Fire Dynamics Simulator Technical Manual
Fire Dynamics Simulator Technical Manual
t
+ u = m
b
(3.1)
Conservation of Momentum (Newtons Second Law)
t
(u) + uu+p = g+f
b
+
i j
(3.2)
15
Conservation of Energy (First Law of Thermodynamics)
t
(h) + hu =
Dp
Dt
+ q
b
q
+ (3.3)
Equation of State for a Perfect Gas
p =
R T
W
(3.4)
Note that this is a set of partial differential equations consisting of six equations for six unknowns, all
functions of three spatial dimensions and time: the density , the three components of velocity u = (u, v, w),
the temperature T, and the pressure p. The sensible enthalpy h is a function of the temperature: h =
T
T
0
c
p
(T
)dT
.
The notation adopted above is intended to render the equations in as simple and compact a form as
possible. However, when all of the terms are expanded, the equations appear unwieldy. For the model
developers, this is unavoidable. For the model user, it is sufcient to understand the basic conservation
laws to appreciate what the model does, and does not, represent. Following is a brief description of the
terms. Note that boldfaced quantities represent vectors. Boldfaced quantities with the subscripts i j represent
tensors, which can be best understood as 33 matrices.
The mass conservation equation is often written in terms of the mass fractions of the individual gaseous
species, Y
t
(Y
) + Y
u = D
+ m
+ m
b,
(3.5)
Summing these equations over all species yields the original mass conservation equation because Y
= 1
and m
= 0 and m
b,
= m
b
, by denition, and because it is assumed that D
= 0. This last
assertion is not true, in general. However, transport equations are solved for total mass and all but one of
the species, implying that the diffusion coefcient of the implicit species is chosen so that the sum of all the
diffusive uxes is zero.
In the momentum equation, the term uu is a diadic tensor formed by multiplying (in the sense of two
31 matrices) the vectors u
T
and u. The term uu is thus a vector formed by applying the vector operator
= (
x
,
y
,
z
) to the tensor. The force term f
b
in the momentum equation represents external forces such
as the drag exerted by liquid droplets. The stress tensor
i j
is dened:
i j
=
_
2 S
i j
2
3
i j
( u)
_
;
i j
=
_
1 i = j
0 i = j
; S
i j
=
1
2
_
u
i
x
j
+
u
j
x
i
_
i, j = 1, 2, 3 (3.6)
The term S
i j
is the strain tensor, written using conventional tensor notation. The symbol is the dynamic
viscosity of the uid.
In the energy equation, note the use of the material derivative, D()/Dt =()/t +u (). The term q
is the heat release rate per unit volume from a chemical reaction. The term q
b
is the energy transferred to
the evaporating droplets. The term q
=kT
+ q
r
(3.7)
where k is the thermal conductivity.
The term in the energy equation is known as the dissipation rate. It is the rate at which kinetic energy
is transferred to thermal energy due to the viscosity of the uid:
i j
u =
_
2 S
i j
S
i j
2
3
( u)
2
_
(3.8)
16
=
_
2
_
u
x
_
2
+2
_
v
y
_
2
+2
_
w
z
_
2
+ (3.9)
_
v
x
+
u
y
_
2
+
_
w
y
+
v
z
_
2
+
_
u
z
+
w
x
_
2
2
3
_
u
x
+
v
y
+
w
z
_
2
_
(3.10)
This term is usually neglected because it is very small relative to the heat release rate of the re. However,
it is included here because it plays a role in the discussion of large eddy simulation.
The equations presented here form the basis for a wide variety of engineering applications, but not
without the further application of simplifying assumptions unique to each eld. The only assumptions made
thus far are that the uid is a perfect gas, that the stress is linearly dependent on the strain, that heat is
conducted according to Fouriers Law, and that gas species diffuse according to Ficks Law. In the sections
to follow, additional assumptions will be imposed on the governing equations to apply them to re and
other low speed thermal processes. The most important of these assumptions involves the treatment of the
diffusion and source terms that differentiates one type of CFD model from another.
3.1.2 The Low Mach Number Assumption and the Equation of State
A distinguishing feature of a CFD model is the regime of ow speeds (relative to the speed of sound) for
which it is designed. High speed ow codes involve compressibility effects and shock waves. Low speed
solvers, however, explicitly eliminate compressibity effects that give rise to acoustic (sound) waves. As
written in Section 3.1.1, the Navier-Stokes equations describe the propagation of information at speeds
comparable to that of the uid ow (for re, 10-20 m/s), but also at speeds comparable to that of sound
waves (for still air, 300 m/s). Solving a discretized form of these equations would require extremely small
time steps in order to account for information traveling at the speed of sound, making practical simulations
difcult.
