, . . . , 5) for most practical systems [108].
Resolving the summations is the mathematical diculty of the method of mo-
ments, since the collision coecients
ij
are nonadditive and assume a complex functional
form. Furthermore, the functional form is dierent in dierent coagulation regimes. The
coagulation regime is classied on the basis of the Knudson number, K
n
, [92], which is
dened as K
n
= 2
f
/d, where
f
is the gas mean free path and d the particle diameter.
There are three regimes: the continuum regime for K
n
1, the free-molecule regime for
K
n
1, and a transition regime for intermediate values of K
n
. In the continuum limit,
39
the collision coecient for coalescent collisions of spherical particles is give by [115],
c
ij
= K
c
_
_
C
i
m
1/3
i
+
C
j
m
1/3
j
_
_
_
m
1/3
i
+m
1/3
j
_
, (3.7)
where subscript c denotes continuum regime, K
c
is a coecient that contains the tem-
perature dependence, and C is the Cunningham slip correction factor [92] that relates
to the Knudson number. Substitution of the expression (3.7) into the source term for
the zeroth moment gives,
G
c
0
= K
c
_
M
2
0
+M
1/3
M
1/3
+K
c
_
M
1/3
M
0
+M
1/3
M
2/3
__
, (3.8)
and
K
c
= 2.514
f
_
p
/6
_
1/3
,
where
p
is the particle material density and the denition of moment (equation 3.3) is
extended to include any real value of r. Introducing the reduced moments
r
= M
r
/M
0
, (3.9)
we obtain,
G
c
0
= K
c
_
1 +
1/3
1/3
+K
c
_
1/3
+
1/3
2/3
__
M
2
0
. (3.10)
Similar but more complex formulae can be obtained in the same manner for the
higher moments (G
c
r
). To achieve closure of the moment equations, the fractional-order
moments (
1/3
, etc.) have to be determined in terms of the integer-order moments
that are solved. The simplest way to accomplish this is to presume the functional form
of the PSDF (for example, a log normal distribution). However, the PSDF for soot
is not known a priori in most combustion systems. An alternative approach to close
the moment equations is to interpolate among instantaneous values of the integer-order
moments; the latter are available at each integration step of the moment equations. The
40
validity of moment interpolation has been established using series analysis [108], and
it is suggested that a higher accuracy is obtained by separating the interpolations for
positive-order and negative-order moments. This method of moments with interpolation
closure is called MOMIC by Frenklach [108].
For a particle ensemble undergoing simultaneous nucleation, coagulation, and
surface growth, the rates as well as the moments of nucleation, coagulation, and surface
growth are additive. Therefore, the dynamics of this particle ensemble are described by
[92, 108],
dM
0
dt
= R
0
G
0
, (3.11)
dM
1
dt
= R
1
+W
1
, (3.12)
dM
r
dt
= R
r
+G
r
+W
r
for r = 2, . . . , r
max
, (3.13)
where R, G, and W are the nucleation, coagulation, and surface growth source terms,
respectively. For nucleation, the source terms are derived based on the rate of particle
nucleation, which come either from empirical correlations or from detailed kinetics of
chemical precursors. For coagulation, the source terms for free-molecule and transition
regimes can be developed in the same way as shown above for the continuum regime,
although the source terms for the transition regime can be more eciently expressed as
the harmonic mean of the limiting values; that is
G
r
=
G
f
r
G
c
r
G
f
r
+G
c
r
for r = 0, 2, 3, . . . , (3.14)
where superscript f and c refer to the free-molecule and continuum regimes, respectively.
For surface growth, the source terms are derived based on the mass deposition rates
from gas-phase chemical species according to the state-of-the-art understanding of this
process. The surface growth terms so derived can also comprehend the particle oxidation
process, which removes mass from particles by reaction with gas-phase species. The latest
development for surface growth terms can be found in Appel et al. [107].
41
The advantages of MOMIC are its rigorous description of particle dynamics and
its numerical economy. One needs to solve just a few dierential equations for mo-
ments. This becomes critical in modeling complex systems, such as turbulent reacting
ows, where particle dynamics constitute only a part of the problem. However, MOMIC
does have shortcomings. The present formulation of MOMIC does not consider particle
fragmentation; that is, when the surface oxidation is at the verge of removing all parti-
cle material, one has to handle the decomposition of the smallest particle into gaseous
species. This may be overcome by solving a time evolution equation for the number
density of the smallest particles as suggested in [108]. Also, the above description of
MOMIC assumes coalescent coagulation and spherical-shaped particles. In many com-
bustion environments, soot agglomerates form large aggregates of mass fractal geometry
[86]. This regime of coagulation is dealt with by an extension of MOMIC [106], where
the moments for the number of primary particles in soot aggregates are introduced and
additional moment equations are solved.
3.5 Soot Calculations in Turbulent Flames
The simulation of soot formation and oxidation is a dicult subject because it
involves highly coupled nonlinear chemical and physical processes and these processes
are not well understood. Soot calculations in turbulent ames are even more challenging
because of the intimate coupling among turbulent mixing, ame structure, soot processes,
and radiative heat loss.
Incorporation of soot into turbulent ame simulations has drawn great attention
because of its practical importance. Kennedy [90] oers a good review of work up to
1997. Early work commonly employed semi-empirical soot models, which represent soot
sub-processes in term of rate expressions. One or two additional transport equations are
solved for soot number density and soot mass fraction. The eects of turbulence on the
mean rate expression were either ignored or were accounted for using an laminar amelet
approach. Radiation was usually treated in an ad hoc manner by adjusting parameters
in the temperature calculations.
42
Recent advances in modeling of turbulence/chemistry interactions and in under-
standing of soot processes have inspired another round of eorts to couple soot into tur-
bulent ame calculations. Various closure techniques developed for TCIs have been ap-
plied to soot calculations in turbulent ames. Laminar amelet libraries with presumed
PDFs for mixture fraction and scalar dissipation rate have been a popular approach to
deal with turbulent uctuations in soot production rate calculations, e.g. [113, 116, 117].
Soot formation rates (rather than soot volume fraction, for example) are usually stored
in the amelet libraries as a function of mixture fraction, dissipation rate, pressure, and
temperature. The conditional moment closure approach has been used to apply soot
models derived from laminar ames to turbulent ames and to predict dierential dif-
fusion of soot particles [118]. A linear eddy model for homogeneous box turbulence has
been used to provide information on the joint PDF of soot volume fraction, temperature,
and mixture fraction [119]. The Probabilistic-Eulerian-Lagrangian (PEUL) model has
been extended to include soot formation, and captures the eects of turbulent uctua-
tions on soot yield [120]. A transported PDF method is the most promising approach,
but remains dicult to implement. Its usage for soot calculation has been investigated
in a partially stirred plug-ow reactor [114]. Large eddy simulation of soot formation
has been formulated and conducted to study the coupled turbulence, soot chemistry,
and radiation interactions [121, 122]. In these eorts, semi-empirical soot models are
still widely used, except in some cases where a laminar amelet approach with a detailed
soot model has been employed. In that case, the detailed soot model is precomputed
by solving one-dimensional amelet equations before its coupling to the turbulent ow
solver. Two transport equations for soot mass, or volume, fraction and soot number den-
sity usually are solved. Radiative heat loss is commonly accounted for in the optically
thin limit, neglecting the absorption of soot particles.
Soot formation and oxidation is such a complicated and nonlinear phenomenon
that small errors in one sub-process may yield large errors in other processes. Detailed
soot models are required to study the sensitivities of soot quantities to various parameters
in dierent sub-processes. These sub-processes take place on dierent time scales and
therefore respond dierently to turbulent uctuations. Soot processes depend strongly
43
on ame temperature, while soot aects ame temperature both through its continuum
emission and its absorption of gas-phase radiation. Reliable and accurate soot calcu-
lations in turbulent ames require a comprehensive modeling approach that includes
detailed chemistry and soot modeling, sophisticated radiation calculations, and realis-
tic accounting for interactions among turbulence, soot formation, radiation, and ame
structure.
3.6 Summary
Issues arising from calculating soot formation and oxidation in turbulent ames
have been discussed. Soot emission from a practical combustion appliance often reects
poor combustion conditions and a loss of eciency. It also can cause health problems.
In some situations, however, soot formation in a ame is desirable since it promotes
radiation and hence the eciency of heat transfer from the ame. In that case, soot
needs to be oxidized before the combustion products are exhausted. Soot formation and
oxidation involve highly-coupled nonlinear chemical and physical processes and these
processes are not well understood. Soot is produced due to high temperature pyrolysis
and oxidation. The temperature range of the combustion system where soot is generated
lies typically between 1500 and 2500K. The net amount of soot is a competitive result
of soot formation and oxidation processes. Current understanding is that there are four
key processes: soot particle inception or nucleation, surface growth, particle coagulation,
and particle oxidation. Models that reect this understanding with varying degrees of
sophistication have been developed and successfully applied to laminar ames. These
models can be categorized into three groups: empirical models, semi-empirical models,
and detailed models. Semi-empirical models enjoy widespread applications in current
simulations for their simplicity, but they suer from the problem that model parameters
need to be adjusted for dierent conditions. Detailed models overcome this problem
with their detailed descriptions of each sub-process. They also enable the accurate
determination of the soot particle size distribution in combustion systems. However,
the determination of the soot particle size distribution is computationally expensive.
The method of moments reduces this cost greatly and has been discussed in detail.
44
Soot calculations in turbulent ames are challenging because of the intimate coupling
among turbulent mixing, ame structure, soot process, and radiative heat loss. This
requires a comprehensive modeling approach that includes detailed chemistry and soot
modeling, sophisticated radiation calculations, and realistic accounting for interactions
among turbulence, soot formation, radiation, and ame structure. Current eorts of
incorporating soot models into turbulent ame calculations have been reviewed.
45
Chapter 4
Radiation Calculations
4.1 Introduction
Thermal radiation is an important heat transfer mode in most combustion sys-
tems. By nature, the rate of radiative heat transfer generally depends on the temperature
to the fourth power or higher. This makes thermal radiation dominant over convection
in most ames, especially when soot particles are present. In addition, radiation exerts
its eects at a distance and allows energy to travel directly from the hot product regions
to cold regions such as the reactant mixture and the surroundings. Accurate description
of radiative heat transfer is a crucial element in simulations of turbulent combustion
systems.
The same reasons that make thermal radiation important in ames also make
its calculation complicated. Since radiation is a long-range phenomenon, conservation
of energy must be applied over the entire domain under consideration instead of an in-
nitesimal volume as for mass and momentum conservation. This leads to an integral
equation containing up to seven independent variables: the frequency of radiation, three
space coordinates, two coordinates describing the direction of radiation, and time. Fur-
thermore, radiative properties of combustion products are usually dicult to measure
and often display erratic behavior with wavelength and temperature. Because of these
diculties, most combustion simulations ignore the eect of radiative heat loss, or treat
it in an ad hoc or overly simplistic manner.
Inadequate treatment of radiation can cause large errors in determining ame
structure and pollutant emission. For example, the prediction of NO
x
emission is very
sensitive to the predicted ame temperature distribution, which is inuenced signicantly
by the radiation model. Soot and radiation are highly coupled processes. Errors in
temperature predictions will over- or under-predict soot formation and oxidation rates
46
and therefore soot yields, which will in turn result in erroneous radiative heat losses.
Detailed description of radiative energy transfer is therefore essential to a successful
comprehensive modeling of turbulent combustion processes.
Radiative heat transfer from a ame can be predicted in principle if the radiative
properties and temperature distributions in the ame are available. However, ame
temperature is an unknown variable that is solved for in ame simulations. As a result,
the calculation of radiative heat transfer is coupled to the energy equation through a
heat source term (e.g., the
Q
rad
term in equation 2.4). This radiative heat source can
be expressed as the divergence of the radiative heat ux, q
R
,
q
R
=
_
0
_
4I
b
_
4
I
d
_
d (4.1)
where stands for wavenumber, solid angle,
(r, s, )d,
where r is a position vector xing the location of a point in space, and s is a unit direction
vector described in terms of the polar angle and the azimuthal angle .
A ray of radiative energy going through a radiatively participating medium, such
as a ame, experiences several phenomena that change the intensity of the radiation.
These include absorption, emission, and scattering from the medium. An energy balance
for a ray of radiation leads to the so-called radiative transfer equation. The RTE describes
the radiative intensity eld within an enclosure containing a participating medium as a
function of spatial location, direction, and a spectral variable (wavenumber). The RTE
can be written as [88],
dI
ds
= s I
I
b
+
s
4
_
4
I
( s
i
)
( s
i
, s)d
i
. (4.2)
Here
is the spectral extinction coecient; this is the sum of the spectral absorp-
tion coecient
( s
i
, s) is
the scattering phase function and describes the probability that a ray from a certain
direction, s
i
, is scattered into a certain other direction, s.
Since radiation intensity propagates at the speed of light, there is no time de-
pendent term in the RTE. The rst term on the right hand side of the RTE is the
augmentation of radiative intensity due to emission. The second term is the attenuation
due to absorption and scattering. And the third term is the augmentation due to scat-
tering of the radiative intensity from other directions. Equation (4.2) is valid to describe
radiative heat transfer in any participating medium. However, the equation can be sim-
plied depending on the specic nature of the participating medium. For example, in
lightly sooting ames, scattering is usually negligible and the RTE becomes,
dI
ds
= s I
(I
b
I
). (4.3)
48
The RTE is a rst-order dierential equation in intensity for a xed direction s
and it is closed by specifying the intensity at the enclosure boundary into direction s.
When the solution of the RTE, I
( s) sdd, (4.4)
and the divergence of this radiative heat ux vector gives the radiative heat source term
in the energy equation (eq. 2.4).
4.3 Solution Methods for the Radiative Transfer Equation
The radiative transfer equation is an integro-dierential equation for spectral
radiative intensity in ve independent variables: three space coordinates and two direc-
tional coordinates with local origin. The problem becomes even more complicated if the
medium is nongray, which introduces the additional spectral variable, and if the geom-
etry of the participating medium enclosure is irregular. Consequently, exact analytical
solutions exist for only a few extremely simple situations: for example, a one-dimensional
plane-parallel gray medium that is either at radiative equilibrium (radiation is the only
mode of heat transfer) or whose temperature eld is known. Even for these simplest
cases, the exact solution can only be written implicitly in the form of an integral equa-
tion [88].
For most combustion systems, the problems are multi-dimensional and nonhomo-
geneous, and the spectral variations of the radiative properties frequently have to be
accounted for. Therefore simplifying approximations are necessary in the solution of the
RTE. There is no single model or approximation that is universal and can be applied
to all dierent types of practical problems. Depending on the nature of the combustion
system, characteristics of the ame, the solution techniques used for other governing
equations, the degree of accuracy required, and the available computer facilities, solu-
tion methods of varying degrees of approximation have been devised. The majority of
the methods that have been used in combustion simulations fall into one of the ve
49
groups: (1) optically thin approximations; (2) spherical harmonic methods; (3) discrete
ordinate methods; (4) zonal methods; and (5) statistical methods. There are also hybrid
methods that combine the features of several dierent methods. Detailed discussions
and references can be found in, e.g. [88, 123]. In the following, a general overview of
these methods is provided.
4.3.1 Optically Thin Approximation
The optically thin approximation has been widely used in combustion simula-
tions. The essence of the optically thin approximation is that the absorption in the
participating medium is neglected. This means that the radiation that is emitted can
travel without interruption to the boundaries of the participating medium. This approx-
imation is appropriate either when the dimensions of the medium are very small or when
the absorption coecient of the medium is very small. The RTE is thus simplied and
only the rst term on the right-hand side of equation (4.2) is kept. The advantage of this
method is its simplicity and computational economy. The disadvantage is that it gen-
erally over-predicts radiative heat loss and fails to account for the interactions between
radiation and ame structure.
4.3.2 Spherical Harmonic Method
The spherical harmonic method, also known as the P
N
method or dierential
method, is based on the idea that the solution of the RTE can be simplied by express-
ing the radiative intensity as a series of products of angular (directional) and spatial
functions. The angular dependence is represented using spherical harmonics, which sat-
isfy Laplaces equation in spherical coordinates, and the spatial functions then are solved.
The solution is facilitated by the orthogonality of spherical harmonics. The number of
terms retained in the series expansion gives the method its order and its name: for
example, the P
1
or the P
3
approximation. It is known from neutron transport theory
that approximations of odd order are more accurate than approximations of the next
higher even order, so that the P
2
approximation is never used. The spherical harmonic
method is mathematically elegant. It transforms the RTE into a set of simultaneous
50
partial dierential equations that are similar in structure to the other pdes that must
be solved for a chemically reacting ow. The drawback of the method is that the ac-
curacy improves slowly for higher-order approximations while mathematical complexity
increases extremely rapidly, and the low-order approximations are only accurate in me-
dia with near-isotropic radiative intensity. Still, the P
1
approximation is used widely in
combustion simulations.
4.3.3 Discrete Ordinate Method
The discrete ordinate method, also known as the S
N
method or ux method, deals
also with the the angular (directional) dependency of radiative intensity. Unlike the P
N
method, the S
N
method discretizes the entire solid angle ( = 4) using a nite number
of ordinate directions and corresponding weight factors. The radiative intensity in each
sub-direction is considered to be uniform and is solved for as a function of position.
Integrals, such as the radiative heat ux and the radiative source term, over solid angle
are then approximated using numerical quadrature. The number of sub-directions, which
is determined by the order of quadrature scheme used, gives the method its order and its
name: for example, the S
2
or the S
4
approximation. Even orders are used because it is
customary to choose sets of directions and weights that are completely symmetric: that
is, sets that are invariant under any rotation of 90
i=1
r
2
i
,
where r
i
is the distance from the center of each particle to the center of gravity of the
soot aggregate. The fractal dimension and prefactor, D
f
and k
f
, appear to be relatively
universal properties of soot aggregates, independent of fuel type and ame condition.
Various soot aggregate measurements have shown that the fractal dimension is in the
range of 1.61.9 and the fractal prefactor 7.09.2 [141, 143]. Suggested average values
are D
f
= 1.82 and k
f
= 8.5 with experimental uncertainty (95% condence) of 0.08 and
0.5, respectively [140].
Based on the mass fractal description of soot aggregates, predictions of soot ra-
diative properties have been made using two kinds of methods: fundamental methods
60
and approximate methods. The fundamental methods usually solve the integral rep-
resentation of Maxwells equations for a dielectric object and involve constructions of
soot aggregates using stochastic simulation and approximations of electromagnetic elds
inside and outside of the aggregates. Examples of the fundamental methods include
Borghese et al.s approach [144], Purcell and Pennypeckers discrete dipole approxima-
tion (DDA) [145], and the Iskander-Chen-Penner (ICP) approach [146]. The ICP model
is considered today to provide the most reliable and accurate predictions. Reviews and
comparisons of various fundamental methods can be found in [147, 142]. The fundamen-
tal methods usually consider a single soot aggregate and are computationally involved.
Therefore, they are often used for the evaluation of approximate methods.
A variety of approximate methods have been proposed to treat soot aggregate
radiative properties. These include Rayleigh scattering, Mie scattering for an equivalent
sphere, and Rayleigh-Debye-Gans (RDG) scattering [142]. The Rayleigh scattering ap-
proximation is valid for the individual primary particle, which is small. When applied
to an aggregate, however, it generally underestimates scattering and extinction. The
Mie calculations assume a single sphere that has the same volume as the aggregate. It
yields unreliable results because a compact sphere is not a good model of the open wispy
structure of a soot aggregate. However, a recent study [143] suggests that the shape
eects on emission predictions can be neglected.
The RDG scattering theory tends to give relatively good results and is considered
to be the most promising approximate method for soot radiative property prediction.
The RDG approximation has gone through several revisions and extensions. The major
assumptions of this approach are as follows: spherical primary particles have constant
diameters and uniform refractive indices; primary particles just touch one another; and
the aggregates are mass-fractal objects. Justications of these assumptions are proved
in [86]. The original version of RDG [148] ignored the eects of multiple- and self-
scattering so that the electric eld of each primary particle is the same as the incident
electric eld. The dierences of the phase shift of scattered light from various points
within a particular primary particle also were ignored. It assumes a complex index of
refraction near unity, which is clearly not very accurate for soot. Jullien and Botet [93]
61
and Martin and Hurd [149] improved the original RDG by applying correction factors
for the phase and multiple scattering eects of real aggregates. Dobbins and Megaridis
[150] applied RDG scattering to polydisperse aggregate populations. K oyl u and Faeth
[151] improved the Dobbins and Megaridis model by considering soot aggregates of large
size and broad size distribution; their model has been evaluated using both experimental
measurements [151] and more exact simulations using the ICP algorithm [152].