Following the work of Rehm and Baum [10], an approximation to the equation of state (3.4) is made by
decomposing the pressure into a background component and a perturbation. The original version of FDS
assumed that the background component of the pressure applied to the entire computational domain, most
often a single compartment. Starting in FDS version 5, it is now assumed that the background component
of pressure can differ from compartment to compartment. If a volume within the computational domain
is isolated from other volumes, except via leak paths or ventilation ducts, it is referred to as a pressure
zone and assigned its own background pressure. The pressure within the mth zone, for example, is a linear
combination of its background component and the ow-induced perturbation:
p(x, t) = p
m
(z, t) + p(x, t) (3.11)
Note that the background pressure is a function of z, the vertical spatial coordinate, and time. For most
compartment re applications, p
m
changes very little with height or time. However, for situations where the
pressure increases due to a re in a tightly sealed enclosure, or when the height of the domain is signicant,
p
m
takes these effects into account [25]. The ambient pressure eld is denoted p
0
(z). Note that the subscript
0 denotes the exterior of the computational domain, not time 0. This is the assumed atmospheric pressure
stratication that serves as both the initial and boundary condition for the governing equations.
The purpose of decomposing the pressure is that for low-Mach number ows, it can be assumed that the
temperature and density are inversely proportional, and thus the equation of state (in the mth pressure zone)
can be approximated
p
m
= TR
(Y
/W
) = TR /W (3.12)
The pressure p in the state and energy equations is replaced by the background pressure p
m
to lter out
sound waves that travel at speeds that are much faster than typical ow speeds expected in re applications.
17
The low Mach number assumption serves two purposes. First, the ltering of acoustic waves means that the
time step in the numerical algorithm is bound only by the ow speed as opposed to the speed of sound, and
second, the modied state equation leads to a reduction in the number of dependent variables in the system
of equations by one. The energy equation (3.3) is never explicitly solved, but its source terms are included
in the expression for the ow divergence, to be derived presently.
The stratication of the atmosphere is derived from the relation
dp
0
dz
=
0
(z)g (3.13)
where
0
is the background density. Using Eq. (3.12), the background pressure can be written as a function
of the background temperature, T
0
(z)
p
0
(z) = p
exp
_
z
z
W g
R T
0
dz
_
(3.14)
where the subscript innity generally refers to the ground. Usually, the temperature stratication of the
atmosphere is specied by the user, in which case p
0
and
0
are derived from Eqs. (3.14) and (3.12),
respectively.
3.1.3 Combination of the Mass and Energy Equations via the Divergence
Because of the low Mach number assumption, the divergence of the ow, u, plays a very important
role in the overall solution scheme. The divergence is obtained by taking the material (total) derivative of
the modied Equation of State (3.12), and then substituting terms from the mass and energy conservation
equations. A few assumptions are needed. First, dene the constant-pressure specic heat of the mixture:
c
p
=
c
p,
Y
where c
p,
is the temperature-dependent specic heat of species . Next, dene the enthalpy
h =
where
h
T
T
ref
c
p,
(T
) dT
(3.15)
Finally, note that the material derivative of the background pressure, p
m
(z, t), is written:
Dp
m
Dt
=
p
m
t
+w
p
m
z
=
p
m
t
w
m
g (3.16)
The divergence can now be written:
u =D+P
p
m
t
(3.17)
where
D =
m
W
W
b
+
R
Wc
p
p
m
_
q
b
q
_
P w
m
g +
W
(Y
/W
)
R
Wc
p
p
m
+
1
_
W
W
c
p
T
_
m
P =
1
p
m
_
R
Wc
p
1
_
Contributions to the divergence of the ow include the heat release rate of the re, q
b
, the net heat ux from thermal conduction and radiation, q
m
DdV
m
u dS
_
_
m
P dV (3.18)
Equation (3.18) is essentially a consistency condition, ensuring that blowing air or starting a re within a
sealed compartment leads to an appropriate decrease in the divergence within the volume.