4.5 Turbulence-Radiation Interactions
Similar to turbulence-chemistry interactions, turbulence-radiation interactions
(TRIs) arise from the highly nonlinear dependence of radiative transfer on temperature
and species concentrations. Turbulence inuences radiative transfer through uctuations
in temperature and species concentrations, and those in turn inuence the blackbody
intensity and the absorption and scattering coecients of the radiatively participating
medium. At the same time, radiation inuences turbulence through its intimate coupling
with ame structure, such as temperature and density distributions. Because radiative
transfer takes place at the speed of the light, turbulence uctuations are reected in-
stantaneously in radiative heat ux and its divergence.
Over the years, both experimental and theoretical studies have shown the im-
portance of TRIs. Experimental observations from the groups of Faeth and Gore
[153, 154, 155, 156, 157, 158, 159] have indicated that, depending on the fuel, radia-
tive emission from a ame may be as much as 50 to 300% higher than would be expected
based on mean values of temperature and absorption coecient. Coxs [160] theoret-
ical analysis using Taylor series expansion of emissive power has shown that emission
from a hot medium increases dramatically due to turbulence. In addition, idealized
numerical studies [161, 162, 163] have demonstrated the eect of TRIs as well as the
eect of dierent parameters on the overall TRI eect. The importance of TRIs is now
well acknowledged and it has become essential to account for these interactions in a
comprehensive modeling of turbulent ames.
62
TRIs arise mathematically from averaging the radiation source term (equation
4.1) that appears in the energy (enthalpy) equation (equation 2.4):
q
R
=
_
0
_
4I
b
_
4
I
d
_
d =
_
0
_
4
I
b
_
4
d
_
d, (4.5)
where the over-bar indicates the Reynolds-averaging operator. Two correlations result
that require modeling: the correlation between absorption coecient and Planck func-
tion,
I
b
, and the correlation between absorption coecient and radiative intensity,
. The term
I
b
can be constructed from single-point statistics of the temperature
and radiating species concentrations, while the term
, cannot be expressed by local scalar variables. This makes the closure of TRI
terms complicated, and the absorption term in particular.
The analysis and approximations made by Kabashnikov and coworkers [164, 165,
166] simplify the situation. They suggested that if the mean free path for radiation is
much larger than the turbulence eddy length scale, then the local radiative intensity
is only weakly correlated with the local absorption coecient; that is,
.
This expression is commonly known as the optically-thin eddy approximation. While
not valid for the full spectrum of combustion gases and for extremely sooty ames,
this approximation is justiable in many applications [88]. To date, nearly all TRI
calculations have employed the optically-thin eddy approximation.
Even with the optically-thin eddy approximation, the remaining correlations
and
I
b
remain dicult to evaluate. As was the case for the mean chemical reaction
rates, these correlations are nonlinear functions of temperature and radiating species
concentrations. They require models and additional equations for their closure, as do
the mean reaction source terms in turbulent ames. It is natural to employ the same
modeling techniques for the closure of the TRI terms as for turbulence/chemistry inter-
actions (TCIs). However, the traditional RANS-based models have been shown to be
incapable of dealing with TRI terms because of the large number of correlations involved
63
and the large magnitude of the uctuations in turbulent ames [163]. In the recent sim-
ulations of turbulent combustion with radiative heat transfer, TRIs are either neglected
all together or are accounted for using the advanced PDF methods.
PDF-based approaches to TRIs can be divided into two groups, as was the case
for TCIs: presumed PDF methods and transported PDF methods. The presumed PDF
methods provide a general framework in which a number of turbulent reaction and mix-
ing models may be implemented, such as the popular laminar amelet models. Typical
choices of PDF shapes include the beta function and the clipped Gaussian distribu-
tion. The presumed PDF method is the most common approach to TRIs; examples
include Song and Viskanta [167], Gore et al. [168], Hartick et al. [169], and Coelho
et al. [170]. Transported PDF methods have the advantage of treating TRIs rigor-
ously and exactly (only the emission part, the correlation between absorption coecient
and Planck function). Investigations of TRIs using this approach have been conducted
by the research group of Modest [171, 172]. Mazumder and Modest [171] employed a
velocity-composition joint PDF method, and Li and Modest [172] used a composition
PDF method in their simulations of methane-air turbulent diusion ames. Li and Mod-
est [88] also investigated the importance of the various TRI correlations to the total TRI
eect by freezing species concentration and temperature elds. They concluded that
consideration of the temperature self correlation alone is not sucient to capture TRIs,
and the absorption coecient-Planck function correlation also must be included.
TRI calculations to date have focused on nonluminous ames and the eects of
soot radiation have been mostly neglected. Gore and Faeth [155] have suggested that
TRI eects should be strong in luminous ames. In addition, the optically-thin eddy
approximation becomes invalid in sooty ames, and models need to be developed to
account for the correlation between absorption coecient and radiative intensity.
4.6 Summary
This chapter has been dedicated to issues that arise when considering thermal
radiation heat transfer in turbulent reacting ow simulations. Radiation is an important
heat transfer mode in combustion applications due to its fourth power dependence on
64
temperature. It is intimately coupled with the determination of ame structure. Ignor-
ing radiation or treating radiation using over-simplied models may result in large errors
in ame predictions, especially ame temperature, pollutant emissions, and soot forma-
tion. Radiation calculations are coupled to ame simulations through a source term in
the energy (enthalpy) equation. The source term is calculated based on the solution
of the radiative transfer equation. The accuracy of the radiation calculations depends
on the accuracy of the methods employed to solve the radiative transfer equation and
the accuracy of the radiative properties of participating media. Various solution meth-
ods have been discussed. Statistical methods are the most accurate, but are extremely
expensive computationally. Approximation methods, especially the spherical harmonic
P
1
method and the discrete ordinates S
2
or S
4
methods, are widely used in CFD-based
combustion simulations. Various models for both gas-phase and soot radiative proper-
ties have been discussed. For gas-phase species, global models are commonly employed;
the recently developed full-spectrum k-distribution method is considered to be the most
accurate and ecient model for gas-phase radiation calculations that is currently avail-
able. For soot particles, the Rayleigh-Debye-Guns approximation with the consideration
of mass fractal nature of soot aggregates is considered to be the most promising model
for soot radiative properties. In turbulent ames, turbulence-radiation interactions play
an important role in determining radiative heat losses but are dicult to deal with. An
optically-thin eddy approximation has commonly been employed, and transported PDF
methods provide a more rigorous and accurate treatment of TRIs. The eect of soot
radiation on TRIs has not been explored.
65
Chapter 5
Two-Stage Lagrangian Simulation of
Oxygen-Enriched Turbulent Jet Non-Premixed Flames
5.1 Introduction
Numerical modeling of turbulent jet ames remains an active research area; its
origins can be traced back to the 1970s [173, 174]. Most CFD-based models solve
discretized conservation equations coupled with a turbulence closure scheme, such as k-
and k--g models. An alternative formulation is the probability density function (PDF)
method [77], which has been employed widely for its superior treatments of turbulence,
chemistry, and turbulence-chemistry interactions. Until recently, computing limitations
precluded such models from including all but the simplest chemistry. Progress with
approaches such as the in situ adaptive tabulation scheme [48] is allowing the use of
larger chemical mechanisms.
The importance of using detailed chemistry and the diculties associated with it
have been discussed in Chapter 2. One strategy for dealing simultaneously with both
the turbulent ow eld and complex chemistry that involves hundreds of species is to
simplify the description of the ow eld physics (Section 2.3). A successful example of
this strategy is the two-stage Lagrangian (TSL) jet-ame model of Broadwell and Lutz
[15]. The TSL model permits the use of arbitrarily complex chemistry while capturing
the main hydrodynamic features of turbulent jet ames. It has proven to be a useful
and computationally ecient model of turbulent nonpremixed jet ames [13].
However, the original TSL model lacks the capability to predict soot and employs
an overly simplistic radiation model. In this chapter, the functionality and utility of
the TSL model are extended by incorporating a detailed soot model and by developing
a sophisticated radiation model that considers the emission and absorption from both
66
gas-phase species and soot particles. The improved model is utilized to simulate oxygen-
enriched turbulent jet ames.
5.2 The Two-Stage Lagrangian Model
The Two-Stage Lagrangian (TSL) model represents turbulent mixing as a two-
stage process that is viewed in a coordinate frame that moves downstream with the
average uid motion [15]. The two-stage approach is based on experimental observations:
reactions occur rst in the diusion layers, and continue in nearly homogeneous regions
that are created by turbulent mixing [13]. The homogeneous regions are represented
by one perfectly stirred reactor. The diusion layer can be modeled either as a one-
dimensional strained diusion ame, or more simply, as another perfectly stirred reactor.
The choice leads to the two versions of the model: the two-reactor version, and the
reactor/diusion-ame version. The TSL model is shown schematically in Fig. 5.1.
The TSL model is simply an extension of the stirred reactor concept [9] to repre-
sent a turbulent reacting jet. In a steady jet, the core reactor starts near the jet nozzle
with pure fuel and moves downstream with the average velocity, receiving air by entrain-
ment. It represents the average time history of the uid. The uid mechanics describing
the growth of the jet and ow into the two reactors is controlled by empirical entrainment
correlations. A numerical integration is performed following the jet ame gases down-
stream in time in a Lagrangian manner. The average compositions and temperature are
computed during the integration.
A key feature of the model is the simple physically based representation of turbu-
lent mixing. The TSL model represents the mixing rate and momentum transport using
empirical correlations. This reduction of ow structure to its essence allows the use of
arbitrarily large chemical mechanisms. The ability of this model to predict NO, NO
2
,
and CO emission trends, along with radiant fraction trends [15], suggests that enough
of the underlying ow physics is captured by the model to provide useful guidance in
understanding complex-chemistry issues associated with jet ames.
It has been found [13] that the two-reactor version of the TSL model yielded
essentially the same results as the reactor/diusion-ame version, but at greatly reduced
67
Fig. 5.1. TSL two-reactor and reactor/diusion-ame models.
68
computational cost. Therefore, the two-reactor model is preferred for the majority of
applications. In this study, the two-reactor version of the model is then employed.
5.3 Detailed Chemical Kinetics
A prominent feature of the TSL model is its ability to use arbitrarily complex
chemistry with great computational eciency. To take full advantage of this feature,
a detailed comprehensive chemical mechanism is employed, including a state-of-the-art
soot mechanism.
This complex mechanism is a compilation of GRI-Mech 3.0 [22] and the mecha-
nism for PAH growth and oxidation from Appel et al. [107]. The combined mechanism
contains 122 chemical species and 677 elementary reactions. The incorporation of the two
mechanisms allows detailed modeling of NO
x
emissions and soot production in natural
gas ames.
The PAH mechanism of Appel et al. (ABF2000 model) [107] is an updated
version of the mechanism of Wang and Frenklach (WF1997 model) [175] based on new
developments in gas-phase reactions and aromatic chemistry. The WF1997 model is
based on GRI-Mech 1.2 and a consistent set of rate coecients, thermodynamic data,
and transport data for reactions of aromatics developed by Wang and Frenklach. It
contains 99 species and 527 reactions. This reaction mechanism can be used for oxidation
of methane, ethane, ethylene and acetylene at ame temperatures. Aromatic chemistry
is included up to the formation of pyrene. The mechanism has been tested against
a number of laminar premixed ames of acetylene and ethylene. It predicts well the
major, minor, and small-aromatic species, but under-predicts two-, three-, and four-
ring aromatics. The ABF2000 model improves on this part, and predicts well the major,
minor, and aromatic species up to pyrene in laminar premixed ames of ethane, ethylene
and acetylene fuels. However, it is based on GRI-Mech 1.2 as well.
GRI-Mech is an optimized detailed chemical reaction mechanism that represents
current best understanding of natural gas combustion, including NO
x
formation and
re-burn chemistry. The latest release, GRI-Mech 3.0, is a compilation of 325 elementary
reactions and associated rate coecient expressions and thermo-chemical parameters for
69
53 chemical species. It diers from the previous releases in that kinetics and target data
have been updated, improved, and expanded. Propane and C
2
oxidation species have
been added, and new formaldehyde (CH
2
O) and NO formation and re-burn chemistry
included. The relations between GRI-Mech 3.0 and its predecessor are shown in Fig.
5.2.
GRI-Mech has been optimized for methane and natural gas as fuels. It is suggested
that GRI-Mech should not be used to model combustion of pure fuels such as methanol,
propane, ethylene, and acetylene even though these compounds are included in the GRI-
Mech species list. The WF1997 and ABF2000 model are based on GRI-Mech 1.2; that
is, the H/O, C
1
H
x
, and C
2
H
x
mechanisms are taken from GRI-Mech 1.2. However,
the reaction coecients from GRI-Mech 1.2 were updated to better describe ethane,
ethylene, and acetylene ames. These modications are retained during the integration
of GRI-Mech 3.0 and the soot mechanism of ABF2000 model. The resultant complex
mechanism is expected to be a good model for ethane- and propane-fueled ames. It is
given in Appendix A.
5.4 Soot Calculation
The soot model employed is a detailed description of the dierent physical and
chemical processes in the formation and oxidation of soot particles [107, 175]. The
processes include formation and oxidation of soot precursor species, nucleation of the
rst soot particles, coagulation, surface growth, and oxidation of soot particles.
The gas-phase chemistry for soot precursor species (PAHs) includes up to the for-
mation of four-ring aromatics (pyrene) and is described by the detailed chemical mech-
anism containing 122 species and 677 reactions described in section 5.3. The occurrence
of the smallest soot particles results from the coagulation of two PAHs (mainly pyrene):
A4 + A4 rst soot particle,
where A4 denotes pyrene. These particles grow further in mass through surface reactions
and condensation of PAH molecules. The heterogeneous surface reactions, including
70
GRI-Mech 1.2
325 elementary reactions
with 53 species
277 elementary reactions
with 49 species
175 elementary reactions
with 32 species
NO chemistry
propane chemistry
GRI-Mech 2.11
GRI-Mech 3.0
Addition
Addition
U
p
d
a
t
e
&
I
m
p
r
o
v
e
Fig. 5.2. The evolution of GRI-Mech.
71
particle oxidation, are treated using a detailed reaction mechanism based on a chemical
analogy to gas-phase aromatic chemistry, mainly the HACA mechanism (section 3.2.1).
Oxidation of soot occurs as a result of surface reactions with oxygen molecules and the
OH radicals. The oxidation rate by O
2
uses the expression by Lin and Lin [97] and the
oxidation rate by OH is based on the data of Neoh et al. [98].
The size distribution of soot particles is sought from this detailed model. The
evolution of soot size distribution, that is the soot dynamics, due to nucleation, coagula-
tion, and surface growth and oxidation is described by the Smoluchowski master equation
using the method of moments (see section 3.4 for details). The soot size distribution is
then determined in terms of soot moments. The agglomeration of soot particles into
aggregates is not considered here.
The governing equations for soot moment transport have been derived according
to the specic conguration of the TSL model. The derivation is similar to that of the
species conservation equations. Instead of the mass of a chemical species, the conserva-
tion of the soot particle number density of a certain size is considered [92]. A formal
derivation of soot moment transport equations is given in the next chapter. Here the
soot moment equations for the TSL two-reactor model are given, neglecting diusion of
soot particles (since it is small):
Core (Homogeneous) Reactor
dM
r,h
dx
=
1 +B
m
h
d m
h
dx
_
M
r,f
M
r,h
_
+
Q
r,h
h
u
, (5.1)
Flame Sheet Reactor
1
__
M
r,
M
r,f
_
+B
_
M
r,h
M
r,f
__
+
Q
r,f
f
= 0 for r = 0, 1, . . . , 5, (5.2)
M
r
=
M
r
Q
r
=
dM
r
dt
.
72
Here m is the mass ow rate, u the axial velocity, the residence time for ame-sheet
reactor, and B is the ratio of the mass ow from the core reactor to the ame-sheet reactor
to the entrainment ow from the surroundings. The subscript h denotes the homogeneous
core reactor, f the ame-sheet reactor, and the surroundings. The source terms
dM
r
/dt are the time-rates-of-change of moments due to nucleation, coagulation, surface
growth, and oxidation (section 3.4).
The term M
r
is used to make the moment equations identical to species con-
servation equations. Compared to species mass fractions, soot moments are very large
numbers. Therefore, the moment equations are rewritten as,
dY
r,h
dx
=
1 +B
m
h
d m
h
dx
_
exp(Y
r,f
)
exp(Y
r,h
)
1.0
_
+
Q
r,h
exp(Y
r,h
)
h
u
, (5.3)
1
__
exp(Y
r,
)
exp(Y
r,f
)
1.0
_
+B
_
exp(Y
r,h
)
exp(Y
r,f
)
1.0
__
+
Q
r,f
exp(Y
r,f
)
f
= 0 for r = 0, 1, . . . , 5,
(5.4)
where,
Y
r
= ln(M
r
).
The solution of the moment equations is then facilitated by adding these equations to the
species equation set in the TSL code. In the calculation, the production rates of those
species that participate in soot formation are corrected for the gas-phase calculation. The
eects of soot particle formation on the energy equation and mixture density are ignored
since the total mass of soot is relatively small. The results of the soot moments are
further processed to obtain soot information of interest, such as equivalent soot volume
fraction, averaged soot diameter, soot mass ow rate, etc.
5.5 Radiation Calculation
In the original TSL model, radiation is treated by assuming that the jet ame
emits energy from its surface to the cold surroundings. Therefore, the loss term in the
energy equation, which has dimensions of energy per unit time per unit mass in the
73
reactor, takes the form of [13]:
q
R
= 4
T
4
T
4
d
,
where is the Stephan-Boltzmann constant, the gray-gas emissivity, T and T
the
reactor and surrounding temperature, respectively, the reactor mixture density, and d
the local ame diameter.
This description of ame radiation is overly simplistic. The gray-gas emissivity,
, is taken to be a constant in the code and its value is adjusted manually to match the
measured radiant fraction, the ratio of radiative heat loss to the total chemical energy
release. This treatment excludes the possibility of considering the contributions of gas-
phase and soot radiation. The above expression for radiation loss applies to the core
homogeneous reactor. When the ame-sheet reactor temperature and mass density are
substituted, it is employed to describe the radiation in the ame-sheet reactor and causes
confusion in the physical interpretation.
Two radiation models have been developed subsequently to improve the radiation
calculations of the TSL models. Both models are based on the solution of the radiative
transfer equation (section 4.2). The rst model employs the optically-thin approximation
(section 4.3). The mathematical formulations are described in a technical report by Wang
et al. [176] and are given in Appendix B. The second model improves on the rst one by
considering absorption eects and solves the RTE (equation 4.3) for a one-dimensional
medium. Separate models were developed for the core reactor and the ame sheet. In
the core reactor, the radiative heat ux is determined by applying the P1 (dierential)
approximation to equation 4.3. In the ame-sheet reactor, a discrete ordinates method
(S2) is implemented. The S2 method has the property of always going to the correct
optically-thin limit [88], a desired attribute for modeling the physically thin ame-sheet
reactor. The detailed development of this second model is described in a technical report
by Wang et al. [177] and is given in Appendix B. An apparent drawback of the second
model is that the radiative heat exchange in the axial direction is neglected because of a
inherent feature of the TSL model: the solution of the TSL model marches downstream
74
from the jet nozzle and hence the temperature eld downstream of the jet ame is not
available for radiation calculations during TSL integration.
In both reactors, gray media characteristics are assumed, employing Planck-mean
absorption coecients to determine both gas-band and soot radiative properties. Curve-
ts [10] are used for the pressure-based Planck mean absorption coecients for CO
2
,
H
2
O, CO, and CH
4
, while the soot Planck-mean absorption coecients are calculated
from Rayleighs theory by assuming that soot are small spherical particles. The details
of the evaluation of Planck-mean absorption coecients are described in Appendix C.