3.1.4 The Momentum Equation
The momentum equation is simplied to make it easier to solve numerically. First, we start with the non-
conservative form of the momentum equation introduced above
_
u
t
+(u )u
_
+p = g+f
b
+
i j
(3.19)
Next, we make the following substitutions:
1. Subtract the hydrostatic pressure gradient of the nth pressure zone,
n
(z, t)g, from both sides. Note
that p =
n
g+ p.
2. Apply the vector identity: (u )u = |u|
2
/2u
3. Divide all terms by the density,
4. Decompose the pressure term, noting that
=
p
+
_
1
_
p
5. Dene H |u|
2
/2+ p/
_
p =
1
_
(
n
)g+f
b
+
i j
_
(3.20)
The numerical solution of the pressure equation obtained by taking the divergence of Eq. (3.20) is greatly
simplied by either neglecting the last term on the left hand side, or in cases where it cannot be neglected,
treating it with some care. The decision to either neglect the extra pressure term or to approximate it depends
on its relative contribution to the creation of vorticity. An evolution equation for the circulation, =
u dx,
reveals the sources of vorticity in the absence of any external forces, f
b
:
d
dt
=
_
1
_
p dx+
g dx+
i j
dx (3.21)
The rst term on the right hand side represents the baroclinic torque. The second term is buoyancy-induced
vorticity. The third term represents the vorticity generated by molecular and subgrid-scale stresses, as in
boundary and shear layers. In most practical large scale re simulations, the re itself occupies a small part
of the computational domain. Hence, the re is often not well resolved by the numerical grid, in which case
19
the vorticity generated in regions where there are large deviations in density is not captured directly. The
mixing of air and combustion products occurs in the plume above the re where buoyancy is the dominant
source of vorticity. In these calculations, the baroclinic torque can be neglected to simplify the numerical
solution. In simulations where detailed ame dynamics are resolvable, the pressure term responsible for the
baroclinic torque cannot be neglected, but for reasons to be made clear below, must be treated differently
than the other pressure term. In neither of these cases is the Boussinesq approximation invoked the uid
is still considered thermally-expandable; the divergence is non-zero; and the mass and energy equations are
not modied.
3.1.5 The Equation for Pressure (Poissons Equation)
The reason for either neglecting the baroclinic torque or decomposing the pressure term in the momentum
equation is to simplify the elliptic partial differential equation obtained by taking the divergence of the
momentum equation
2
H =
( u)
t
F ; F =u+
_
1
_
p
1
_
(
0
)g+f
b
+
i j
_
(3.22)
Note that the pressure p appears on both sides of Eq. (3.22). The pressure on the right hand side is taken from
the previous time step of the overall explicit time-marching scheme. It can be neglected if the baroclinic
torque is not considered important in a given simulation. The pressure on the left hand side (incorporated
in the variable H) is solved for directly. The reason for the decomposition of the pressure term is so that
the linear algebraic system arising from the discretization of Eq. (3.22) has constant coefcients (i.e. it is
separable) and can be solved to machine accuracy by a fast, direct (i.e. non-iterative) method that utilizes
Fast Fourier Transforms (FFT). No-ux or forced-ow boundary conditions are specied by asserting that
H
n
=F
n
u
n
t
(3.23)
where F
n
is the normal component of F at the vent or solid wall, and u
n
/t is the prescribed rate of change
in the normal component of velocity at a forced vent. Initially, the velocity is zero everywhere.
At open external boundaries the pressure-like term H is prescribed depending on whether the ow is
outgoing or incoming
H =|u|
2
/2 outgoing
H = 0 incoming
(3.24)
The outgoing boundary condition assumes that the pressure perturbation p is zero at an outgoing boundary
and that H is constant along streamlines. The incoming boundary condition assumes that H is zero innitely
far away. At the boundary between two meshes, the pressure boundary condition is similar to that at an
external open boundary, except that where the ow is incoming, H is taken from the adjacent mesh.
3.1.6 Large Eddy Simulation (LES)
The most distinguishing feature of any CFD model is its treatment of turbulence. Chapter 1 contains a brief
history of turbulence modeling as it has been applied to the re problem. Of the three main techniques of
simulating turbulence, FDS contains only Large Eddy Simulation (LES) and Direct Numerical Simulation
(DNS). There is no Reynolds-Averaged Navier-Stokes (RANS) capability in FDS.
LES is a technique used to model the dissipative processes (viscosity, thermal conductivity, material
diffusivity) that occur at length scales smaller than those that are explicitly resolved on the numerical grid.
This means that the parameters , k and D in the equations above cannot be used directly in most practical
20
simulations. They must be replaced by surrogate expressions that model their impact on the approximate
form of the governing equations. This section contains a simple explanation of how these terms are modeled
in FDS. Note that this discussion is quite different than what it typically found in the literature, thus the
reader is encouraged to consider other explanations of the technique in the references that are listed in a
review article by Pope [26].
Recall from Section 3.1.1 that there is a small term in the energy equation known as the dissipation rate
, the rate at which kinetic energy is converted to thermal energy by viscosity. To understand where this
term originates, form an evolution equation for the kinetic energy of the uid by taking the dot product of
the momentum equation (3.2) with the velocity vector
1
:
Du
Dt
u =
D
_
|u|
2
/2
_
Dt
= f
b
up u+ (
i j
u) (3.25)
As mentioned above is a neglible quantity in the energy equation. However, its functional form is useful
in representing the dissipation of kinetic energy from the resolved ow eld. Following the analysis of
Smagorinsky [27], the viscosity is modeled
LES
= (C
s
)
2
_
2 S
i j
S
i j
2
3
( u)
2
_1
2
(3.26)
where C
s
is an empirical constant and is a length on the order of the size of a grid cell. The bar above
the various quantities denotes that these are the resolved, or ltered, values, meaning that they are computed
on a numerical grid. The other diffusive parameters, the thermal conductivity and material diffusivity, are
related to the turbulent viscosity by
k
LES
=
LES
c
p
Pr
t
; (D)
l,LES
=
LES
Sc
t
(3.27)
The turbulent Prandtl number Pr
t
and the turbulent Schmidt number Sc
t
are assumed to be constant for a
given scenario.
The model for the viscosity,
LES
, serves two roles: rst, it provides a stabilizing effect in the numerical
algorithm, damping out numerical instabilities as they arise in the ow eld, especially where vorticity is
generated. Second, it has the appropriate mathematical form to describe the dissipation of kinetic energy
from the ow. Note the similar mathematical form of
LES
and the dissipation rate, , dened in Eq. (3.10).
In the parlance of the turbulence community, the dissipation rate is related to the turbulent kinetic energy
(most often denoted by k) by the relation k
3/2
/L, where L is a length scale.
There have been numerous renements of the original Smagorinsky model [28, 29, 30], but it is difcult
to assess the improvements offered by these newer schemes for res. There are two reasons for this. First,
the structure of the re plume is so dominated by the large-scale resolvable eddies that even a constant eddy
viscosity gives results comparable to those obtained using the Smagorinsky model [31]. Second, the lack of
precision in most large-scale re test data makes it difcult to assess the relative accuracy of each model.
The Smagorinsky model with constant C
s
produces satisfactory results for most large-scale applications
where boundary layers are not well-resolved [?]. In fact, experience to date using the simple form of LES
described above has shown that the best results are obtained when the Smagorinsky constant C
s
is set as low
as possible to maintain numerical stability. In other words, the most realistic ow simulations are obtained
when resolvable eddies are not damped by excessive amounts of articial viscosity.
1
In this section it is convenient to work with the Lagrangian form of the conservation equations.
21
3.1.7 Direct Numerical Simulation (DNS)
There are some ow scenarios where it is possible to use the molecular properties , k and D directly.
Usually, this means that the numerical grid cells are on the order of 1 mm or less, and the simulation
is regarded as a Direct Numerical Simulation (DNS). For a DNS, the viscosity, thermal conductivity and
material diffusivity are approximated from kinetic theory because the temperature dependence of each is
important in combustion scenarios. The viscosity of the species is given by
=
26.6910
7
(W
T)
1
2
v
kg
m s
(3.28)
where
A) and
v
is the collision integral, an empirical func-
tion of the temperature T. The thermal conductivity of species is given by
k
c
p,
Pr
W
m K
(3.29)
where the Prandtl number Pr is 0.7. The viscosity and thermal conductivity of a gas mixture are given by
DNS
=
; k
DNS
=
(3.30)
The binary diffusion coefcient of species diffusing into species is given by
D
=
2.6610
7
T
3/2
W
1
2
D
m
2
s
(3.31)
where W
=2(1/W
+1/W
)
1
,
= (
)/2, and
D
is the diffusion collision integral, an empirical
function of the temperature T [32]. It is assumed that nitrogen is the dominant species in any combustion
scenario considered here, thus the diffusion coefcient in the species mass conservation equations is that of
the given species diffusing into nitrogen
(D)
,DNS
= D
0
(3.32)
where species 0 is nitrogen.