5.6 Results and Discussions
With the improvement in radiation model and the incorporation of a detailed
soot model, the modied TSL code is used to simulate a set of oxygen-enriched turbu-
lent jet nonpremixed ame experiments conducted in our laboratory. The jet ame rig
used in the experiments was designed to produce a vertical jet ame in a nearly quies-
cent air/oxygen coow [12]. The burner consists of a 3-mm i.d. fuel tube centered in a
200-mm i.d. stainless-steel ame chamber. Before entering the chamber, the air/oxygen
oxidizer passes through a 25-mm thick glass bead bed and a 100-mm long ceramic hon-
eycomb to produce a uniform, laminar coow. Tests were run to determine the eects
of fuel type, fuel-jet momentum ux (or velocity), and oxygen index on soot, radia-
tion, and NO/NO
x
in turbulent jet ames. Fuel types used were propane, natural gas
(approximate composition: 94% methane, 4% ethane, 0.7% nitrogen), and a 90%/10%
methane/ethene blend. The momentum uxes of the jets were matched for comparison
purposes as indicated in Table 5.1. Also shown in the table are jet exit velocity and jet
Reynolds number based on cold fuel properties. Oxygen index is dened as the volume
fraction (percentage) of oxygen in the oxidizer and the values used were 21 (air), 30, 40,
55, 75, 90, and 100% (pure oxygen).
Laser extinction measurements were made along the centerline of the ame cham-
ber using a 632.8-nm intensity-stabilized helium-neon laser to determine equivalent soot
volume fraction as dened by Kent and Bastin [178]. Equivalent soot volume fraction is
the soot volume fraction over a path length equal to the nozzle diameter that produces
75
Table 5.1. Test Conditions of Oxygen-Enriched Flames
Fuel Momentum ux (N) Velocity (m/s) Jet Reynolds No.
Propane 6.95 10
4
7.28 4,940
Propane 6.25 10
3
21.8 14,810
Natural gas 6.25 10
3
35.6 6,646
Natural gas 3.12 10
2
79.6 14,802
Blend 6.25 10
2
35.4 6,697
Blend 3.12 10
2
79.1 14,992
the same extinction as the actual prole. Using Rayleigh theory, the equivalent soot
volume fraction is calculated by,
F
v
=
6d Im
_
m
2
1
m
2
+2
_ ln
_
I
I
0
_
, (5.5)
where m is the complex index-of-refraction (here m = 1.94 0.54i was used [12] and it
is close to the value used by Lee and Tian [137]), is the wavelength, d is the fuel nozzle
diameter, and I and I
0
are the measured and initial laser intensities.
Radiation measurements were made using a wideband radiometer with a zinc
sulde window that allows transmission at the 12 m wavelengths of water bands. Ra-
diative heat ux measurements were made at 100-mm intervals along the length of the
ame chamber. The measurements were corrected for blackbody emission from the duct,
calculated from the measured duct temperature, and then integrated to determine the
total radiative heat loss from the ame. Radiant fractions are calculated by dividing the
total radiative heat loss of the ame by the chemical heat release. From test compar-
isons, it is estimated that the conned ames in the experiments have radiant fractions
5 to 25% higher than free ames.
A chemiluminescent analyzer was used to measure NO/NO
x
emissions. CO
2
emissions were measured using an infrared analyzer, allowing for the calculation of NO
x
76
emission indices, dened as the mass of pollutant formed per mass of fuel burned, using
the relationships in [179]. As is customary, the indices were calculated assuming the NO
ultimately becomes NO
2
when converting molar quantities to mass quantities. The gas
samples were taken using a quartz probe at the exit plane of the ame chamber.
5.6.1 Performence of Radiation Sub-Models
The radiation models are evaluated by running the TSL code for a propane tur-
bulent diustion ame at an initial velocity of 21.8 m/s and with variations in oxygen
index. The results are shown in Fig. 5.3.
These preliminary evaluations show that the second radiation model, which con-
siders absorption eects, performs better than the rst optically-thin model, which in
turn performs better than the original radiaion model. In Fig. 5.3, the radiant fraction
curve predicted by the second absorption model follows the trend shown in the exper-
iments, as does the curve from the rst optically-thin model. However, the distance
between the predicted and experimental curves diminishes when using the absorption
model. This is an expected result since absorption within the ame decreases the radi-
ation loss to the surroundings.
5.6.2 NO
x
Emissions
Figure 5.4 shows the variation of NO
x
emission indices with the oxygen index
for four dierent ames, from both the experiments (curves with symbols) and the TSL
simulations (curves without symbols). For all ames, the experimental NO
x
emission
levels increase signicantly with the initial oxygen enhancement, reaching maximum
values at oxygen indices of 75%, or less. Increasing NO emission with oxygen index is
consistent with increased temperatures (see Table 5.2), promoting NO formation through
the Zeldovich mechanism [9] in two ways: the temperature increases the value of the
rate constant and promotes a greater pool of O-atoms. Furthermore, oxygen enrichment
increases the overall oxygen availability to form O-atoms. Beyond an oxygen index of
approximately 70%, NO
x
emissions decrease as a result of the unavailability of nitrogen
and rapidly drop to zero as the oxygen index approaches 100%. Although absolute values
77
Oxygen Index (%)
R
a
d
i
a
n
t
F
r
a
c
t
i
o
n
0 20 40 60 80 100
0.1
0.2
0.3
0.4
0.5
Experiments
Original Model
Improved Model 1
Improved Model 2
Propane Jet Flames, v = 21.8 m/s, d = 3 mm
Fig. 5.3. The performance of radiation models.
78
are over-predicted, this trend of NO
x
emission variation is captured by the TSL code as
shown. To assess the role of super-equilibrium O-atom concentration in oxygen-enhanced
combustion, equilibrium calculations were performed simultaneously with the kinetic
calculations. For the propane ames with air (21% O
2
) and with an initial jet velocity
of 21.8 m/s, peak kinetic O-atom concentrations exceed equilibrium peak concentrations
by a factor of approximately three; however, kinetic O-atom concentrations approach
the equilibrium values with increasing oxygen index and the two are nearly identical at
an index of 40%.
Table 5.2. Calculated Peak Temperatures by TSL
Temperature (K)
Propane, Propane,
O
2
index(%) Blend Natural gas v
0
= 21.8 m/s v
0
= 7.28 m/s
21 2129 2108 2124 2111
30 2513 2494 2513 2503
40 2712 2695 2719 2712
55 2868 2851 2882 2878
75 2981 2967 3003 3000
90 3037 3022 3063 3059
100 3064 3052 3093 3090
We also note that the NO
x
emission indices for the methane-dominated ames are
somewhat higher than those for the propane ames. When mass-based emission indices
are converted to an energy basis, i.e., mass of NO
x
per unit of energy released, the eect
of fuel type is diminished. For example, the TSL predictions at an oxygen index of 75%
are 5.36 10
3
gNO
x
/kJ for the natural gas ame and 5.43 10
3
gNO
x
/kJ for the high
velocity propane ames, nearly identical results.
79
+
+
+
+
+
+
Oxygen Index (%)
E
I
N
O
x
(
g
/
k
g
f
u
e
l
)
0 20 40 60 80 100
0
50
100
150
200
250
300
350
TSL Model
Blend
Natural Gas
Propane(21.8 m/s)
Propane(7.28 m/s)
Blend
Natural Gas
Propane(21.8 m/s)
Propane(7.28 m/s)
Experiments
+
Fig. 5.4. NO
x
emission indices from TSL model and experiments: fuel, velocity, and
oxygen index eects.
80
The general trend that NO
x
emission indices are greater for the high-velocity
propane ame compared to the low-velocity case is consistent with previous studies [180]
in our laboratories. This trend is also consistent with the TSL predictions, where the
high-velocity ames have slightly higher peak temperatures. Comparing the predicted
axial temperature proles (Fig. 5.5), we see that the high- and low-velocity ames are
clearly not self-similar, so temperature-time histories appear to be signicantly dierent
for the two conditions. Buoyancy eects are likely to be more prominent in the low-
velocity ames. Estimated ame Froude numbers [9] are 0.163 and 0.488 for the low-
and high-velocity ames, respectively, for combustion in air.
5.6.3 Soot Volume Fractions
Figure 5.6 shows the axial proles of equivalent soot volume fraction for propane
ames with an initial velocity of 21.8 m/s and with varying oxygen indices. These
proles represent the general shapes produced by the other ames measured in the
experiments. Two apparent trends can be readily seen. First is that the location of the
peak soot volume fraction moves towards the jet nozzle as the oxygen index is increased,
a demonstration that the addition of oxygen shortens the ame length. The other trend
is that the total amount of soot, roughly proportional to the area under the prole,
increases with initial oxygen enhancement, peaks at about 30 to 40% oxygen index, and
then decreases with further oxygen addition. The initial increase can be attributed to
an increase in ame temperature, which promotes the fuel pyrolysis and soot formation.
The decrease likely results from the combined condition of high temperature and high
oxygen concentration, which promotes soot oxidation. Also contributing is the reduced
residence time resulting from the shortened ame length at high oxygen index.
The predicted axial proles of equivalent soot volume fraction for the same ames
as above are shown in Fig. 5.7. Two dierences from the experiments are noted readily:
the narrower sooting region, and the much lower magnitude of total soot amount. These
can be explained by the inherent simplications embodied in the TSL model and by
the uncertainties associated with soot modeling. The TSL model treats the jet ame as
a homogeneous core reactor surrounded by a ame-sheet reactor. As the core reactor
81
Height (cm)
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 10 20 30 40 50
0
500
1000
1500
2000
2500
3000
v
0
= 7.28 m/s
v
0
= 21.8 m/s
Core Reactor
Flame-Sheet Reactor
Fig. 5.5. Temperature proles from TSL model for propane ames with 40% oxygen
index
82
+
+
+
+
+
+
+
+
+
+
X
X
X
X
X
X
X X
Height (cm)
E
q
u
i
v
a
l
e
n
t
S
o
o
t
V
o
l
u
m
e
F
r
a
c
t
i
o
n
0 20 40 60 80
0.0E+00
5.0E-06
1.0E-05
1.5E-05
2.0E-05
2.5E-05
3.0E-05
40%
30%
55%
75%
90%
100%
21%
Fig. 5.6. Axial prole of equivalent soot volume fraction from experiments for propane
ames with v
0
= 21.8 m/s.
83
convects downstream, it grows as uid enters after being processed by the ame-sheet
reactor. Conceptually, the favorable conditions for soot production, such as proper
equivalence ratio and temperature, can be expected to exist only in a narrow region, while
in a real turbulent jet diusion ame, favorable soot production conditions can be found
in various locations throughout the ame due to turbulent mixing and transport. Also,
our current approach of modeling the ame sheet as a well-stirred reactor is seriously
decient as it predicts no soot formation, while in real ames the ame sheet is a major
source of soot. In spite of these important dierences, the sooting trends (peak equivalent
soot volume fraction location and magnitude) shown in Fig. 5.6 are captured by the TSL
code.
We now discuss the sooting trends associated with fuel type and fuel jet velocity.
To facilitate comparison and discussion, the sooting level of each ame is represented
by its peak equivalent soot volume fraction since the axial proles of the soot volume
fraction assume the same general shape for all the ames in the study. The values of
the peak soot volume fractions are further normalized using the values from the propane
ame at the high-velocity and 40%-oxygen-index conditions for both experimental data
and TSL predictions. These normalized data are plotted in Figs. 5.8 and 5.9.
Figure 5.8 shows the sooting trends as functions of oxygen index for various fuels.
Curves with symbols are from the experiments. As can be seen from the gure, soot
quantities for all the ames increase with initial oxygen enhancement and then decrease
as oxygen content is increased further. The highest soot values occur in the range of 30
to 40% oxygen index. As for fuel-type eect, the propane ame produces much more
soot than the methane/ethene blend ame, which produces slightly more soot than the
natural gas ame. We can see from this gure that the experimental trends are predicted
remarkably well by the TSL model.
Figure 5.9 shows the sooting trends for two initial jet velocities as functions of
oxygen index for propane ames. The low-momentum ame clearly generates more soot
than the high-momentum ame. This likely results from, rst, the longer residence
time in the low-momentum ame (Fig. 5.12), a critical variable in soot production
[181, 182, 183], and second, the higher rates of entrainment and fuel-oxidizer mixing in
84
Height (cm)
E
q
u
i
v
a
l
e
n
t
S
o
o
t
V
o
l
u
m
e
F
r
a
c
t
i
o
n
0 10 20 30
0.0E+00
5.0E-08
1.0E-07
1.5E-07
2.0E-07
2.5E-07
3.0E-07
21%
30%
40%
55%
75%
90%
100%
Fig. 5.7. Axial prole of equivalent soot volume fraction from TSL model for propane
ames with v
0
= 21.8 m/s.
85
Oxygen Index (%)
N
o
r
m
a
l
i
z
e
d
P
e
a
k
F
v
*
0 20 40 60 80 100
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Propane
Blend
Natural Gas
Fig. 5.8. Normalized peak equivalent soot volume fraction from experiments (curves
with symbols) and TSL model: fuel and oxygen index eects.
86
Oxygen Index (%)
N
o
r
m
a
l
i
z
e
d
P
e
a
k
F
v
*
0 20 40 60 80 100
0
0.2
0.4
0.6
0.8
1
1.2
1.4
v
0
= 7.28 m/s
v
0
= 21.8 m/s
Fig. 5.9. Normalized peak equivalent soot volume fraction from experiments (curves
with symbols) and TSL model: fuel jet velocity and oxygen index eects.
87
the high-momentum ame, which together with the higher temperature promote soot
oxidation. Demonstrated in this gure is the ability of the model to capture the sooting
trend with jet velocity and oxygen index. For the experimental results, peak soot values
increase moderately and plateau with oxygen index in the case of low-velocity ames,
while for the high-velocity ames, the peak soot values increase sharply with initial
oxygen addition and then decrease steadily. This variation in trend with oxygen index,
caused by a change in velocity, is captured quite well by the model.
5.6.4 Flame Radiation
Flame radiant fractions as functions of oxygen index are shown in Figure 5.10 and
5.11, which collect the eects of fuel type, fuel velocity, and oxygen index in one graph
each for the experiments (Fig. 5.10) and for the TSL model (Fig. 5.11). Generally,
propane ames have larger radiant fractions than methane-dominant ames; increasing
the velocity in the propane ame decreases the radiant fraction. In all of the ames, radi-
ant fraction increases monotonically with oxygen index and the rate of increase decreases
at higher oxygen indices. By comparing the predictions (Fig. 5.11) and experimental
results (Fig. 5.10), we see that the trends are predicted remarkably well. Note, however,
that soot is under-predicted in the calculation and thus makes a negligible contribution
to the radiation loss. To obtain a more realistic prediction of soot contribution, soot
volume fractions calculated from the TSL code were increased to match experimental
soot volume fractions before calculating the radiant loss. These articially enhanced
values were obtained by multiplying the original calculated soot volume fractions by an
enhancement factor that reects the average ratio of measured to predicted soot volume
fractions for each value of oxygen index. The results of these calculations for propane
ames at low velocity (denoted soot up in Fig. 5.11) thus reect a more realistic con-
tribution from the soot radiation. For these ames, enhancement factors ranged from 62
to 101.
Since the TSL model successfully duplicated the trends shown in the experiments,
we can analyze these trends with the aid of TSL code. Three factors are particularly
important in determining radiant fractions: the temperature through a fourth-power
88
Oxygen Index (%)
R
a
d
i
a
n
t
F
r
a
c
t
i
o
n
0 20 40 60 80 100
0
0.1
0.2
0.3
0.4
0.5
Propane (7.28 m/s)
Propane (21.8 m/s)
Natural Gas
Blend
Fig. 5.10. Radiant fractions from experiments: fuel, velocity, and oxygen index eects.
89
Oxygen Index (%)
R
a
d
i
a
n
t
F
r
a
c
t
i
o
n
0 20 40 60 80 100
0
0.1
0.2
0.3
0.4
0.5
Natural Gas
Blend
Propane (21.8 m/s)
Propane (7.28 m/s)
Propane (7.28 m/s soot up)
Fig. 5.11. Radiant fractions from TSL model: fuel, velocity, and oxygen index eects.
90
inuence on radiant loss, the absorption coecients of the radiating gases and particles,
and the time available for heat to be lost from the ame. The calculated temperatures
for all the ames are tabulated in Table 5.2, while global residence times were calculated
from the TSL code as follows:
=
_
L
0
1
u
dx (5.6)
where u is the local jet velocity and L is the ame length. Figure 5.12 shows residence
times where the ame length is dened as the distance from the nozzle exit to the
location of the peak temperature. From Table 5.2 and, we see the trend of increasing
temperature (T
4
) with oxygen index, which promotes larger radiant fractions; residence
times, however, decrease with oxygen index (Fig. 5.12), which has the opposite eect
on radiant fractions. Since the peak temperatures are quite similar at the same oxygen
indices for all of the ames, the ordering of radiant fractions shown in Fig. 5.10 and 5.11
is most likely explained by the ordering of the residence times shown in Fig. 5.12.
Other factors aecting radiant fraction that are linked to the oxygen index are the
CO
2
and H
2
O mole fractions, both of which increase with oxygen index, and the soot
volume fraction, which shows a more complicated dependence on oxygen index (see Figs.
5.6 and 5.7). To better understand this particular issue, the individual contributions of
gas band radiation and soot radiation were analyzed. Figure 5.13 contrasts the calculated
contributions from gas-band radiation and from soot radiation for the propane ame with
v
0
= 21.8 m/s and the natural gas ame with v
0
= 35.6 m/s, both of which operated
at nearly the same Reynolds number and momentum ux. The soot volume fractions in
the TSL computation were increased by a constant factor such that peak values matched
those measured in the experiments. The gure shows that the soot radiation has a strong
contribution to heat loss in the propane ame, while the soot contribution is small for the
weakly-sooting natural gas ame. Also shown in the gure are the temperature proles
for these two ames. We note that the ame length of the natural gas ame is shorter
than the propane ame and that the peak radiative heat uxes from gas phase are not
at the locations where peak temperatures occur.
Figure 5.14 shows the calculated contribution of soot radiation to the total radia-
tion loss as function of oxygen index for high- and low-velocity propane ames. In both
91
Oxygen Index (%)
R
e
s
i
d
e
n
c
e
T
i
m
e
(
s
)
0 20 40 60 80 100
0
0.01
0.02
0.03
0.04
0.05
Propane (7.28 m/s)
Propane (21.8 m/s)
Blend
Natural Gas
Fig. 5.12. Calculated global residence times as functions of oxygen index for dierent
fuels.
92
Height (cm)
R
a
d
i
a
t
i
o
n
H
e
a
t
F
l
u
x
(
W
/
c
m
2
)
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 10 20 30 40
0
5
10
15
20
25
0
500
1000
1500
2000
2500
3000
Temperature
Heat Flux (soot)
Heat Flux (gas)
Propane
NG
NG
Fig. 5.13. Comparisons between calculated soot and gas-phase radiation heat ux for
propane and natural gas ames with v
0
= 21.8 m/s, 40% oxygen index, and soot volume
fractions increased to match experiments.
93
cases, the fraction of heat loss coming from soot radiation reaches a maximum value of
nearly 15% at an oxygen index of about 30-35%. Beyond the maximum contribution,
the low-velocity ame has a substantially larger soot contribution than the high-velocity
ame. With pure oxygen, only small contributions ( 1% and 3%) to the total radiation
are attributed to soot.
5.6.5 Articially Enhanced In-Flame Soot
Recent advancements in oxy-fuel burner design [7] have proven it possible to in-
crease soot formation beyond the levels generated in ames of more conventional burners.
A numerical exercise was therefore conducted to assess the potential of soot enhancement
in boosting radiant heat transfer and reducing NO
x
emissions. In these calculations, pre-
dicted soot volume fractions are multiplied by factors such that the calculated peak soot
volume fractions are 0.5, 1, 1.5, 2, 3, and 5 times the experimental value associated with
the particular ame condition. Results of this exercise are shown in Fig. 5.15 for a
propane ame with an oxygen index of 30%. The 30%-oxygen condition was selected
as this condition results in the maximum soot contribution to the radiant fraction (see
Fig. 5.14). In Fig. 5.15, we see that the ve-fold increase in soot results in a substantial
reduction in NO
x
emission indices. For the same soot enhancement (5), temperatures
fall approximately 200 K while the radiant fraction increases from about 0.32 to 0.40,
a boost of approximately 25%. Although specic to a particular test condition, results
suggest that soot enhancement may be an eective tool in an overall strategy for lower-
ing NO
x
and maximizing heat transfer, while the modeling exercise illustrates how the
TSL model can be used to estimate the inuence of combustion modications that alter
in-ame soot concentrations.
5.7 Summary
The eects of oxygen index, jet velocity, and fuel type on soot, radiation, and
emission characteristics of oxygen-enriched turbulent jet nonpremixed ames have been
investigated numerically. The numerical studies demonstrate the usefulness of the two-
stage Lagrangian model, with the improved radiation model and the detailed soot model,
94
Oxygen Index (%)
R
a
t
i
o
o
f
S
o
o
t
t
o
T
o
t
a
l
R
a
d
i
a
t
i
o
n
0 20 40 60 80 100
0
0.04
0.08
0.12
0.16
v
0
= 7.28 m/s
v
0
= 21.8 m/s
Fig. 5.14. Calculated soot contribution to total radiation for propane ames with peak
soot volume fractions increased to match experiments.
95
Soot Enhancement Factor
E
I
N
O
x
(
g
/
k
g
f
u
e
l
)
o
r
R
a
d
i
a
n
t
F
r
a
c
t
i
o
n
(
%
)
P
e
a
k
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 1 2 3 4 5
0
10
20
30
40
50
60
2000
2400
2800
3200
3600
4000
Radiant Fraction
EINO
x
Peak Temperature
Fig. 5.15. Eects of articially enhanced soot on NO
x
emission and radiation for a
propane ame with 30% oxygen index and v
0
= 21.8 m/s.
96
to predict general trends for turbulent jet nonpremixed ames. The following are the
conclusions that can be drawn from the numerical investigation:
The modied TSL code is capable of predicting the general trends of pollutant
emission, soot formation, and radiative heat transfer in turbulent jet nonpremixed
ames as functions of oxygen index, fuel type, and initial jet velocity.
The modied TSL code can help to understand and explain results obtained from
experiments.
The modied TSL code can provide information that cannot be measured in the
experiments.
The modied TSL code can provide guidance for conducting further experiments.
The modied TSL code fails to provide quantitative predictions of ame structure
due to its simplistic treatment of the hydrodynamics of ow eld.
97
Chapter 6
CFD Modeling of Oxygen-Enriched
Turbulent Jet Non-premixed Flames
6.1 Introduction
In the previous chapter, the TSL model was shown to perform well in predictions
of general trends in turbulent jet ames. However, it fails to provide quantitatively
accurate predictions: for example, it over-predicts NO
x
emissions (Fig. 5.4) and under-
predicts soot (Fig. 5.7). The main reason for the deciency of the TSL model is its
phenomenological approach to modeling the turbulent ow eld. The TSL model repre-
sents a jet ame as a series of perfectly-stirred reactors moving downstream. Compared
to a real ame, the perfect mixing in those reactors leads to faster chemical reactions and
therefore higher temperatures, which promote ever faster reactions. As a result, each
reactor rapidly approaches chemical equilibrium and we see shorter ame lengths and
over-predicted NO
x
emissions compared to experimental measurements. In addition,
turbulent mixing in the TSL model is controlled by empirical entrainment correlations.
The uncertainties of the correlations not only compound the diculties in correctly
capturing ame structures, but also compromise the applicability of the TSL model to
arbitrary jet ames.
To achieve more rigorous descriptions of the hydrodynamics and turbulence/chemistry
interactions in turbulent ames, and for more accurate and productive numerical pre-
dictions, a CFD-based modeling approached has to be employed. Challenges in this
approach include incorporation of detailed chemistry, soot formation and oxidation, and
radiative heat transfer. These have been discussed in Chapters 2, 3, and 4. In this chap-
ter, an engineering CFD code is modied and enhanced with state-of-the-art models in
detailed chemistry, soot, and radiation calculations. The resulting CFD model then is
exercised to generate new insight into oxygen-enriched turbulent jet ames
98
6.2 The CFD model of ow elds
The underlying ow eld solver, GMTEC, is an engineering computational uid
dynamics code. It solves the conservation equations of mass, momentum, energy or
enthalpy, and chemical species for a compressible multiple-species ideal gas mixture in
a global Cartesian inertial coordinate system. The four independent variable are three
position coordinates x
i
(i = 1, 2, 3) and time t. The principlal dependent variables, for
each of which a pde is solved, are thermodynamic pressure p, three velocity components
u
i
(i = 1, 2, 3), mass-specic internal energy e or mass-specic enthalpy h, and N
s
species
mass fractions Y
is neglected.
Mixture mass density and temperature T are specied algebraically using ap-
propriate equations of state. A thermal equation of state that relates mixture mass
density to pressure, temperature, and composition is used to deduce the density. A
caloric equation of state that relates the energy variable to pressure, temperature, and
composition is used to deduce the temperature.
Key uid properties required are mixture mass-specic constant-volume and constant-
pressure specic heats C
v
and C
p
, mixture thermal conductivity , mixture dynamic
viscosity , and species Schmidt numbers Sc
=
Y
kin,
+f
turb
, (6.1)
where Y
=
Y
C
1
dt +f
turb
, (6.3)
where dt is the computational time step in the CFD solver, and Y
and Y
A
s
ij
f
s
i
x
s
j
should not be large compared to unity. It is possible that the scale factors can be
manipulated to give lesser or greater weight to individual dependent or independent
variables. In our calculations, the independent variables are the beginning-of-time-step
species mass fractions, temperature, and pressure (initial condition of equation 2.10),
and the dependent variables are the end-of-time-step species mass fractions (solution
of equation 2.10). One choice for scale factors is the maximum possible values of each
variable in a ame. For species, a simpler alternative is to use a single value, for example,
0.1 or 0.01, for all species. Both choices work well in our experience.
ISAT will perform best if the tabulated function f (x) is approximately linear in
x, since a piecewise-linear approximation is used. A transformation of variables can be
performed before calling the ISAT routines to reduce the level of nonlinearity. In our
calculations, since chemical reaction rates are generally proportional to the exponential
of the inverse of temperature, exp(C
T
/T) is used as the independent temperature
variable for tabulation. C
T
is a constant and is set to the inverse of the largest activation
temperature in the reaction mechanism used. Transformations of other variables are
possible, but are not used in our calculation.
The error tolerance is used to specify the accuracy of the scaled linear interpola-
tion. The smaller the error tolerance, the more accurate the interpolation. It is also used
to determine the extent of the region of accuracy (ROA, Chapter 2.4), or to determine
104
whether to grow a current ROA or to add a new ROA to the table. Therefore, specifying
too small an error tolerance results in a large table and more CPU time to build and
to retrieve from the table. For our applications, and with the scalings and transforma-
tions specied above, an error tolerance in the range of 10
4
to 10
3
generally yields a
satisfactory compromise between accuracy and CPU time.
For a given problem and a given value of error tolerance, the primary parameter
controlling the eciency and memory demands of ISAT is the number of binary trees,
n
t
. In attempting to retrieve, ISAT examines up to n
t
ROAs to see if the query point is
contained within them. Larger values of n
t
lead to smaller tables because the number of
adds is decreased (more grows) and less time is needed to build the table. However, the
time to retrieve increases with n
t
because more trees are being traversed and more ROAs
are being examined. In general, the optimal value of n
t
depends on the complexity of the
problem, the storage available, the number of queries required, and the error tolerance.
For the 16-species methane mechanism with an error tolerance of 10
3
, n
t
= 128 gives
good performance. For the 122-species soot mechanism, with an error tolerance of 10
3
,
and additional considerations discussed below, n
t
= 1024 oers a good start.
The mapping gradient matrix, A, provides the coecient to perform the piecewise
linear interpolation:
f (x
q
) = f (x
0
) +Ax,
or in scalar format
f
q
i
= f
0
i
+
N
ind
j=1
A
ij
(x
q
j
x
0
j
) for i = 1, 2, . . . , N
d
,
where N
ind
and N
d
are the number of independent and dependent variables, respectively.
No explicit expression is generally available for the mapping gradient matrix so that it
has to be computed numerically.
A straightforward approach is to use a nite-dierence approximation to com-
pute A. That involves N
ind
function evaluations (or direct integrations) corresponding
to perturbations in each independent variable. This is computationally expensive but
105
relatively accurate. An alternative is to make use of the Jacobian matrix that is com-
puted internally in the ODE solver used to solve equation (2.10). The Jacobian matrix
is related to the mapping gradient matrix [48]. This approach is less expensive com-
putationally but results in a less accurate approximation of the gradient matrix. Both
approaches have been implemented in our calculations.
In the case where all the elements of the mapping gradient matrix are set to
zero, ISAT returns a piecewise constant approximation to f (x
q
). The advantage of this
simplication is that there is no need to compute the gradient matrix. However, it
requires many more table entries for a specied error tolerance. In our calculations of
nonpremixed ames, the piecewise constant approximation has been found inadequate.
The problem with using ISAT with a large chemical mechanism is that if all the
species are used as independent variables, then the table will become so large that there
is not enough computer memory to hold it; moreover, the CPU time required to build
and retrieve from the table and to compute the mapping gradient is so high that the
benet of using ISAT is lost. To illustrate the problem, consider the memory that is
required for each entry in an ISAT table. Each entry consists of the tabulation point
(N
ind
elements), the reaction mapping (N
d
elements), the mapping gradient (N
ind
N
d
elements), a unitary matrix (N
ind
N
ind
elements), and a diagonal matrix (N
ind
elements). If N
ind
= N
d
and double-precision real variables are used, then each entry
needs 8(2N
2
ind
+3N
ind
) bytes. If we have a 100-species mechanism and all the species
are used as independent variables, then each table entry occupies approximately 160
k-bytes of memory. A large memory space of 1.6 G-bytes could hold only 10
4
table
entries, which is not enough for nonhomogeneous cases such as nonpremixed ames. To
build and search such a large table is time consuming. And to compute the mapping
gradient needed for each table entry, a nite-dierence method requires performing the
direct integration 100 times, a formidable task.
However, it is not necessary to retain all the species as independent variables,
because they are not, in fact, independent. For example, the sum the species mass
fractions must be equal to unity. Therefore, it is more rigorous to use a reduced set of
species mass fractions as independent variables. However, the number and the identity of
106
the species that should be retained are unknown in general. Formal techniques developed
for chemistry reduction can be used to solve this problem, such as the quasi-steady-
state and partial-equilibrium assumptions, the rate-controlled constrained- equilibrium
method, and so on (Chapter 2.3). A simpler solution is to use a heuristic approach
based on physical reasoning to determine the independent variables. For example, fuel,
O
2
, CO
2
, H
2
O, and CO were selected as independent variables in a study of HCCI
engine auto-ignition using ISAT with a 40-species n-heptane mechanism [56]. In our
calculations, this heuristic approach is adopted and the following species are chosen for
the detailed soot mechanism: fuel, O
2
, CO
2
, H
2
O, CO, H
2
, O, OH, NO, N
2
O, C
2
H
2
,
A
1
, A
2
A
3
, and A
4
. NO and N
2
O are included for NO
x
kinetics. The A
i
s are soot
precursor species (PAHs) and are chosen together with C
2
H
2
to characterize the soot
kinetics.
The performance of ISAT will deteriorate as the inhomogeneity of the system
increases. This is because the number of table entries required is large and the reuse fre-
quency for each recorded table entry is low. The degree of inhomogeneity in nonpremixed
ames is relatively large compared to premixed ames. To improve the performance of
ISAT for nonpremixed ames, additional measures have to be taken. One idea is to
avoid using ISAT where and when the ISAT performance is poor. For example, direct
integration may be used in some regions, while ISAT is used in the others; in a transient
calculation, we may start to use ISAT after certain period of time to bypass the initial
transient. Another idea is to build several tables simultaneously. Each table is assigned
to a certain region of the ame simulated according to its temperature range, to its
mixture fraction values, or to a combination of the two. Eective use of ISAT always
is a compromise between accuracy and computational cost. In dierent tables, we can
use dierent sets of independent variables, error tolerances, numbers of trees, and ways
of computing the gradient matrix. We may even devise a scheme where we use ISAT
for one part of the chemical mechanism and direct integration for the remainder. In our
calculations, the approach of multiple tables according to specied temperature ranges
has been implemented.
107
The performance of ISAT is measured by the fractions of queries that result in
retrieves, adds, and grows. It has been shown [56] that the ratio of CPU time for an add
to that for a direct integration, t
A
/t
DI
, varies little with chemical mechanism, ISAT
control parameter set, and CFD mesh size, and that it is close to three. The break-even
point for ISAT can be estimated as,
t
A
N
A
= t
DI
N
Q
,
where N
A
and N
Q
are the total numbers of adds and queries, respectively. If we neglect
the grows, then break-even point in terms of the fraction of queries that result in retrieves
is,
N
R
N
Q
= 1.0
t
DI
t
A
67%,
where N
R
is the total number of retrieves. That is, if N
R
> 70%N
Q
, then we would
anticipate a net saving in CPU time compared to the direct integration. As an example,
Table 6.1 shows the statistics of using ISAT in modeling the Sandia D ame [10] with
the 16-species methane/air mechanism. Three tables were used with the temperature
ranges for tables 1 to 3 being: T < 800K, 800K T 1500K, and T > 1500K. For
all tables, the number of trees is 128, the error tolerance is 10
3
for the piecewise-linear
approximation and 10
4
for the piecewise-constant approximation. Table 6.1 shows
that the piecewise constant approximation is not suitable for our applications: the ISAT
table is full (indicated by the replace column) and the ISAT calculation results are not
accurate compared to the direct-integration results (not shown).
Table 6.2 shows the accuracy of ISAT by comparing the results from ISAT and
from direct integration (DI). Three locations in the ames are monitored and the temper-
atures (T) and axial velocities (w) at those locations are tabulated. The three locations
are chosen such that one is in the ame zone (location 2), one is on the fuel-rich side
of the ame zone (location 1), and one is on the fuel-lean side (location 3). One ISAT
table is used, the number of trees is 128, and the error tolerance is 10
3
. From this and
more detailed comparisons, we can conclude that ISAT is suciently accurate to replace
direct integration for our chemical source term calculations. Both the nite-dierence
108
Table 6.1. Statistics of Using ISAT
Queries Retrieves Grows Adds Replaces
Piecewise
Linear
Table 1 2.84 10
6
2.84 10
6
5.20 10
3
1.54 10
2
0
2 1.53 10
6
1.50 10
6
3.74 10
4
5.30 10
2
0
3 1.04 10
6
9.79 10
5
5.58 10
4
8.06 10
2
0
Piecewise
Constant
Table 1 2.68 10
6
2.52 10
6
2.66 10
4
2.42 10
4
1.05 10
5
2 1.54 10
6
1.27 10
6
8.05 10
3
2.42 10
4
2.34 10
5
3 1.06 10
6
8.48 10
5
7.69 10
3
2.42 10
4
1.77 10
5
(FD) and the Jacobian-matrix (JM) approaches are used to compute the gradient ma-
trix: Table 6.3 shows the ISAT statistics for the two approaches. The dierence in the
number of adds and grows is due to dierences in the mapping gradients since all the
other ISAT controls are the same. The values of the mapping gradient given by the FD
method allow larger ROAs, and hence lead to fewer adds. Not shown in the table is
that the CPU time per add is larger with the FD method than with the JM method.
This is not surprising since the FD method obtains the mapping gradient by conducting
additional direct integrations. The CPU time per grow is nearly the same for the two
methods.
The performance of ISAT ultimately is measured by the time savings of using
ISAT compared to direct integration. Figure 6.1 shows the ratio of ISAT CPU time to
that of direct integration as a function of time step. The CPU time is counted only for
chemical source term calculations. The ame is the Sandia D ame and the mechanism
is the 16-species methane/air mechanism. Three tables are used and for each table, the
full set of species is used as independent variables, the nite-dierence method is used
for the gradient matrix evaluation, the number of tree is 128, and the error tolerance is
10
3
. ISAT is put into use after 1000 iterations starting from the initial conditions. In
the gure, the solid line represents the ratio of CPU time at each time step, while the
109
Table 6.2. Comparisons Between ISAT and Direct Integration
Location 1 Location 2 Location 3
T w T w T w
DI 417.73 4.577 1415.53 8.600 972.84 9.204
ISAT FD 417.53 4.578 1415.23 8.597 973.04 9.206
ISAT JM 417.32 4.576 1415.83 8.601 972.06 9.202
Table 6.3. Statistics of the Two Approaches to the Mapping Gradient Matrix Calcula-
tion
Queries Retrieves Grows Adds
Finite-Dierence 5.466 10
6
5.370 10
6
9.444 10
4
1.48 10
3
Jacobian-Matrix 5.466 10
6
5.368 10
6
9.533 10
4
2.47 10
3
110
Time Steps
T
o
t
a
l
C
P
U
T
i
m
e
(
s
)
C
P
U
t
i
m
e
r
a
t
i
o
o
f
I
S
A
T
t
o
D
I
0 1000 2000 3000
0
3000
6000
9000
12000
15000
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Direct Integration
ISAT
Current
Cumulative
Fig. 6.1. Performance of ISAT in terms of CPU time and CPU time ratio
111
dashed line represents the ratio of cumulative CPU time. The CPU time ratio drops
quickly even at the beginning and approaches an asymptotic speedup of approximately
10 for this problem.
6.4 Detailed Soot Modeling
The detailed soot model used in our CFD simulations is the same as that used
in the TSL simulations (Chapter 5.3, 5.4), except for the implementation of the method
of moments. The dierence comes from the spatial congurations of the two models. In
the TSL model, the geometry is one-dimensional (axial direction) and the two reactors
interact with each other in the radial direction (Chapter 5.2), while in the CFD modeling,
the geometry is three-dimensional with unstructured meshes (Chapter 6.2). In this
section, the instantaneous and the mean transport equation for soot moments are derived,
the calculations of moment source terms are discussed, and the implementation of the
soot model is validated.
6.4.1 Soot Moment Transport Equations
The denition of soot moments (equation 3.3) and the method of moments have
been introduced in Chapter 3.4. To derive the instantaneous transport equations for
soot moments, consider the conservation of soot size distribution, N
, in a xed control
volume, V , shown in Figure 6.2. Here, N
is expressed as,
Rate of Increase
of N
in V
=
Rate of Production
of N
in V
at V Boundary
. (6.4)
In integral form, the conservation of N
is written as,
d
dt
_
V
N
dV =
_
V
dV
_
A
A, (6.5)
112
V
N
t
dV =
_
V
dV
_
V
V . (6.6)
This leads to the dierential form of the conservation equation for N
,
N
t
+
J
i,
x
i
=
Q
. (6.7)
The ux of the soot particles of size can be expressed as,
J
i,
= N
u
i,
= N
(u
i
+u
i,diff,
), (6.8)
where u
i
is the bulk ow velocity and u
i,diff,
is the diusion velocity of the soot
particles of size and it can be expressed as
u
i,diff,
= u
i,,
+u
i,T,
+u
i,P,
+u
i,f,
, (6.9)
where u
i,,
, u
i,T,
, u
i,P,
, and u
i,f,
are the diusion velocities due to the concen-
tration gradient, temperature gradient, pressure gradient, and external forces (such as
gravity), respectively. Terms u
i,,
and u
i,T,
are usually considered in laminar cases,
while u
i,P,
and u
i,f,
are usually ignored.
The production rate of the soot particles of size can be regarded as the sum of
the contributions from soot particle nucleation, coagulation, surface growth, oxidation,
and condensation,
=
dN
dt
s
=
dN
dt
nucleation
+
dN
dt
coagulation
+
dN
dt
surface growth
+
dN
dt
oxidation
+
dN
dt
condensation
.
(6.10)
The various contribution terms have been discussed in Chapter 3.4.
114
The conservation equation for N
t
+
u
i
N
x
i
=
(u
i,diff,
N
)
x
i
+
dN
dt
s
, for = 1, 2, . . . , . (6.11)
To obtain the moment transport equations, one multiplies equation (6.11) by the r
th
power of the soot particle mass of size , m
r
is independent
of position and time, the soot moment transport equations are written as,
M
r
t
+
u
i
M
r
x
i
=
(u
i,diff
M
r
)
x
i
+
dM
r
dt
s
for r = 0, 1, . . . , . (6.12)
Since Favre-averaged quantities are employed in the CFD code, equations (6.12)
and/or (6.11) are reformulated as follows:
(M
r
)/
t
+
(u
i
M
r
)/
x
i
=
(u
i,diff
M
r
)
x
i
+
dM
r
dt
s
. (6.13)
By invoking the continuity equation, this becomes
(M
r
)
t
+
(u
i
M
r
)
x
i
+M
r
u
i
x
i
=
(u
i,diff
M
r
)
x
i
+
dM
r
dt
s
for r = 0, 1, . . . , .
(6.14)
Introducing a new quantity, M
mr
,
M
mr
M
r
, (6.15)
equation (6.13) becomes,
(M
mr
)
t
+
(u
i
M
mr
)
x
i
+ =
(u
i,diff
M
mr
)
x
i
+
dM
r
dt
s
for r = 0, 1, . . . , .