22
3.2 Combustion Model
There are two types of combustion models used in FDS. The default model makes use of the mixture fraction,
a quantity representing the fuel and the products of combustion. For the second model, individual gas species
react according to specied Arrhenius reaction parameters. This latter model is most often used in a direct
numerical simulation (DNS) where the diffusion of fuel and oxygen can be modeled directly. However,
most often for large eddy simulations (LES), where the grid is not ne enough to resolve the diffusion of
fuel and oxygen, the mixture fraction-based combustion model is assumed.
3.2.1 Mixture Fraction Combustion Model
Given a volume containing a mixture of gas species, a mixture fraction can be dened that is the ratio of
the mass of a subset of the species to the total mass present in the volume. In combustion, the mixture
fraction is a conserved quantity traditionally dened as the (mass) fraction of the gas mixture that originates
in the fuel stream. Thus, at a burner surface the mixture fraction is 1 and in fresh air it is 0. In a region
where combustion has occurred this fraction will be comprised of any unburned fuel and that portion of
the combustion products that came from the fuel. The mixture fraction is a function of space and time,
commonly denoted Z(x, t). If it can be assumed that, upon mixing, the reaction of fuel and oxygen occurs
rapidly and completely, the combustion process is referred to as mixing-controlled. This implies that all
species of interest can be described in terms of the mixture fraction alone. The correspondence between the
mass fraction of an individual species and the mixture fraction is called its state relation. FDS versions
2 through 4 assumed that the gas mixture could be uniquely determined by the mixture fraction alone, an
assumption that implies that fuel and oxygen react instantaneously upon mixing.
For many applications, mixed is burned is a reasonable assumption. However, for re scenarios where
it cannot be assumed that fuel and oxygen react completely upon mixing, for example in under-ventilated
compartments. The mixture fraction itself remains a valid quantity, but it can no longer be assumed that it
completely denes the composition of the gas mixture. If fuel and oxygen are to mix and not burn, at least
two scalar variables are needed to describe the extent to which the fuel and oxygen react. The strategy for
moving beyond the mixed is burned model is as follows. Instead of solving a single transport equation for
the mixture fraction Z, multiple transport equations are solved for components of the mixture fraction Z
.
Fuel mass is still conserved, since Z
CO
2
CO
2
+
H
2
O
H
2
O+
CO
CO+
S
S+
N
2
N
2
+
M
M (3.33)
Note that the nitrogen in the fuel molecule is assumed to form N
2
only. Addition product species can be
specied as some number of moles of an average molecular weight species M. These products are presumed
to not consume oxygen during their formation. Soot is assumed to be a mixture of carbon and hydrogen
with the hydrogen atomic fraction given by X
H
. The stoichiometric coefcient,
S
, represents the amount of
fuel that is converted to soot. It is related to the soot yield, y
S
, via the relation:
S
=
W
F
W
S
y
S
; W
S
= X
H
W
H
+(1X
H
)W
C
(3.34)
23
Figure 3.1: State relations for methane.
Likewise, the stoichiometric coefcient of CO,
CO
, is related to the CO yield, y
CO
, via:
CO
=
W
F
W
CO
y
CO
(3.35)
The yields of soot and CO are based on well-ventilated or post-ame measurements. The increased
production of CO and soot in an under-ventilated compartment will be addressed in the following sections.
The mixture fraction, Z, can be dened in terms of the mass fraction of fuel and the carbon-carrying
products of combustion:
Z =Y
F
+
W
F
xW
CO
2
Y
CO
2
+
W
F
xW
CO
Y
CO
+
W
F
xW
S
Y
S
(3.36)
The mixture fraction satises the conservation equation
DZ
Dt
= DZ (3.37)
obtained by taking a linear combination of the transport equations for the fuel and the carbon carrying
products. If it is assumed that combustion occurs so rapidly that the fuel and oxygen cannot co-exist, then
both simultaneously vanish at a ame surface:
Z(x, t) = Z
f
; Z
f
=
Y
O
2
sY
I
F
+Y
O
2
(3.38)
and all species can be related to Z via the state relations shown in Fig. 3.1. In versions of FDS prior to
5, this one-step, instantaneous reaction of fuel and oxygen was assumed. However, starting in version 5, a
more generalized formulation has been implemented and is described next.
24
A Single-Step Reaction, but with Local Extinction
The physical limitation of the single-step reaction model described in the previous section is that it assumes
fuel and oxygen burn instantaneously when mixed. For large-scale, well-ventilated res, this is a good as-
sumption. However, if a re is in an under-ventilated compartment, or if a suppression agent like water mist
or CO
2
is introduced, or if the shear layer between fuel and oxidizing streams has a sufciently large local
strain rate, fuel and oxygen may mix but may not burn. The physical mechanisms underlying these phenom-
ena are complex, and even simplied models still rely on an accurate prediction of the ame temperature
and local strain rate. Subgrid-scale modeling of gas phase suppression and extinction is still an area of active
research in the combustion community.
Simple empirical rules can be used to predict local extinction based on the oxygen concentration and
temperature of the gases in the vicinity of the ame sheet. Figure 3.2 shows values of temperature and
oxygen concentration for which burning can and cannot take place. A derivation of the model, based on the
critical ame temperature concept, is given in Appendix C. Note that once the local state of the gases falls
Figure 3.2: Oxygen-temperature phase space showing where com-
bustion is allowed and not allowed to take place.
into the No Burn zone, the state relations (Fig. 3.1) are no longer valid for values of Z below stoichio-
metric, since now some fuel may be mixed with the other combustion products. Essentially, there are now
two reactions to consider the null reaction, where fuel and oxygen simply mix and do not burn; and the
complete reaction, where fuel and oxygen react and form products according to Eq. (3.33). Note that the
term complete does not imply that no soot or CO is formed, but rather that their respective production
rates are proportional to the fuel consumption rate.
With the denition of the mixture fraction, Eq. (3.36), in mind, consider a partitioning of Z into the
following components:
Z
1
= Y
F
(3.39)
Z
2
=
W
F
xW
CO
2
Y
CO
2
+
W
F
xW
CO
Y
CO
+
W
F
xW
S
Y
S
(3.40)
such that Z = Z
1
+Z
2
. Transport equations are required for both Z
1
and Z
2
. At the burner surface, Z
1
is
25
assigned the mass ux of fuel, while the mass ux for Z
2
is zero. In other words, no combustion products
are generated at the fuel source. Where fuel and oxygen co-exist, a reaction occurs if conditions are favorable
in the sense shown by Fig. 3.2. If a reaction occurs, Z
1
is converted to Z
2
representing the conversion of fuel
to products. The heat release rate of the re is obtained by multiplying the fuel consumption rate by the heat
of combustion.
We expand upon this by allowing for the fuel stream to be diluted with an inert gas (nitrogen) and we
note that since we are using our extra parameter to determine extinction, that the yields of CO and soot are
still xed. Thus we can restate Eq. (3.39) and Eq. (3.40) as:
Z
1
=
Y
F
Y
I
F
(3.41)
Z
2
=
W
F
_
x
CO
(1X
H
)
S
W
CO
2
Y
CO
2
Y
I
F
(3.42)
where Y
I
F
is the fuel mass fraction at the burner surface (Y
I
N
2
would be the diluent mass fraction at the burner
surface).
Since the mixture fraction variables result from linear combinations of the species transport equations,
the converse is also true that species mass fractions are linear combinations of the mixture fraction vari-
ables. The mass fractions of the species in the mixture, Y
(Z
1
, Z
2
), are found via:
Y
F
= Y
I
F
Z
1
Y
N
2
= (1Z) Y
N
2
+Y
I
N
2
Z
1
+
N
2
W
N
2
W
F
Y
I
F
Z
2
Y
O
2
= (1Z) Y
O
2
O
2
W
O
2
W
F
Y
I
F
Z
2
Y
CO
2
=
CO
2
W
CO
2
W
F
Y
I
F
Z
2
Y
H
2
O
=
H
2
O
W
H
2
O
W
F
Y
I
F
Z
2
(3.43)
Y
CO
=
CO
W
CO
W
F
Y
I
F
Z
2
(3.44)
Y
S
=
S
W
S
W
F
Y
I
F
Z
2
(3.45)
Y
M
=
M
W
M
W
F
Y
I
F
Z
2
(3.46)
The stoichiometric coefcients are dened:
N
2
=
a
2
O
2
=
CO
2
+
CO
+
H
2
O
z
2
CO
2
= x
CO
(1X
H
)
S
M
= b
H
2
O
=
y
2
X
H
S
(3.47)
CO
=
W
F
W
CO
y
CO
(3.48)
S
=
W
F
W
S
y
S
(3.49)
It is important to note that the denitions of Z
1
and Z
2
, unlike in the single parameter model, do not imply
anything regarding the rate of combustion, only that the combustion occurs in a single step.