(6.16)
Equation (6.16) is more convenient to manipulate for obtaining the mean moment trans-
port equations, by virtue of the following relations:
M
mr
=
M
mr
=
M
r
M
r
=
M
mr
. (6.17)
115
To obtain the mean moment transport equations, the Reynolds decompositions,
M
mr
=
M
mr
+M
mr
, and u
i
= u
i
+u
i
,
are substituted into equation (6.16), and then the whole equation is Reynolds averaged.
This leads to
(
M
mr
)
t
+
( u
i
M
mr
)
x
i
+
(u
i
M
mr
)
x
i
=
dM
r
dt
s
. (6.18)
The diusion term, u
i,diff
M
mr
, has been neglected since it is small compared to
the turbulent transport term, u
i
M
mr
. With a gradient transport model, the turbulent
transport term is modeled as,
u
i
M
mr
=
T
P
rT,s
M
mr
x
i
, (6.19)
where
T
is an eective turbulence viscosity that, in the case of a standard k- turbulence
model, is given by,
T
= C
k
2
/, (6.20)
C
= 0.9). Here P
rT,s
is an eective turbulent Schmidt
number and is set to unity in our calculations.
The mean moment transport equations are nally written as
(
M
mr
)
t
+
( u
i
M
mr
)
x
i
=
x
i
_
T
P
rT,s
M
mr
x
i
_
+
dM
r
dt
s
for r = 0, 1, . . . , . (6.21)
In our calculations, r = 0, 1, . . . , 5; that is, six moment equations are retained.
Since soot is part of the mixture, total mass conservation is enforced as follows:
N
s
=1
Y
+Y
s
= 1, (6.22)
where Y
and Y
s
are species and soot mass fractions, respectively. According to the
denition of soot moments (equation 3.3), the rst moment, M
1
, represents the total
116
soot mass per unit volume. Therefore, the soot mass fraction is given by
Y
s
=
M
1
t
, (6.23)
where
t
= + M
1
is the density of the mixture including soot particles. Here is the
density of the gas-phase mixture and is computed by the ideal-gas equation of state.
Note that the dimensions of and M
1
must be consistent in the calculations.
6.4.2 Soot Moment Source Terms
The source terms in the soot moment transport equations consist of contributions
from soot particle nucleation, coagulation, surface growth, surface condensation, and soot
oxidation (Chapter 3.4 and 5.4). Their computation can proceed in the same manner as
the gas-phase species source terms (Chapter 2.2): that is, an operator-splitting approach
is invoked and a set of six odes (equation 2.10) is solved. In the current implementation,
the eects of turbulent uctuations on soot moment source terms are neglected.
In the soot model, the calculations of the moment source terms involve several
gas-phase species and therefore modify the production rate of those species. As a result,
the solution of the six odes for soot moments is coupled to the solution of the N
s
odes
for the gas-phase species source terms:
_
_
dY
dt
= S
for = 1, 2, . . . , N
s
dM
r
dt
= S
r
for r = 0, 1, . . . , 5.
However, soot moments characteristically are very large numbers compared to the gas-
phase species mass fraction: for example, M
0
may be on the order of 10
16
. The following
transformation is made to eliminate potential numerical problems:
Y
r
= ln(M
r
+C
r
) for r = 0, 1, . . . , 5, (6.24)
117
where C
r
is an arbitrary positive constant and is specied such that as M
r
approaches
zero, Y
r
will not approach . Then, the set of odes for soot moments becomes
dY
r
dt
=
S
r
M
r
+C
r
for r = 0, 1, . . . , 5. (6.25)
6.4.3 Validation
Validation of the the implementation of the moment transport equations includes
validating soot moment source term calculations and validating soot moment transport
calculations. The moment source term calculations have been validated by using the
SENKIN program [188]. The soot model has been implemented into SENKIN by Fren-
klach [189] and equations similar to (6.25) are solved. For the same initial conditions,
(e.g., temperature, pressure, and species composition), the corresponding subroutines
in our implementation generated the same results as SENKIN. This indicates that the
moment source terms are computed correctly in our implementation.
The moment transport calculations have been validated by utilizing the fact that
the moment transport equations are similar to the species transport equations. The
subroutines in our implementation were set up such that the transport of one of the
gas-phase species in the mechanism, e.g. the OH species, is solved instead of the soot
moments. The results were then compared with those calculated by the species trans-
port equation solver. The same results were obtained indicating that the soot moment
transport equations are implemented correctly.
6.5 Detailed Radiation Modeling
Radiative heat loss is accounted for through a source term in the energy equation:
for example, the
Q
rad
term in equation (2.4). In our CFD calculations, the heat source
term is determined by the solution of the radiative transfer equation (RTE, Chapter 4.2)
for a radiatively participating medium with consideration of absorption eects (equation
4.3). Two radiation models have been implemented for the determination of the heat
source term. The rst model assumes the participating medium is gray and employs the
spherical harmonic P
1
approximation (Chapter 4.3) for the solution of RTE. This model
118
is called the P
1
-Gray model. The second model also employs the P
1
approximation for
the RTE solution, but the nongray characteristics of the medium are taken into account
by using the full-spectrum k-distribution (FSK) method (Chapter 4.4). This model is
called the P
1
-FSK model. In the following, the two models and their validations are
described.
6.5.1 P
1
-Gray model
The averaged radiation source term in the averaged energy equation (2.17) can
be written as (according to equation 4.5),
q
R
) =
_
0
_
4
I
b
)
)
_
d,
where the angle braces are used to replace over-lines as the averaging operators, and G
=
_
4
I
d.
Here G
represents the total radiative intensity impinging on a point from all directions.
If we neglect turbulence-radiation interactions, the averaged source term becomes,
q
R
) =
_
0
_
4
)I
b
)
)G
)
_
d. (6.26)
This expression is dicult to evaluate because of the spectral dependence of the ab-
sorption coecient,
. To overcome this
diculty, a gray medium assumption is commonly invoked, and the Planck-mean absorp-
tion coecient is used to represent the total absorption coecient of the gray medium.
The Planck mean absorption coecient is a spectrally averaged absorption coecient
weighted by the Planck function and is dened as,
p
_
0
I
b
d
_
0
I
b
d
.
119
With the gray medium assumption, equation (6.26) becomes
q
R
) = 4
p
)I
b
)
p
)G). (6.27)
Since turbulence-radiation interactions are neglected, the mean Planck-mean absorption
coecient,
p
), and the mean black-body intensity, I
b
), can be evaluated in terms of
mean temperature and mole fractions of radiating species. As in the radiation model in
the TSL modeling (Chapter 5.5), curve-ts are used for gas-phase Planck-mean absorp-
tion coecients and Rayleighs theory for that of soot particles.
The determination of the mean incident radiation, G), involves the solution of
the mean RTE for mean the total radiative intensity, I). The spherical harmonic P
1
method is employed for this purpose. With the Planck mean absorption coecient,
the RTE for the spectral radiative intensity (equation 4.3) can be integrated over the
spectrum and the RTE for the total radiative intensity is then written as,
dI
ds
=
p
I
b
p
I. (6.28)
The mean total RTE without turbulence/radiation interactions is then written as
dI)
ds
=
p
)(I
b
) I)). (6.29)
The P
1
method transforms the mean RTE into the following form:
p
)
G) 3
p
) (G) 4I
b
)) = 0. (6.30)
This governing equation for G) is a Helmholtz equation with variable coecients and
source term; it can be solved using the same solver as that used in the CFD code for
the conservation equations with the convection disabled. The boundary condition for
equation (6.30) is one of the third kind and is written as [88],
2
3
n G) +
p
)G) = 4
p
)I
b
), (6.31)
120
where n is the unit surface normal of a boundary wall and points towards the radiating
medium; here unity surface emissivity has been assumed.
6.5.2 P
1
-FSK model
The full-spectrum k-distribution (FSK) method has been developed for accurate
calculations of nongray radiative heat transfer in radiatively participating media. The
essence of the method is to reorder the erratic spectral absorption coecient into a
monotonically increasing function over the entire spectrum (Chapter 4.4). In this section,
the FSK method is coupled with the P
1
solution method for the RTE to evaluate the
mean radiative source term in the mean energy equation.
We start from the averaged source term, equation (6.26),
q
R
=
_
0
_
4
I
b
_
4
d
_
d, (6.32)
where the averaging operator is dropped out for convenience and the incident radiation
function is written in terms of the spectral intensity I
0
, )), where k is an ab-
sorption coecient variable and
0
, ) is the mediums spectral absorption coecient
evaluated at a reference state
0
. The
: for example,
= (T, P, ), and the state variables are temperature, T, pressure, P, and mole frac-
tions of radiating species, . For an inhomogeneous medium, which is of interest in
combustion simulations, the absorption coecients at dierent local states are assumed
to be correlated. One implication of this assumption is that at every wavenumber where
0
) has one and the same value k
0
,
. The
spectral absorption coecient in equation (6.32) can then be expressed as,
) =
_
0
k
(k
0
)(k
0
, ))dk. (6.33)
121
By substituting this expression into equation (6.32), it becomes
q
R
=
_
0
4k
(k
0
)dk
_
0
I
b
(k
0
, ))d
_
0
d
_
0
k
(k
0
)dk
_
0
I
(k
0
, ))d.
(6.34)
New quantities, I
k
and f(T,
(k
0
, ))d, (6.35)
f(T,
0
, k) =
1
I
b
_
0
I
b
(k
0
, ))d, (6.36)
where f(T,
0
, k) is a full-spectrum k-distribution. The rst parameter in f(T,
0
, k), T,
represents the temperature at which I
b
and I
b
are evaluated and this is called the Planck
temperature. The second parameter,
0
, represents the state at which
is evaluated
and it also contains a temperature component, which we call the temperature. The
third parameter, k, is the dependent variable. It can be shown from the denition of
f(T,
0
, k) that the integration of f(T,
0
, k) over k equals unity,
_
0
f(T,
0
, k)dk = 1. (6.37)
The cumulative k-distribution, g, is then written as,
g =
_
k
0
f(T,
0
, k)dk. (6.38)
This relation can be inverted to obtain the variation of k with g, and that is called
the k-g distribution and is denoted as k(T,
0
, g). Note that for the same absorption
coecient, e.g.
0
), dierent k-g distributions can be constructed with dierent
Planck temperature, T.
In terms of the new quantities, equation (6.32) becomes,
q
R
=
_
0
4k
(k
0
)I
b
f(T,
, k)dk
_
0
d
_
0
k
(k
0
)I
k
dk. (6.39)
122
By introducing another new quantity, I
g
0
,
I
g
0
=
I
k
f(T
0
,
0
, k)
, (6.40)
and an a function to account for the dierent Planck temperatures in constructing k-gs
for the same
),
a(T, T
0
, g
0
) =
f(T,
0
, k)
f(T
0
,
0
, k)
=
dg(T,
0
, k)
dg(T
0
,
0
, k)
. (6.41)
equation (6.32) becomes,
q
R
=
_
1
0
4k
(k
0
)a(T, T
0
, g
0
)I
b
dg
0
0
d
_
1
0
k
(k
0
)I
g
0
dg
0
=
_
1
0
4k
(k
0
)a(T, T
0
, g
0
)I
b
dg
0
_
1
0
k
(k
0
)G
g
0
dg
0
,
(6.42)
where
g
0
=
_
k
0
f(T
0
,
0
, k)dk. (6.43)
In equation (6.42), the spectral variation of in space, where the variation is irregular,
is transformed into the variation of k in g space, where the variation can be shown to
be smooth. Many fewer quadrature points are required for the integrations in equa-
tion (6.42) than those in equation (6.32). As a result, the number of spectral RTEs
(transformed into the g space as well) that need to be solved (for I
g
or G
g
) is reduced
substantially.
To transform or reorder the spectral RTE (equation 4.3) into the g space, we
multiply the equation by (k (
0
, )) and integrate it over the entire spectrum. This
leads to,
dI
g
0
ds
= k
(k
0
)[a(T, T
0
, g
0
)I
b
(T) I
g
0
]. (6.44)
123
Applying the P
1
method to this equation, we obtain the governing equation for spectral
incident radiation in the g space, G
g
0
,
1
k
(k
0
)
G
g
0
3k
(k
0
)
_
G
g
0
4a(T, T
0
, g
0
)I
b
_
= 0, (6.45)
with the boundary condition of the form,
2
3
n G
g
0
+k
(k
0
)G
g
0
= 4k
(k
0
)a(T, T
0
, g
0
)I
b
. (6.46)
The spectral G
g
0
is solved at representative spectral g
0
locations, which are usu-
ally determined by a Gaussian quadrature scheme. A number of 10 quadrature points or
fewer is generally sucient for the quadrature accuracy because of the smooth behavior
of k in g space. Placing more points on large values of g can increase the accuracy of
the radiation calculations because radiation is substantial at large k-g values.
After solving for G
g
0
, the radiative heat loss is calculated by
q
R
=
nq
i=1
W
g
0i
_
k
g
0i
[4a
g
0i
(T, T
0
, g
0
)I
b
(T) G
g
0i
]
_
, (6.47)
where nq is the number of quadrature points, W is the quadrature weight, W
g
0i
, a
g
0i
,
and G
g
0i
are values at a quadrature point.
To solve equation (6.45), k
(k
0
) needs to be determined. That comes from equa-
tion (6.33). One can construct a k-g distribution, k
(T,
(k
0
) = k
(T,
, g) = k
(T
0
,
, g
0
). (6.48)
This way of determining k
(k
0
) leads to one version of the FSK method, the full-spectrum
correlated-k (FSCK) method. Another way to determine k
(k
0
) is to use the scaling
approximation, which requires the spectral and the spatial dependence of the absorption
124
coecient to be separable [88],
k
(k
0
) = k
(T,
0
, g)u(
0
), (6.49)
where u(
0
) is called the scaling function and its calculation is outlined in [88]. This
leads to another version of the FSK method, the full-spectrum scaled-k (FSSK) method.
The FSSK method is employed in our calculations since it is shown to consistently
outperform the FSCK method [88].
6.5.3 Validation
The implementation of the P
1
-Gray model was tested by devising a problem that
has an analytical solution and then using the GMTEC code with the P
1
-Gray model
to solve the problem numerically. The governing equation of the P
1
-Gray model is
a Helmholtz equation with variable coecients and source term (equation 6.30) and
a boundary condition of the third kind. The problem devised is to solve on a two-
dimensional square (of size L by L) an equation of the form
2
GK
2
G+S = 0, (6.50)
where K and S are constants and G is the quantity to be solved for. The boundary
conditions for the problem are:
at x = 0 and y = 0 : n G = 0
at x = L and y = L : n G+C
1
G = 0,
(6.51)
where C
1
is a constant and n denotes the surface normal at the boundaries.
Upon choosing a specic form of S,
S = cos
x
1
L
cos
y
1
L
(6.52)
125
where
1
and
1
are constant, the analytical solution to the problem is
G(x, y) =
L
2
cos
x
1
L
cos
y
1
L
L
2
K
2
+
2
1
+
2
1
, (6.53)
and this satises equation (6.50) and the boundary condition (6.51).
The numerical and analytical solutions were then compared and the implementa-
tion of the P
1
-Gray model was therefore validated.
The validation of the P
1
-FSK model was based on the validation of the P
1
-Gray
model. The dierence between the governing equations of the two models is mainly that
the P
1
-FSK governing equation (6.45) is spectral and needs to be solved at dierent
spectral g
0
locations. By setting the k
c
1
c
2
k
,h
/I
b,h
, and this leads to I
b,h
I
b,h
. Since
p,h
L
h
,h
L
h
, we have,
I
h,gray
I
h,nongray
. (6.57)
Another way to express this is that the optical thickness of a nongray medium is
much larger than that of a gray medium. This means that a nongray medium absorbs
more radiation than a gray medium. From this analysis, we conclude that the gray
medium assumption will over-predict the radiative heat loss from a ame. Its eect on
ame temperature depends on the distributions of radiating species and the conguration
of the ame, as will be discussed in the following. This analysis also suggests that the
gray-medium assumption may result in a larger error than might result from neglecting
radiation all together.
The eects of nongray radiation calculations are illustrated by simulating an
oxygen-enriched ame using the CFD code described above. Figure 6.5 plots the ax-
ial proles of the radiation heat ux at the peripheral side of the ame and the proles
of the centerline ame temperature, calculated by the two radiation models. The heat
ux proles show that the gray radiation model substantially over-predicts the radiation
loss downstream in the ame. This is expected since the spectrally radiatively partici-
pating species have much higher mass concentrations downstream than upstream in the
ame. The slightly larger upstream heat ux from the nongray model may be the re-
sult of two competing factors: on one hand, the nongray model predicts less radiation
loss and therefore higher ame temperature than the gray model; on the other hand,
the higher ame temperature generates larger radiation heat ux. Figure 6.6 shows the
radial proles of ame temperature at four axial locations. It shows that the gray model
predicts a larger ame temperature upstream but a lower temperature downstream. The
temperature dierence at x/d = 80 is about 80K and at x/d = 320 about 70K. Note
that the upstream temperature dierence occurs only at the ame sheet while the down-
stream temperature dierence occurs throughout the domain. The resulting temperature
distribution will have an impact on predictions of NO
x
and soot formation and oxidation.
132
Height (m)
R
a
d
a
t
i
o
n
H
e
a
t
F
l
u
x
(
W
/
m
2
)
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 0.2 0.4 0.6 0.8 1
0
2000
4000
6000
8000
10000
12000
0
500
1000
1500
2000
2500
3000
3500
4000
Flux
Temperature
P
1
-Gray
P
1
-FSK
Fig. 6.5. Axial prole of radiation heat ux and temperature calculated by the two
radiation models: gray vs. nongray
133
Radial Distance (m)
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 0.02 0.04 0.06
0
400
800
1200
1600
2000
2400
2800
P
1
-Gray
P
1
-FSK
x/d = 80
x/d = 160
x/d = 240
x/d = 320
Fig. 6.6. Radial proles of temperature calculated by the two radiation models at four
axial locations: gray vs. nongray
134
6.6.1.2 Eects of Soot Radiation
Compared to gas-phase species, which radiate at discrete spectral locations, soot
emits and absorbs thermal radiation in a continuous spectrum over the visible and in-
frared regions. It has been shown that soot emission often is considerably stronger than
the emission from the gas-phase species, and it is expected that, because of the presence
of soot particles of varying sizes, the spectral oscillation from gas-phase radiation tend
to be damped out so that the gray medium assumption becomes reasonable.
In oxygen-enriched ames, a great deal of soot is generated because of the in-
creased ame temperature. The soot distribution used in the radiation calculations is
obtained as follows: the predictions of soot formation and oxidation are performed with
the detailed soot model [175] with the method of moments in the CFD code. Transport
equations for soot moments are solved and the distributions of soot volume fraction are
obtained. This is done with a detailed mechanism. The calculated distribution of soot
volume fraction (SVF) is then manipulated such that the manipulated SVF distribution
gives the exactly the same axial prole of equivalent soot volume fraction as obtained
from experiments. Figure 6.7 shows the distribution of this manipulated SVF. This SVF
distribution is then imposed on the same computational grid with the soot calculations
turned o and a reduced chemical mechanism used.
Soot radiation is incorporated into the FSK calculations of gas-phase radiation
by assuming gray soot properties, which is expected to be a good assumption for a soot
ensemble of varying particle sizes. The Planck-mean approach [88] is used to evaluate
the soot absorption coecient, and that is added directly to the k values in the k-g
distribution obtained from the FSK method for gas-phase species.
Figure 6.8 shows the axial proles of radiation heat ux and centerline ame
temperature calculated by two versions of the P
1
-FSK model, one with soot and one
without soot. Figure 6.9 plots the the radial proles of the ame temperature at four axial
locations. These gures show that soot radiation is very strong for this sooting ame with
a maximum soot volume fraction of 1.710
06
. For example, the maximum temperature
dierence at location x/d = 160 is more than 300K between calculations with and
without the soot contribution, and the dierence is nearly 200K at location x/d = 240,
135
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Z
Fv
1.6E-06
1.5E-06
1.4E-06
1.3E-06
1.2E-06
1.1E-06
9.5E-07
8.4E-07
7.4E-07
6.3E-07
5.3E-07
4.2E-07
3.2E-07
2.1E-07
1.1E-07
Height (m)
Height (m)
E
q
u
i
v
a
l
e
n
t
S
o
o
t
V
o
l
u
m
e
F
r
a
c
t
i
o
n
0 0.2 0.4 0.6 0.8 1
0.0E+00
4.0E-06
8.0E-06
1.2E-05
1.6E-05
2.0E-05
Experiment
Manipulated
Fig. 6.7. (Manipulated) distribution (axial prole left, contour right) of soot volume
fraction for a propane ame with 40% oxygen index and 21.8 m/s jet velocity
136
while the maximum temperature dierence between gray and nongray calculations is less
than 80K.
The large upstream heat ux obtained from the P
1
-FSK model with the soot
contribution in Figure 6.8 represents a breakdown of the P
1
method when applied to a
strongly sooting turbulent jet ame. This is explained in the following section.
6.6.1.3 Applicability of the P
1
method to Jet Flames
Part of the reason that the upstream radiation heat ux is so large is the over-
predicted ame temperature that results from the simple turbulent combustion model,
which leads to fast combustion rates and therefore near equilibrium ame temperatures.
However, even with the upstream ame temperature reduced articially by several hun-
dred Kelvin, the upstream heat ux is still very high (not shown). This suggests that
the abnormal large upstream heat ux shown in Figure 6.8 may come from the P
1
approximation.
To examine the accuracy of the P
1
approximation, two types of calculations are
conducted. First, the CFD code is set up with a small cylindrical hot zone along the
centerline of the ame and in the middle of the cylindrical computational domain (a
wedge), as shown in Figure 6.10. The hot zone has a constant high temperature of 2200K
and a constant large Planck-mean absorption coecient of 5000 1/m. The temperature
and the Planck-mean absorption coecient for the rest of the computational domain are
set to 300K and a small constant value of
rest
, respectively. The temperature values
at the boundaries are the same as those of the cases shown in Figure 6.8. The radiation
heat ux at the peripheral boundary is then calculated using the P
1
-gray model in the
code. The second type of calculation is black-surface to black-surface heat transfer using
view factors. The hot zone is taken as a black small hot cylinder radiating towards the
interior of the black large cold cylinder. The radiation heat ux at the interior surface
of the large cylinder is determined by using view factors. The rst calculation should
generate a heat ux axial prole that has the same shape as that generated by the second
calculation (the peak values may not match exactly).
137
Height (m)
R
a
d
i
a
t
i
o
n
H
e
a
t
F
l
u
x
(
W
/
m
2
)
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 0.2 0.4 0.6 0.8 1
0
3000
6000
9000
12000
15000
0
500
1000
1500
2000
2500
3000
3500
4000
Flux
Temperature
P
1
-FSK w/o Soot
P
1
-FSK w/ Soot
Experiment
Fig. 6.8. Axial prole of radiation heat ux and temperature calculated by the two
radiation models: soot eects
138
Radial Distance (m)
T
e
m
p
e
r
a
t
u
r
e
(
K
)
0 0.01 0.02 0.03 0.04 0.05 0.06
0
500
1000
1500
2000
2500
x/d = 80
x/d = 160
x/d = 240
x/d = 320
P
1
-FSK w/o Soot
P
1
-FSK w/ Soot
Fig. 6.9. Radial proles of temperature calculated by the two radiation models at four
axial locations: soot eects
139
Fig. 6.10. Computational conguration to test the P
1
approximation
140
Figure 6.11 plots the axial proles of radiation heat ux calculated by the P
1
-gray
model in the CFD code and by the view factors between two black cylindrical surfaces.
The numbers in the plot, 0.1, 0.3, and 1.0, are the values of the Planck-mean absorption
coecient outside the hot zone,
rest
, for the P
1
-gray model. It can be seen from the
gure that as the value of
rest
decreases, or as the absorption coecient ratio of the
hot zone to the rest of the domain increases, the performance of the P
1
approximation
becomes progressively worse. This is the reason that the upstream heat ux in Figure 6.8
is so large when soot radiation is considered. Mathematically, if the absorption coecient
ratio is very large, the P
1
governing equation (equation 6.30 and 6.45) for regions of small
absorption coecient becomes the Laplace equation. This is true for regions upstream
and outside a jet ame. Physically, since the P
1
method approximates the radiation
intensity in a radiatively participating medium using the rst-order spherical harmonic,
it is appropriate for media where radiation comes from all directions (approximately
isotropic radiation). This is not the case for regions upstream and outside of a jet
ame. Although the P
1
method may perform better downstream in the ame where the
distribution of the radiating species is not restricted to a narrow region as upstream in
the ame, the P
1
method is in general not suitable for quantitatively accurate radiation
calculations in jet-ame-like congurations.
6.6.2 Simulation of an Oxygen-Enriched Flame
6.6.2.1 Key Sensitivities in Soot Predictions
Initial simulations (not shown) yielded soot volume fractions nearly three orders
of magnitude lower than the measured values; and computed soot levels were found
to be particularly sensitive to the soot surface growth model. To investigate the key
sensitivities in soot predictions, four test cases were run (Table 6.5). Two surface growth
mechanisms were investigated: those by Appel et al. [107] and by Kazakov et al. [105].
The steric factor accounts for the fraction of surface active sites that are available for
surface reactions. The correlation in [107] gives steric factor values that are less than 0.4
for any temperature. The correlation was adjusted in Case 3 and 4 so that the values
of are in the range 0.9-1.0, as suggested by El-Leathy et al. [198]. In all four cases,
141
Height (m)
R
a
d
i
a
t
i
o
n
H
e
a
t
F
l
u
x
(
W
/
m
2
)
0 0.2 0.4 0.6 0.8 1
0
1000
2000
3000
4000
5000
6000
7000
8000
P
1
-Gray
Black-Surface
1.0
0.3
0.1
Fig. 6.11. Axial proles of radiation heat ux: study of the P
1
approximation
142
the P
1
-gray model with soot radiation was used, except for Case 4, where soot radiation
was turned o.
Table 6.5. Case specications for soot sensitivity study.
Case Surface growth mechanism Steric factor, Soot radiation
1 Appel et al. [107] Original ( < 0.4) On
2 Kazakov et al. [105] Original ( < 0.4) On
3 Kazakov et al. [105] Modied ( 1.0) On
4 Kazakov et al. [105] Modied ( 1.0) O
Figure 6.12 shows axial proles of equivalent soot volume fraction (F
v,eq
), where
F
v,eq
is the value over a path length equal to the nozzle diameter that gives the same
extinction as the actual prole of the centerline laser extinction measurements [12].
Cases 1 and 2 reveal the sensitivity of soot yields to the soot surface-growth mechanism.
Kazakov et al.s model [105], recommended by El-Leathy et al.s empirical study [198],
produces a soot yield almost one order of magnitude higher than that of Appel et al.s
[107] model. Cases 2 and 3 illustrate the sensitivity of soot yields to the steric factor. A
steric factor close to unity produces a soot yield almost one order of magnitude higher
than for steric factors less than 0.4. The change in steric factor also changes the shape of
the axial prole; the original correlation (lower steric factor) better resembles the shape
of the measured prole. The over-predicted upstream soot in Case 3 is likely due to the
over-prediction of upstream ame temperature from the simple turbulent combustion
model.
Cases 3 and 4 show the sensitivity of soot yields to ame temperature and reveal
the interactions between soot formation and soot radiation. Turning o soot radiation
(Case 4) increases ame temperature and enhances soot formation. Although this case
provides a peak value of F
v,eq
and a shape of the axial prole that are closer to the
143
Height (m)
E
q
u
i
v
a
l
e
n
t
S
o
o
t
V
o
l
u
m
e
F
r
a
c
t
i
o
n
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.0E+00
2.0E-06
4.0E-06
6.0E-06
8.0E-06
1.0E-05
1.2E-05
1.4E-05
1.6E-05
1.8E-05
Case 1
Case 2
Case 3
Case 4
Experiment
Fig. 6.12. Axial proles of equivalent soot volume fraction.
144
experiment, the result is fortuitous. An important factor not accounted for adequately
with the characteristic-time-scale turbulent combustion model is the eect of turbulent
uctuations in temperature and species composition. Given that the diusion of soot
particles is small compared to gas-phase species, the turbulent uctuations can be very
important in determining the maximum soot volume fractions. The remnant soot down-
stream in Cases 3 and 4 (Fig. 6.12) indicates insucient soot oxidation. This likely
results from the lower temperatures in the downstream region of the ame, a conse-
quence of over-predicted radiation heat loss by the gray model, as shown below. The
maximum values of F
v,eq
for each of the modeled cases and the experiment are listed in
Table 6.6.
Table 6.6 also presents global NO
x
emissions indices (EINO
x
) and radiant frac-
tions (
R
). Radiant fractions increase as soot yield increases (Cases 1, 2, and 3). The
calculated radiant fractions are all larger than the measured value; this over-prediction
likely results from the gray-gas model and over-predicted upstream ame temperatures.
As a result, the calculated NO
x
indices are under-predicted, although they respond faith-
fully to the radiant fraction changes. In the absence of the characteristic time-scale
combustion model, the calculated NO
x
index (not shown) is nearly twice the measured
value. Like soot, NO
x
formation is a slow nonlinear process that is quite sensitive to
temperature and species concentrations and their uctuations.
Table 6.6. Comparisons of global quantities soot models.
Case Peak F
v,eq
EINO
x
R
1 2.54 10
07
11.5 0.35
2 1.52 10
06
9.6 0.37
3 9.48 10
06
4.4 0.43
4 1.83 10
05
10.1 0.34
Experiment 1.62 10
05
29.8 0.24
145
6.6.2.2 Key Sensitivities in Radiation Predictions
Figure 6.13 shows the axial proles of radiant heat ux at the peripheral side
of the ame (r = 0.1 m, Fig. 6.14). The curves without symbols are obtained using
the P
1
-Gray and P
1
-FSK models with and without soot radiation, as indicated. For
these comparisons, the soot distribution used in the radiation calculations is obtained
as described in section 6.6.1.2. The so obtained F
v
distribution then is imposed on the
same computational grid with the soot calculations turned o.
The eects of nongray radiation and soot radiation can be seen by comparing these
four curves in Fig. 6.13. Nongray radiation more strongly inuences the downstream
radiative heat ux than upstream heat ux. In this oxygen-enriched ame, the eect of
soot radiation on the heat ux is much larger than the nongray-gas eects. The eects
of nongray-gas radiation, however, still are substantial even in the presence of strong
gray soot radiation. The peak radiative heat ux computed by the P
1
-FSK model with
soot radiation agrees well with the experimental value. The upstream and downstream
discrepancies between experiments and calculations result from the simple turbulent
combustion model and the spherical harmonic P
1
approximation.
The upper row of Fig. 6.14 shows computed contours of temperature (using the
P
1
-FSK model with and without soot radiation), soot F
v
(adjusted), and mass fractions
of H
2
O and CO
2
species (using P
1
-FSK with soot). It can be seen that soot is distributed
in the fuel-rich regions upstream of the nominal ame zone, while spectrally radiating
and absorbing H
2
O and CO
2
species are found within and downstream of the ame
zone. The gas-phase species are distributed more broadly downstream than upstream
in the ame. As a result, the temperature at the ame tip drops substantially because
of soot radiation (more than 300 K, see temperature contours), and the nongray eects
are stronger for downstream radiation than for upstream radiation (Fig. 6.13). The
resulting changes in temperature distribution, in turn, aect the rate of soot formation
and oxidation, ame structure, and NO
x
emission. In addition, the presence of soot
shortens ame length as indicated by the axial location of the peak temperatures (the
peak location moves from 0.58 m without soot to 0.54 m with soot, a 5% reduction).
146
Height (m)
R
a
d
i
a
n
t
H
e
a
t
F
l
u
x
(
W
/
m
2
)
0 0.2 0.4 0.6 0.8 1
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
Experiment
P
1
-FSK
P
1
-Gray
P
1
-FSK-Soot
P
1
-Gray-Soot
Fig. 6.13. Axial proles of radiant heat ux.
147
T
2400
2250
2100
1950
1800
1650
1500
1350
1200
1050
900
750
600
450
300
T (w/ soot)
0 0.05 0.1
X
0
0.2
0.4
0.6
0.8
1
H
2
O
r (m)
z
(
m
)
0.28
0.26
0.24
0.22
0.20
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.00
CO
2
1.6E-06
1.5E-06
1.4E-06
1.3E-06
1.2E-06
1.1E-06
1.0E-06
9.0E-07
8.0E-07
7.0E-07
6.0E-07
5.0E-07
4.0E-07
3.0E-07
2.0E-07
Soot Fv
3.5E-03
3.2E-03
3.0E-03
2.7E-03
2.5E-03
2.2E-03
2.0E-03
1.8E-03
1.5E-03
1.3E-03
1.0E-03
7.5E-04
5.0E-04
2.5E-04
0.0E+00
P1-Gray (w/o soot)
NO
P
1
-FSK (w/ soot)
NO
P
1
-FSK (w/o soot)
NO NO
P
1
-Gray (w/ soot)
Fig. 6.14. Computed contours of ame temperature, soot volume fraction, and species
mass fractions of CO
2
, H
2
O, and NO.
148
6.6.2.3 NO
x
emissions
The lower row of Fig. 6.14 shows four computed contours of NO mass frac-
tion using the two radiation models with and without soot radiation. The predicted
NO
x
emission indices are tabulated in Table 6.7 together with the corresponding global
radiant fractions. The eect of soot radiation on NO
x
formation is straightforward: soot
radiation greatly decreases the ame temperature and therefore NO
x
emission. The P
1
-
Gray model (with and without soot radiation), however, predicts a greater radiative
heat loss and a higher NO
x
emission index than the P
1
-FSK model. This can be under-
stood from the heat ux curves in Fig. 6.13. The larger radiative heat loss predicted
by the P
1
-Gray model occurs far downstream in the ame, while NO is mainly formed
in the upstream ame zone and its peak values follow peak ame temperature (see NO
contours). The extreme sensitivity of NO formation to ame temperature justies the
use of a detailed radiation model in simulations of oxygen-enriched ames. Remaining
discrepancies between experiments and computations result from over-predicted ame
temperature (simple turbulent combustion model), neglect of the eects of turbulent
uctuations in temperature and species concentrations, and the P
1
approximation used
to solve the RTE.
Table 6.7. Comparisons of global quantities radiation models.
Radiation model EINO
x
R
P
1
-Gray 54.2 0.37
P
1
-FSK 33.8 0.31
P
1
-Gray-Soot 13.6 0.50
P
1
-FSK-Soot 10.7 0.43
Experiment 29.8 0.24
149
6.7 Summary
Motivated by TSLs failing to provide quantitative predictions of ame structure
due to its simplistic treatment of the hydrodynamics of the turbulent low eld, a CFD
model of uid dynamics was extended to serve as a comprehensive combustion simulation
tool by incorporating a detailed chemistry model, a detailed soot model, and a detailed
radiation model. The detailed chemistry model is based on the use of the CHEMKIN
libraries and the calculations of chemistry is accelerated by using the ISAT software.
The eective use of ISAT for detailed chemistry in nonhomogeneous systems is discussed
and outlined. The detailed soot model is adopted from Frenklachs detailed soot model
with the method of moments. It is coupled to the three-dimensional CFD code through
transport equations of soot moments. The implementation is described and validated.
Two detailed radiation models are implemented, the P
1
-gray model and the P
1
-FSK
model. Both models employ the spherical harmonic P
1
method for the solution of the
radiative transfer equation on three-dimensional unstructured meshes. The P
1
-gray
model employs the gray medium assumption and Planck-mean absorption coecient
for radiative property evaluations. The P
1
-FSK model addresses the nongray nature
of the radiative heat transfer by using the full-spectrum k-distribution method. Their
implementations into the CFD code are described and validated. By using these two
detailed radiation models, several issues regarding radiation calculations in turbulent
ames, such as gray versus nongray radiation, soot radiation, and the applicability of the
P
1
method, are discussed. With the detailed chemistry model, the detailed soot model,
and the detailed radiation models, the modied GMTEC code becomes a comprehensive
simulation tool for gas-phase turbulent ames. It is exercised to simulated oxygen-
enriched ames, which is an ideal target for testing a comprehensive combustion model.
The following conclusions can be drawn from the simulation of a propane-fueled oxygen-
enriched turbulent nonpremixed jet ame:
The prediction of soot is extremely sensitive to the soot surface growth model
(surface-growth mechanism and steric factor in particular). Soot formation is
closely coupled with ame temperature through soot radiation.
150
The presence of soot shortens the ame length and changes the ame structure.
Soot is distributed in the fuel-rich regions upstream of the nominal ame zone,
while spectrally radiating gas-phase species are found within and downstream of
the ame zone. This separation of radiating media has great impact on the ame
structure, temperature, and NO
x
emissions.
Soot radiation decreases ame temperature and NO
x
emission substantially, espe-
cially in the ame-tip region.
The eects of nongray gas-phase radiation are important even in the presence of
strong gray soot radiation, and must be included to capture the correct distribution
of radiative heat loss.
Nongray gas eects are stronger downstream than upstream and, therefore, are
particularly important for the burn-out of soot.
Both soot and nongray gas radiation are required for accurate predictions of soot
and NO
x
formation within and upstream of the ame zone.
Remaining discrepancies between experiment and model are attributed to the sim-
ple turbulent combustion model, neglect of turbulent uctuations in composition
and temperature. and the P
1
approximation used to solve the RTE. These are the
subjects of ongoing research.
151
Chapter 7
Summary and Conclusions
7.1 Summary
In general, the present work aims at a comprehensive approach for the simula-
tions of turbulent reacting ows. In particular, it focuses on the modeling of detailed
chemistry, detailed soot formation and oxidation, and modeling of detailed radiative
heat transfer in gas-phase turbulent ames. In addition, the present work centers on
numerical investigations of oxygen-enriched turbulent nonpremixed ames.
At present, combustion simulation tools remain limited in their capacity to deal
simultaneously with all the complex interacting physical-chemical phenomena involved
in turbulent reacting ows, such as turbulent transport, nite-rate chemical reactions,
radiative heat transfer, and particulate behavior. For nite-rate chemistry, the common
practice is the use of reduction techniques to simplify chemistry. For radiative heat
transfer, the common practice is the use of the optically thin model with the gray medium
assumption. And for soot formation, the common practice is the use of various empirical
or semi-empirical models. However, detailed chemistry is required to address issues such
as pollutant emissions, extinction and ignition phenomena, and unsteady phenomena
including combustion instabilities. Radiation needs to be treated accurately due to its
dominance in ames and its inuence on ame structures, temperatures, and pollutant
emissions. Detailed soot models are desired to increase the accuracy and the range of
applicability of soot calculations and to address the close coupling between soot and
radiation.
With the continuous improvements in physical understanding, numerical meth-
ods, and computer capabilities, turbulent combustion simulations become more and more
152
sophisticated. In this research, two comprehensive combustion simulation tools are de-
veloped, one based on an empirical description of the turbulent ow eld and the other
based on CFD modeling of the ow eld.
The empirical-description-based model is simple but computationally ecient.
It is an extension of the Two-Stage Lagrangian model of turbulent jet ames. The
extension includes the incorporation of a detailed soot model and the improvement of
the radiation model. The soot model is a detailed one adopted from Appel-Bockhorn-
Frenklachs soot model and the dynamics of soot particles are described by the method
of moments adapted to the TSL formulation. The original constant-emissivity radiation
model is improved by solving the radiative transfer equation on the spatial conguration
of the TSL model using the spherical harmonic P
1
method and the discrete ordinate
S
2
method. The gray medium assumption is employed and the Planck-mean absorption
coecient is used to determine the radiative properties of both gas-phase species and
soot particles. With the comprehensive TSL model, the characteristics of soot, radiation
and NO
x
emissions in oxygen-enriched ames are studied.
For a more realistic description of the ow eld and a more productive numerical
simulation, a CFD-based approach has to be employed. The CFD-based comprehensive
model is based on an engineering CFD code (GMTEC) and it solves the compressible
ow equations on unstructured meshes. GMTEC is extended by incorporating a detailed
chemistry model, a detailed soot model, and a detailed radiation model. The detailed
chemistry model is based on the use of the CHEMKIN libraries and the calculations
of chemistry is accelerated by using the ISAT software. The eective use of ISAT for
detailed chemistry in nonhomogeneous systems is outlined. The detailed soot model is
adopted from Frenklachs detailed soot model with the method of moments. It is coupled
to the three-dimensional CFD code through transport equations of soot moments. Two
detailed radiation models are implemented, the P
1
-gray model and the P
1
-FSK model.
Both models employs the spherical harmonic P
1
method for the solution of radiative
transfer equation on three-dimensional unstructured meshes. The P
1
-gray model em-
ploys the gray medium assumption and Planck-mean absorption coecient for radiative
153
property evaluations. The P
1
-FSK model addresses the non-gray nature of the radia-
tive heat transfer by using the full-spectrum k-distribution method. The CFD-based
comprehensive model is then exercised to simulated an oxygen-enriched ame.
7.2 Conclusions
The two comprehensive models have proven to be successful in the simulation of
oxygen-enriched turbulent ames. The advantage of the TSL model is its computational
economy, and the advantage of the CFD-based model is its capability of performing
detailed, quantitative predictions.
The TSL model is capable of general trend predictions. The following conclusions
can be drawn from the simulation of a set of oxygen-enriched turbulent nonpremixed jet
ames:
The TSL model is capable of predicting the general trends of pollutant emission,
soot formation, and radiative heat transfer in turbulent jet non-premixed ames
as functions of oxygen index, fuel type, and initial jet velocity.
The TSL model can help to understand and explain results obtained from experi-
ments.
The TSL model can provide information that cannot be measured in the experi-
ments.
The TSL model can provide guidance for conducting further experiments.
The TSL model fails to provide quantitative predictions of ame structure due to
its simplistic treatment of the hydrodynamics of ow eld.
The CFD-based model is capable of capturing the strong coupling among soot,
radiation, ame structure, and NO
x
emission in oxygen-enriched ames. It can be
used to identify the key sensitivities in soot and NO
x
formations, to study the eects of
nongray gas-phase and soot radiation, and to study the inuence of mixing, fuel type, and
oxygen index on the soot formation, NO
x
emission, and thermal radiation characteristics
154
of oxygen-enriched turbulent ames. The following conclusions can be drawn from the
simulation of a propane-fueled oxygen-enriched turbulent nonpremixed jet ame:
The prediction of soot is extremely sensitive to the soot surface growth model
(surface-growth mechanism and steric factor in particular). Soot formation is
closely coupled with ame temperature through soot radiation.
The presence of soot shortens the ame length and changes the ame structure.
Soot is distributed in the fuel-rich regions upstream of the nominal ame zone,
while spectrally radiating gas-phase species are found within and downstream of
the ame zone. This separation of radiating media has great impact on the ame
structure, temperature, and NO
x
emissions.
Soot radiation decreases ame temperature and NO
x
emission substantially, espe-
cially in the ame-tip region.
The eects of nongray gas-phase radiation are important even in the presence of
strong gray soot radiation, and must be included to capture the correct distribution
of radiative heat loss.
Nongray gas eects are stronger downstream than upstream and, therefore, are
particularly important for the burn-out of soot.
Both soot and nongray gas radiation are required for accurate predictions of soot
and NO
x
formation within and upstream of the ame zone.
Remaining discrepancies between experiment and model are attributed to the sim-
ple turbulent combustion model, neglect of turbulent uctuations in composition
and temperature. and the P
1
approximation used to solve the RTE. These are the
subjects of ongoing research.
155
7.3 Future Work
The present work did not consider turbulence-chemistry-radiation interactions in
a rigorous way and this is a limitation of the present work. A simple turbulent com-
bustion model, a characteristic time-scale model, was used to provide a slightly better
modeling of mean reaction rates than just using the mean quantities (temperature and
species mass fractions) directly for the valuations of mean reaction rates. The eects
of turbulence uctuations are not considered and therefore the turbulence-chemistry
interactions are not accounted for rigorously. In addition, the turbulence-radiation in-
teractions are ignored totally. Since PDF methods are capable of treating exactly the
nonlinear processes such as chemical reaction and radiative emission, a transported joint
composition PDF method may be implemented to remedy this shortcoming. In this ap-
proach, a particle-tracking-based Monte Carlo method is employed for the solution of a
modeled PDF transport equation on an unstructured mesh. GMTECs Favre averaged-
ow solver provides the mean ow eld information used by the mixing models. This
hybrid nite-volume/PDF Monte Carlo method is considered one of the most promising
method for the evaluations of the mean species source terms and the mean radiation loss
term in the energy equation [199].
The detailed soot model used in the present work employs concentration moments
of soot particle size distribution functions (PSDF): this method of moments replaces the
innite number of dierential equations to be solved for the PSDF by a small number
of equations for the corresponding moments. It allows the distribution of soot particles
to be computed using essentially the same approach that is used for gas-phase chemical
species. In the present study, the source terms for moment equations are determined from
mean temperature and concentration elds and an interpolative scheme is employed for
moment closure. Stochastic particle models for soot particle growth and oxidation can be
developed based on the current understanding of the physics of soot formation/oxidation.
The composition PDF method can be used as the framework for the closure of the
moment equations, where in addition to the scalars (reactants and temperature), several
moments of soot distributions may be carried. To date, few researchers have employed
moment transport equations for soot calculations in turbulent ames.
156
The present work on soot calculations did not consider the formation of aggregated
or agglomerated soot particles. Soot aggregates make the consideration of radiation
scattering necessary. This involves the evaluation of the scattering properties of large soot
particles. The Rayleigh-Debye-Gans (RDG) approximation may be used to determine
the total scattering coecient in the aggregate regime based on the information from
soot moment calculations. Work can be done in the following areas regarding radiation
calculations: rst, evaluations of the radiative properties of agglomerated soot particles;
second, sensitivity studies to evaluate how much radiation contribution comes from large
agglomerated soot; and third, assessment of dierent radiation models on accuracy and
eciency, on feedback to ame, soot production, and NO
x
emissions, and so on.
Chemical source term acceleration is essential for detailed chemistry calculations
in turbulent ames, since the numerical integration of chemical source terms dominates
the computational eorts. The in situ adaptive tabulation software is implemented in
the present work for chemistry acceleration. The use of ISAT with large mechanisms
has been explored in the current study and measures have been taken to accommodate
mechanisms with more than a hundred species. An alternative approach to ISAT to
achieve acceleration with arbitrarily complex chemistry may be developed based on the
rate-controlled constrained-equilibrium (RCCE) method. The basic idea of RCCE is that
the evolution of a complex chemical system can be described with acceptable accuracy
by a relatively small number of rate-controlling reactions that impose slowly changing
constraints to those assumed intermediate equilibrium states due to fast reactions. It
would be worthwhile to explore the possibility of using RCCE as a more accurate and
ecient approach to large chemistry accelerations.
The next logical step of the current research would be to go beyond the Reynolds-
averaged Navier-Stokes (RANS) modeling, in which all uctuations about the mean must
be modeled, and move to Large Eddy Simulation (LES) of turbulent reacting ows, in
which the large scales (super-lter-scale) are computed explicitly while the eects of
unresolved (sub-lter-scale) uctuations remain to be modeled. It would be appropriate
to adapt the best available RANS approaches such as PDF methods as subgrid-scale
models for chemistry, radiation, and turbulence/chemistry/radiation interactions in LES.
157
Signicant improvements can be realized because a portion of the velocity and scalar eld
uctuations is explicitly resolved. The hybrid nite-volume/PDF Monte Carlo method,
which combines the nite volume methods eciency and robustness in solving ow
elds and the PDF methods exactness in dealing with chemical reactions and thermal
radiation, can provide the starting point for the study of subgrid-scale modeling in LES.
158
References
[1] J.C. Tannehill, D.A. Anderson, and R.H. Pletcher. Computational Fluid Mechanics
and Heat Transfer. Taylor & Francis, second edition, 1997.
[2] S. Candel, D. Thevenin, Darabiha N., and Veynante D. Progress in numerical
combustion. Combust. Sci. and Tech., 149:297337, 1999.
[3] D.C. Haworth. Applications of turbulent combustion modelling. In Turbulent
Combuston, Von Karman Institute Lecture Series. Von Karman Institute for Fluid
Dynamics, Rhode-Saint-Genese, Belgium, 1721 March 2003.
[4] A.M. Eaton, L.D. Smoot, S.C. Hill, and C.N. Eatough. Components, formulations,
solutions, evaluation, and application of comprehensive combustion models. Prog.
Energy Combust. Sci., 25:387436, 1999.
[5] C.E. Baukal. Oxygen-Enhanced Combustion. CRC Press, Boca Raton, 1998.
[6] P.B. Eleazer and A.G. Slavejkov. Clean ring of glass furnaces through the use of
oxygen. 54th Glass Problems Conference, October 26-27 1993.
[7] A.G. Slavejkov et al. Method and device for low-NO
x
high eciency heating in
high temperature furnaces. United States Patent No. 5,575,637, 1996.
[8] W.J. Horan, A.G. Slavejkov, and L.L. Chang. Heat transfer optimization in TV
glass furnaces. Ceramic Engineering Science Proceedings, 17(2):141151, 1996.
[9] S.R. Turns. An Introduction to Combustion: Concepts and Applications. McGraw-
Hill, New York, second edition, 2000.
[10] R.S. Barlow. website. http://www.ca.sandia.gov/TNF/, accessed on January 1,
2004.
[11] R.S. Barlow and J.H. Frank. Eects of turbulence on species mass fractions in
methane/air jet ames. Proc. Combust. Institute, 27:10871095, 1998.
[12] N.E. Endrud. Soot, radiation, and pollutant emissions in oxygen-enriched turbulent
jet ames. Masters thesis, The Pennsylvania State University, 2000.
[13] A.E. Lutz and J.E. Broadwell. TSL: Two-stage lagrangian model for mixing and
reactions in turbulent jet. Technical Report GRI-97/0367, Gas Research Institute,
Chicago, IL., 1997.
[14] S. Subramaniam and D.C. Haworth. A pdf method turbulent mixing and com-
bustion on three-dimensional unstructured deforming meshes. Internl. J. Engine
Research, 1:171190, 2000.
159
[15] J.E. Broadwell and A.E. Lutz. A turbulent jet chemical reaction model: No
x
production in jet ames. Combust. and Flame, 114:319335, 1998.
[16] K.K. Kuo. Principles of Combustion. John Wiley & Sons, New York, 1986.
[17] P.A. Libby and F.A. Williams, editors. Turbulent Reacting Flow. Academic Press,
1994.
[18] R.J. Kee, F.M. Rupley, and J.A. Miller. The chemkin thermodynamic data base.
Technical Report Report SAND87-8251 B, Sandia National Laboratories, March
1987.
[19] D.R. Stull and H. Prophet. Janaf thermochemical table. Technical Report Second
edition, NSRDS-NBS 37, National Bureau of Standards, June 1971. The third
edition is available from NIST.
[20] S. Gordon and B.J. McBride. Computer program for calculation of complex chem-
ical equilibrium compositions, rocket performance, incident and reected shocks,
and chapman-jouguet detonations. Technical Report NASA SP-273, NASA, 1976.
[21] R.J. Kee, F.M. Rupley, and J.A. Miller. Chemkin-II: A fortran chemical kinetics
package for the analysis of gas-phase chemical kinetics. Technical Report Report
SAND89-8009, Sandia National Laboratories, March 1991.
[22] G.P. Smith, D.M. Golden, Frenklach M., N.W. Moriarty, B. Eiteneer, M. Golden-
berg, C.T. Bowman, R.K. Hanson, S. Song, W.C. Gardiner Jr., V.V. Lissianski,
and Z. Qin. website. http://www.me.berkeley.edu/gri mech/, accessed on January
1, 2004.
[23] C.K. Westbrook. Topical review: chemical kinetics of hydrocarbon ignition in
practical combustion systems. Proc. Combust. Institute, 28:15631577, 2000.
[24] P.H. Renard, D. Thevenin, J.C. Rolon, and S. Candel. Dynamics of ame/vortex
interactions. Prog. Energy and Combust. Sci., 26(3):225282, 2000.
[25] A.A. Amsden. Kiva-3V: a block-structured kiva program for engines with vertical
or canted valves. Technical Report Report LA-13313-MS, Los Alamos National
Laboratory, Los Alamos, NM, 1997.
[26] B. Khalighi, S.H. El Tahry, D.C. Haworth, and M.S. Huebler. Computation and
measurement of ow and combustion in a four valve engine with intake variations.
SAE Technical Paper, No. 950287, 1995.
[27] D. Yosse, M.R. Belmont, S.J. Maskell, and S.J. Ashcroft. A comparison of the
relative eects of fuel composition and ignition energy on the early stage of com-
bustion in a natrual gas spark ignition using simulation. J. Automobile Engineering
(Proc. I. Mech. E. Part D), 214(D4):383393, 2000.
160
[28] A.D. Harvey and H. Pitsch. Modeling turbulent reacting methane thermochlori-
nation ows. Proceedings of the 2000 Summer Program, pages 181191, 2000.
[29] S.M. Correa. Power generation and aeropropulsion gas turbines: from combustion
science to combustion technology. Proc. Combust. Institute, 27:17931870, 1998.
[30] Smooke M.D., Mitchell R.E., and Keyes D.E. Numerical-solution of two-
dimensional axisymmetric laminar diusion ames. Combust. Sci. Techno., 67(4
6):85122, 1989.
[31] C.J. Montgomery, M.A. Cremer, J-Y Chen, C.K. Westbrook, and L. Maurice.
Reduced chemical kinetic mechanisms for hydrocarbon feuls. J. Prop. Power,
18(1):192198, 2002.
[32] J. Warnatz, U. Mass, and R.W. Dibble. Combustion: Physical and Chemical Fun-
damentals, Modeling and Simulaiton, Experiments, Pollutant Formation. Springer-
Verlag, Berlin Heidelberg, third edition, 2001.
[33] I.S. Wichman. On the use of operator-splitting methods for the equations of com-
bustion. Combust. and Flame, 83:240252, 1991.
[34] B. Yang and S.B. Pope. An investigation of the accuracy of manifold methods and
splitting schemes in the computational implementation of combustion chemistry.
Combust. and Flame, 112:1632, 1998.
[35] M.G. Mungal. Experiments in combustion. In Turbulent Combuston, Von Karman
Institute Lecture Series. Von Karman Institute for Fluid Dynamics, Rhode-Saint-
Genese, Belgium, 1721 March 2003.
[36] Seshadri K. and F. A. Williams. Reduced chemical systems and their applications
in turbulent combustion. In P. A. Libby and F. A. Williams, editors, Turbulent
Reacting Flows. Acdemic Press, 1994.
[37] J. C. Keck. Rate-controled constrained-equilibrium theory of chemical reactions
in complex systems. Prog. Energy Combust. Sci., 16:125154, 1990.
[38] D. Hamiroune, P. Bishnu, M. Metghalchi, and J. C. Keck. Rate-controlled
constrained-equilibrium method using constraint potentials. Combust. Theory &
Modelling, 2:8194, 1998.
[39] Lam S. H. and Goussis D. A. Understanding complex chemical kinetics with
computational singualr perturbation. Proc. Combust. Institute, 22:931, 1988.
[40] A. Massias, D. Diamantis, E. Mastorakos, and Goussis D. A. An algorithm for
the construction of global reduced mechanisms with csp data. Combust. Flame,
117:685708, 1999.
[41] U. Maas and Pope S. B. Simpling chemical kinetics: Intrinsic low-dimensional
manifolds in composite space. Combust. Flame, 83:239264, 1992.
161
[42] U. Maas and Pope S. B. Implementation of simplied chemical kinetics based on
intrinsic low-dimensional manifolds. Proc. Combust. Institute, 24:103, 1992.
[43] U. Maas. Ecient calculaton of instrinsic low-dimensional manifolds for simpli-
cation of chemical kinetics. Comp. and Visual. in Sci., 1:69, 1998.
[44] M.D. Smooke, editor. Reduced Kinetic Mechanisms and Asymptotic for Methane-
Air Flames. Lecture Notes in Physics 384. Springer-Verlag, Berlin Heidelberg,
1991.
[45] N. Peters and B. Rogg. Reduced Kinetic Mechanisms for Application in Combus-
tion. Lecture Notes in Physics Monograph 15. Springer-Verlag, Berlin Heidelberg,
1993.
[46] J.-Y. Chen. A general procedure for constructing reduced reaction mechanisms
with given independent relations. Combust. Sci. Technol., 57(13):8994, 1988.
[47] J.-Y. Chen. Development of reduced mechanisms for numerical simulation of tur-
bulent combustion. In Workshop on Numerical Aspects of Reduction in Chemical
Kinetics. Centre dEnseignement et de Recherche en Mathematiques, Informatique
et Calcul Scientique, Ecole Nationale des Ponts et Chaussees, September 1997.
[48] S.B. Pope. Computationally ecient implementation of combustion chemistry
using in situ adaptive tabulation. Combust. Theory & Modelling, 1(1):4163, March
1997.
[49] J.-Y. Chen, W. Kollmann, and R.W. Dibble. Pdf modelling of turbulent non-
premixed methane jet ames. Combust. Sci. Technol., 64(46):315346, 1989.
[50] I. Veljkovic, P.E. Plassmann, and D.C. Haworth. A scientic on-line database
for ecient function approximation. Submitted to 2003 International Conference
on Computational Science, Sanit Petersburg, Russian Federation and Melbourne,
Australia, 2-4 June, 2003, 2003.
[51] F.C. Christo, A.R. Masri, E.M. Nebot, and S.B. Pope. An integerated pdf/neural
network approach for simulating turbulent reacting systems. Proc. Combust. In-
stitute, 26:4348, 1996.
[52] J.-Y. Chen, J.A. Blasco, N. Fueyo, and C. Dopazo. An economical strategy for
storage of chemical kinetics: Fitting in situ adaptive tabulation with articial
neural networks. Proc. Combust. Institute, 28:115121, 2000.
[53] T. Turanyi. Application of repro-modelling for the reduction of combustion mech-
anisms. Proc. Combust. Institute, 25:948, 1994.
[54] S.R. Tonse, N.M. Moriarty, N.J. Brown, and M. Frenklach. Prism: Piecewise
reusable implementation of solution mapping: An economical strategy for chemical
kinetics. Israel J. Chem., 39:97, 1999.
162
[55] J.B. Bell, N.J. Brown, M.S. Day, J.F. Frenklach, M. Grcar, R.M. Propp, and
S.R. Tonse. Scaling and eciency of prism in adaptive simulations of turbulent
premixed ames. Proc. Combust. Institute, 28:107113, 2000.
[56] M. Embouazza, D.C. Haworth, and N. Darabiha. Implementation of detailed chem-
ical mechanism into multidimentional cfd using in situ adaptive tabulation: Ap-
plication to hcci engines. SAE Technical Paper, 2002-01-2773, 2002.
[57] P.N. Brown, G.D. Byrne, and A.C. Hindmarsh. Vode: A variable-coecient ode
solver. SIAM J. Sci. Stat. Comput., 10:10381051, 1989.
[58] D.E. Knuth. The Art of Computer Programming, volume I: Fundamental Algror-
ithms. Addison-Wesley, third edition, 1997.
[59] Q. Tang, J. Xu, and S.B. Pope. Pdf calculations of local extinction and NO
production in piloted-jet turbulent methane/air ames. Proc. Combust. Institute,
28:133139, 2000.
[60] V. Saxena and S.B. Pope. Pdf simulations of turbulent combustion incorporating
detailed chemistry. Combust. Flame, 117:340350, 1999.
[61] R.J. Blint and D.C. Haworth. Modelling oxidation of exhaust gases diluted by air
under exhaust manifold conditons. Proc. Combust. Institute, 28:24512457, 2000.
[62] S.B. Pope. ISAT-CK (Version 3.0) Users Guide and Reference Manual. Ithaca
Combustion Enterprise, LLC, Ithaca, NY, 2000. http://www.IthaComb.com.
[63] D. Veynante and L. Vervisch. Turbulent combustion modeling. Prog. Energy
Combust. Sci., 28:193266, 2002.
[64] K.N.C. Bray. Challenges in turbulent combustion. In Turbulent Combuston,
Von Karman Institute Lecture Series. Von Karman Institute for Fluid Dynam-
ics, Rhode-Saint-Genese, Belgium, 19-23 March 2001.
[65] D.C. Haworth and T.J. Poinsot. Numerical simulation of lewis number eect in
turbulent premixed ames. J. Fluid Mech., 244:405436, 1992.
[66] L. Vervisch and T.J. Poinsot. Direct numerical simulations of non-premixed tur-
bulent ames. Annu. Rev. Fluid Mech., 30:655691, 1998.
[67] S.B. Pope. Turbulent Flows. Cambridge University Press, Cambridge, 2000.
[68] N. Peters. Turbulent Combustion. Cambridge University Press, Cambridge, 2000.
[69] I.G. Sheperd, J.B. Moss, and K.N.C. Bray. Turbulent transport in conned pre-
mixed ames. Proc. Combust. Institute, 19:423431, 1982.
[70] D.B. Spalding. Mixing and chemical reaction in steady conned turbulent ames.
Proc. Combust. Institution, 13:649657, 1971.
163
[71] B.F. Magnussen and B.H. Hjertager. On mathematical models of turbulent com-
bustioin with special emphasis on soot formation and combustion. Proc. Combust.
Institute, 16:719729, 1976.
[72] N. Peters. Laminar amelet concepts in turbulent combustion. Proc. Combust.
Institute, 21:12311250, 1986.
[73] K.N.C. Bray, P.A. Libby, G. Masuya, and J.B. Moss. Turbulence production in
premixed turbulent ames. Combust. Sci. and Tech., 25:127140, 1981.
[74] K.N.C. Bray, P.A. Libby, and J.B. Moss. Unied modeling approach for premixed
turbulent combustionpart 1: General formulation. Combust. Flame, 61:87102,
1985.
[75] F.E. Marble and J.E. Broadwell. The coherent ame model for turbulent chemical
reactions. Technical Report TRW-9-PU, 1977. Project Squid, Tech. Rep.
[76] J.M. Duclos, D. Veynante, and T.J. Poinsot. A comparison of amelet models for
premixed turbulent combustion. Combust. Flame, 95:101117, 1993.
[77] S.B. Pope. Pdf methods for turbulent reacting ows. Prog. Energy Combust. Sci.,
11:119192, 1985.
[78] M. Muradoglu, P. Jenny, S.B. Pope, and D.A. Caughey. A consistant hybrid
nite volume/particle method for the pdf equations of turbulent reactive ows. J.
Comput. Phys., 154:342371, 1999.
[79] A.Y. Klimenko. Multicomponent diusion of various scalars in turbulent ows.
Fluid Dyn., 25:327334, 1990.
[80] R.W. Bilger. Conditional moment closure for turbulent reacting ows. Phys. Fluid,
A5:436444, 1993.
[81] A.Y. Klimenko and R.W. Bilger. Conditional moment closure for turbulent com-
bustion. Prog. Energy Combust. Sci., 25:595687, 1999.
[82] A.R. Kerstein. Linear-eddy modeling of turbulent transport, part 7: Finite-rate
chemistry and multi-stream mixing. J. Fluid. Mech., 240:289313, 1992.
[83] V. Sankaran and S. Menon. Structure of premixed turbulent ames in the thin-
reaction-zones regime. Proc. Combust. Institute, 28:203209, 2000.
[84] W.C. Reynolds. The potential and limitations of direct and large eddy simulation.
In Whither Turbulence, Turbulence at Crossroads, Lecture Notes in Physics, page
313. Springer Verlag, New York, 1989.
[85] In D. Olivari, editor, Advances in Turbulence Modeling, Von Karman Institute
Lecture Series. Von Karman Institute for Fluid Dynamics, Rhode-Saint-Genese,
Belgium, 2327 March 1998.
164
[86] G.M. Faeth and
U.
O. K oyl u, G.M. Faeth, T.L. Farias, and M.G. Carvalho. Fractal and projected
structure properties of soot aggregates. Combust. and Flame, 100:621633, 1995.
[141]
U.
O. K oyl u and G.M. Faeth. Structure of overre soot in buoyant turbulent dif-
fusion ames at long residence times. Combust. and Flame, 89:140, 1992.
[142]
U.
O. K oyl u and G.M. Faeth. Optical properties of overre soot in bouyant tur-
bulent diusion ame at long residence time. ASME J. Heat Trans., 116:152159,
1994.
[152] T.L. Farias, M.G. Carvalho,
U.
b
I k I s
ds
dI
= =
r
,
where, I
is the radiative intensity. The subscript b represents blackbody, and wave number.
The integration of the above equation over all solid angle gives the divergence of the
radiative heat flux, which is the net loss of radiative energy per unit time per unit volume:
b
I k q = 4
r
r
.
Integration over the spectrum gives the total radiation loss:
=
0
4
d I k q
b
r
r
.
The integral in the above equation is hard to carry out because of the irregular
dependence of absorption coefficient on wave number. For convenience, the Planck-mean
absorption coefficient is introduced [88]:
=
0
4
0
0
d I k
T
d I
d k I
k
b
b
b
P
,
and the total energy loss becomes:
4
4 T k q
P
=
r
r
185
Here,
P
k is the linear Planck-mean absorption coefficient with units of 1/m; and is the
Stephan-Boltzmann constant and equals 5.67x10
-8
W/m
2
K
4
. The units associated with the total
energy loss are:
[ ]
3
4
4 2
1
m
W
K
K m
W
m
q =
=
r
In the original TSL code, the energy equation has the dimensions of energy per unit time
per unit mass of the reactor. Therefore, dividing the above equation by the mass density of the
reactor gives the radiation loss term in the energy equation. This formulation of the radiation
term is more rigorous and more physically based than the original one for both the core and
flame-sheet reactors.
P
1
-S
2
Model
Core Reactor
Black
Ts=300K
=1
T1
T2
T1=T
H
T2=T
fs
d
dx
F-S reactor
Core Reactor
Black
Ts=300K
=1
T1
T2
T1=T
H
T2=T
fs
d
dx
F-S reactor
Figure B1 The core and flame-sheet reactor of the radiation model
186
The core reactor in the TSL model is treated as a segment (of length dx) of an infinitely
long cylinder of homogeneous, isothermal combustion gases. This long cylinder is surrounded by
a cylindrical shell of other homogeneous gases at the temperature and with the thickness of the
flame-sheet reactor. Bounding the cylindrical shell, the cold surroundings of the jet flame forms
a black wall at a temperature of 300 K. This configuration is sketched in Figure B1. We will
treat the radiation in the core and flame-sheet reactors separately and focus on the calculations of
the radiative heat flux, the values of which on the boundaries give the radiative heat loss terms
used in the energy conservation equations in the TSL code.
Homogeneous (Core) reactors
For the core reactor, we apply the P
1
or differential approximation to model the radiation
field within. We start here with the general formulation for non-scattering media [88]:
G I q
b
=
4
r
r
q G
r
r
= 3
where is the radiative heat flux and is given by q
v
=
4
) , ( d s s r I q
v v
,
and where is the radiative intensity at spatial location ) , ( s r I
v
r
v
and along the direction . The
position vector
s
r
v
is determined by (r,
c
, x) in cylindrical coordinates and by polar angle and
azimuthal angle with respect to the local coordinates. The term I
s
b
is black body intensity and G
is known as incident radiation function, is defined by
=
4
) , ( d s r I G
v
,
and expresses the total intensity impinging on a point from all directions.
187
The subscript indicates that the space coordinates have been non-dimensionalized by
the extinction coefficient =(T), i.e., d =d, and is the absorption coefficient and is a
function of temperature in this problem.
For a cylindrical system (r,
c
, x) and assuming that the temperature varies only in the
radial direction, the governing equations of the P
1
approximation reduce to
G I
d
q d
b
=
4
) ( 1
, (B1a)
q
d
dG
= 3
. (B1b)
We now define the following non-dimensional variables. The optical thickness, , is
already in non-dimensional form,
[ ]
1
= = m dr d
4
4
4 4 4
4
s s
b
b
s s
T
T
T
E
e
T
G
g
T
q
=
=
=
=
where T
s
is the cold surroundings temperature and set to 300 K in the calculation. Equations
(B1a) and (B1b) become
) ( 4
) ( 1
g e
d
d
b
=
(B2a)
=
4
3
d
dg
(B2b)
Eliminating g, equations (B2a) and (B2b) combine to yield
d
de
d
d
d
d
b
= +
4 )
1
3 (
1
2 2
2
(B3a)
Since this is a second-order ordinary differential equation of the function , we need two
boundary conditions on to complete the formulation. One of the boundary condition comes
188
from the symmetry of the problem, that is, at the center of the cylinder, the radiative heat flux
equals zero:
at 0 , 0 = = (B3b)
Finding the second boundary condition poses some difficulty. It is tempting to assume
that the cylinder is bounded by a black or a gray diffuse wall at known temperature as usually
done in most applications. Although the radiation originating from inside the cylinder impinges
onto the flame sheet reactor shell and penetrates it without any reflection, which suggests a black
body nature, it would be inappropriate to assume that the homogeneous cylinder is bounded by a
black wall at the temperature of the flame sheet reactor. Moreover, because of the discontinuities
of temperature and absorption coefficient at the interface between the two reactors, it is hard to
treat the two reactors together as one cylinder.
To find the second boundary condition, imagine that there is a hypothetical surface, or a
wall, surrounding the cylinder of the homogeneous reactor. The boundary condition at this
surface is
at , (B3c)
>
= =
0
) ( 4 2 , 2 /
s n
w
d n s s I q G d r
where is the radiative intensity at the surface along the direction. With the hypothetical
surface being not a black or gray wall, we have to figure out the distribution of intensity at
this surface to obtain the second boundary condition.
) (s I
w
s
) (s I
w
Note that the intensity is the radiation emerging from the hypothetical surface
directed toward the inside of the cylinder. This radiation comes from both the flame sheet reactor
and the cold surroundings. Since the thickness of the flame sheet reactor is rather small, we can
treat it as a flat slab instead of a ring surrounding the core reactor. One side of the slab is black
and at T
) (s I
w
s
= 300 K, which represents the cold surroundings. Inside the slab is contained
189
homogeneous, isothermal, and isotropic combustion gases with a constant absorption coefficient,
. The slab has a thickness of and is sketched in Figure B2.
=constant
T=T
2
=T
fs
Black, T=Ts
) , ( s I
r
r
0
=constant
T=T
2
=T
fs
Black, T=Ts
) , ( s I
r
r
0
Figure B2 The second boundary condition for the core reactor
Solving the equation of transfer for in this emitting, absorbing, but non-scattering
medium, we find that
) (s I
w
= e I I I s I
bw b b w
) ( ) (
2 2
,
where
4
s
bw
T
I
= ,
) ( ,
2
4
2
2 fs b
T T
T
I =
=
2
,
and where
2
is the absorption coefficient in the flame sheet reactor.
Substituting into the right hand side of equation (B3c) gives ) (s I
w
w
J q G = 4 2 ,
where
[ ]
+ = ) ( 2
2 bw b bw w
E E E J ,
190
. ,
4
2 2
4
T E T E
b s bw
= =
On non-dimensionalizing and eliminating G, the second boundary condition for equation
(B3a) turns out to be:
at =
d
=
1
d/2, ) ( 4 )
1
2 ( j e
d
d
b
= + +
(B3d)
where
1
is the absorption coefficient in the core reactor
The problem is now well posed. The governing equation and the boundary conditions are
summarized below:
d
de
d
d
d
d
b
= +
4 )
1
3 (
1
2 2
2
(B3a)
at =0, (B3b) 0 =
at =
d
, ] ) ( [ 4 ) ( )
1
2 ( ) ( j e
d
d
d b d d
= + +
(B3d)
The absorption coefficients in the core and flame sheet reactor are evaluated using
spectrally averaged properties for gas species and soot (Appendix C).
Equation (B3a) can be solved using finite-difference methods, in which the solution
interval (0,
d
) is broken into N equally-spaced subintervals each of size h, such that Nh =
d
.
The grid points are h i
i
= and i =0, 1, 2, , N.
Applying the central difference approximation transforms the governing equation and the
boundary conditions into N+1 linear algebraic equations:
For i =0,
0
0
=
For i =1, 2, , N-1,
) ( 4 )
2
( ) 3 2 ( )
2
(
1
2
1
2 2 2 2 2
1
2
bi bi i i
i
i i i i i
i
i
e e h
h
h h
h
=
+ + + +
+ +
191
For i =N,
[ ]
) ( )
2
( 8
) ( 4 2 ) 1 ( 2 ) 2 4 3 ( ) 2 (
2
1
2 2 2 2
1
2
j e
h
h
e e h h h h h h
bN
d
d
bN bN d N d d N d
+
= + + + + +
The equation set is a tri-diagonal linear algebraic equation system and is diagonally
dominant, so we will have no difficulty in solving this equation set. In the solution array of ,
the element
N
, after multiplied by , gives the radiative heat flux loss from the core
reactor.
4
s
T
Flame Sheet Reactors
We employed the S
2
approximation to calculate the radiation field in the flame sheet
reactor, which is modeled as an absorbing, emitting, but non-scattering medium confined between
two cylinders with radii d/2 and d/2+. For this case, the equation of radiative heat transfer is
given by [88]
2 2 2
sin sin
cos sin
b
I I
I
r r
I
= +
, (B4)
where is the polar angle measured from the x-axis of the cylindrical system (r,
c
, x); is the
azimuthal angle measured from the local radial direction r;
2
is the absorption coefficient in the
flame sheet reactor and assumed to be a constant. I
b2
is the black body intensity at the temperature
of the flame sheet reactor, T
2
=T
fs
.
Applying the discrete ordinates concept to equation (B4), we derived the general
formulation for the cylindrical system using the S
2
approximation as follows:
G I
q
d
dq
b
= +
2
4
(B5a)
192
q
d
dG
= 4
(B5b)
In non-dimensional form, the above equation set becomes:
( g e
d
d
b
= )
2
4
(B6a)
=
d
dg
(B6b)
where
, ,
4
,
4 4
4
4 4
s
b
s
b
s s
T
I
T
T
e
T
G
g
T
q
=
=
. 300 , ,
2
4
4
2
2
K T T T
T
T
e
s fs
s
b
= =
On eliminating g, equations (B6a) and (B6b) combine to form the following governing
equation for the non-dimensional radiative heat flux :
0 4 4
1 1
2
2 2
2
= =
d
de
d
d
d
d
b
(B9)
As this is a second-order ordinary differential equation, we need two boundary
conditions. Note that equation (B9) is the general governing equation for a one-dimensional
cylindrical system and also can be employed in the core reactor with appropriate boundary
conditions.
If we assume that the radiative intensity distributions at the two boundaries are known,
then the boundary conditions for S
2
approximation are:
at ) 2 (
2 1
d = = ,
1
4 2 I q G = + ,
at ) 2 (
2 2
+ = = d ,
2
4 2 I q G = ,
193
where d is the diameter of the core reactor; is the thickness of the flame sheet reactor; and I
1
and I
2
, both with diffuse nature, are the radiative intensities at the two boundaries.
The outer boundary of the flame sheet reactor is approximated by a black wall; therefore,
at =
2
,
4
2
s
bs
T
I I
= = ,
bs
I q G = 4 2 . (B9b)
For the inner boundary condition, we consider again the hypothetical diffuse surface at
the interface between the core and flame sheet reactor. We have calculated the radiative intensity
going towards the core reactor,
= e I I I s I
bw b b w
) ( ) (
2 2
,
and from the heat flux computation in the core reactor, we obtain the net radiative heat
flux at the hypothetical surface,
4
) (
s N
T q =
where
N
comes from the solution array for the core reactor. From the definition of the
radiative heat flux, we can then calculate the radiative intensity going towards the flame sheet
reactor, i.e., I
1
, as
w
J q J I + = = ) (
1 1
,
where
[ ]
+ = ) ( 2
2 bw b bw w
E E E J .
Finally, the boundary condition at =
1
is,
1
4 2 J q G = + . (B9a)
194
Non-dimensionalizing equation (B9a) and (B9b), we obtain the two boundary conditions for
equation (B9):
at
1 1
4 2 4 j g = + = , (B10a)
at 4 2 4
2
= = g , (B10b)
where
+ = + = ) 1 ( 2 1
2 1 b w w N
e j j j
We again employ a finite-difference method to solve equation (B9) with the two
boundary conditions, equation (B10a) and equation (B10b):
For i =1,
) ( ) 2 / ( 8 2 ] 2 ) ( 2 ) 2 4 4 ( [
1 2 1
2
1 2
2
1 1
2
1
2 2
1
2
j e h h h h h h h
b
= + + + + +
For i =2, 3, , N
0 ) 2 / ( )] 2 4 2 ( [ ) 2 / (
1
2 2 2 2
1
2
= + + + + +
+ i i i i i i i i i
h h h h
For i =N+1,
) 1 ( ) 2 / ( 8 ] 2 ) ( 2 ) 2 4 4 [( 2
2 2
2
2 1
2
2
2 2
2
2 2
2
= + + + + + +
+ b N N
e h h h h h h h
where,
) ( ) 1 ( 2 1
1 , , 2 , 1 ) 1 (
) (
) 2 / ( ) 2 / (
1 2 2 1
4
4
2
2
1
1 2
2 2 2 1
+ + =
=
+ = + =
=
+ = =
b N
s
b
i
e j
T
T
e
N i i h
N h
d d
K
195
Appendix C
Planck-Mean Absorption Coefficients in the TSL Model
The determination of the Planck-mean absorption coefficient takes into account the effect
of both gas band and soot particle radiation. For gas band radiation, CO
2
and H
2
O are the most
important radiating species in hydrocarbon flames. The inclusion of CO
2
and H
2
O radiation can
reduce peak temperature by approximately 50 K in a strained laminar flame [10]. CO and CH
4
radiation contribute much less to the flame temperature reduction than do CO
2
and H
2
O.
However, considering all four species in a radiation model is suggested [10].
The TNF website [10] provides curve fits for the pressure-based Planck-mean absorption
coefficients (k
p
) for CO
2
, H
2
O, CO, and CH
4
, based on the results from RADAL program by
Grosshandler of NIST. The following expression is used to calculate k
p
for H
2
O, and CO
2
in units
of (m-1atm-1). These curve fits were generated for temperatures between 300K and 2500K and
may be very inaccurate outside this range.
k
p
= c0 + c1*(1000/T) + c2*(1000/T)
2
+ c3*(1000/T)
3
+ c4*(1000/T)
4
+ c5*(1000/T)
5
The coefficients are:
H
2
O CO
2
c0 -0.23093 18.741
c1 -1.12390 -121.310
c2 9.41530 273.500
c3 -2.99880 -194.050
c4 0.51382 56.310
c5 -1.86840E-05 -5.8169
A fourth-order polynomial in temperature is used for CH4:
k
p,ch4
= 6.6334 - 0.0035686*T + 1.6682e-08*T
2
+ 2.5611e-10*T
3
- 2.6558e-14*T
4
A fit for CO is given in two temperature ranges:
196
K
p,co
= c0+T*(c1 + T*(c2 + T*(c3 + T*c4)))
The coefficients are:
T <= 750 K T > 750 K
c0 4.7869 10.09
c1 -0.06953 -0.01183
c2 2.95775e-4 4.7753e-6
c3 -4.25732e-7 -5.87209e-10
c4 2.02894e-10 -2.5334e-14
Besides the gas species radiation, soot particles also play an important role in
hydrocarbon flame radiation because the luminous emission of the flame comes mainly from soot
particles. Since soot particles are usually small (the averaged soot diameter is about 1x10
-9
m)
compared to the wavelength of the luminous emission (~0.4-0.7 x10
-6
m), if we neglect the
existence of large soot particles and the short wavelength emissions, we can apply Rayleighs
theory to the absorption coefficient evaluation of soot particles [88]:
v
m
f
m
m
I k
= =
6
2
1
2
2
,
where, is the emission wavelength; m is the complex index of refraction, m= n - ik; and f
v
is the
soot volume fraction.
By expanding the complex refraction index mand choosing appropriate spectral averaged
values for the refractive index n and absorption index k, the soot absorption coefficient is
approximated as:
2 2 2 2 0
0
4 ) 2 (
36
k n k n
k n
C
f
C k
v
+ +
=
= =
Substituting the above relation into the definition of the Planck-mean absorption coefficient
gives:
197
2 0
83 . 3 C T C f k
v P
=
where, C
2
is the second Planck function constant and equals 1.4388 cmK.
Since we neglect scattering phenomenon and assume that soot particles are small, the
effect of gas band and soot radiation are additive. Therefore, the summation of the Planck-mean
absorption coefficients over the gas species and soot particles gives the value that is used in the
modified TSL code.
Vita
Liangyu Wang was born in the Spring city of China, Jinan, Shandong Province,
on June 21, 1970. In 1988 he was graduated from Aliated High School of Shandong
Teachers University and went to the oldest university in China (established in 1895),
Tianjin (Peiyang) University. He received the Bachelor of Science in Mechanical Engi-
neering in 1992 and the Master of Science in Mechanical Engineering in 1995. He worked
as a research assistant at the National Engine Combustion Laboratories in China, before
he came to the United States in 1998 and enrolled in the Ph. D. program in Mechanical
Engineering at the Pennsylvania State University. From 1998 to 1999, he worked as a
teaching assistant for the Department of Mechanical and Nuclear Engineering. Since
1999, he has been employed as a research assistant at the Propulsion Engineering Re-
search Center, The Pennsylvania State University.