CO Production (Two-Step Reaction with Extinction)
The previous section describes the complete reaction as the conversion of fuel to products such that the
production rate of each product species is proportional to the fuel consumption rate. This means that for
each fuel molecule, xed amounts of CO
2
, H
2
O, CO, and soot are formed and these products persist in the
26
plume indenitely with no further reaction. This is not an unreasonable assumption if the purpose of the re
simulation is to assess the impact of the re on the larger space. However, in under-ventilated res, soot and
CO are produced at higher rates, and exist within the fuel-rich ame envelope at higher concentrations, than
would otherwise be predicted with a single set of xed yields that are based on post-ame measurements.
To account for the production of CO and its eventual oxidation at the ame envelope or within a hot upper
layer, an additional reaction is now needed:
C
x
H
y
O
z
N
a
M
b
+
O
2
O
2
H
2
O
H
2
O+(
CO
+
CO
)CO+
S
S+
N
2
N
2
+
M
M (3.50)
CO
_
CO+
1
2
O
2
CO
2
_
(3.51)
The brackets around the second reaction are there merely to emphasize that the sum of the two reactions
equal Eq. (3.33). There are two stoichiometric coefcients for CO the rst,
CO
= x (1 X
H
)
S
CO
,
represents CO that is produced in the rst step of the reaction that can potentially be converted to CO
2
assuming the conditions are favorable.
CO
is equivalent to
CO
2
in Eq. (3.33). The second coefcient,
CO
, is the so-called well-ventilated, or post-ame, value that was introduced in the previous section.
The proposed model of CO production still does not contain the necessary kinetic mechanism to predict the
post-ame concentration of CO without the prescription of the measured value of the post-ame CO yield.
Rather, the proposed model includes the production of large amounts of CO in the rst step of a two-step
reaction, followed by a partial conversion to CO
2
if there is a sufcient amount of oxygen present.
To describe the composition of the gas species, the mixture fraction, Z, must now be decomposed into
three components:
Z
1
=
Y
F
Y
I
F
(3.52)
Z
2
=
W
F
_
x(1X
H
)
S
W
CO
Y
CO
Y
I
F
(3.53)
Z
3
=
W
F
_
x(1X
H
)
S
W
CO
2
Y
CO
2
Y
I
F
(3.54)
Here, x(1X
H
)
S
represents the number of carbon atoms in the fuel molecule that are oxidized; that is, the
carbon atoms that are not converted to soot. The decomposition of Z into three components is numerically
convenient. However, the recovery of the individual species mass fractions requires some care. The mass
fraction of any species in the mixture, Y
(Z
1
, Z
2
, Z
3
), is still found via linear combinations of the mixture
fraction variables:
Y
F
= Z
1
Y
I
F
Y
N
2
= (1Z) Y
N
2
+Y
I
N
2
Z
1
+
N
2
W
N
2
W
F
Y
I
F
(Z
2
+Z
3
)
Y
O
2
= (1Z) Y
O
2
W
O
2
Y
I
F
W
F
(
O
2
Z
2
+
O
2
Z
3
)
Y
CO
2
=
CO
2
W
CO
2
W
F
Y
I
F
Z
3
Y
H
2
O
=
H
2
O
W
H
2
O
W
F
Y
I
F
(Z
2
+Z
3
) (3.55)
Y
CO
=
(
CO
+
CO
)W
CO
2
W
F
Y
I
F
Z
2
(3.56)
Y
S
=
S
W
S
W
F
Y
I
F
(Z
2
+Z
3
) (3.57)
Y
M
=
M
W
M
W
F
Y
I
F
(Z
2
+Z
3
) (3.58)
The stoichiometric coefcients are dened:
27
N
2
=
a
2
O
2
=
CO
+
H
2
O
2
O
2
=
CO
2
+
CO
+
H
2
O
z
2
CO
2
= x
CO
(1X
H
)
S
M
= b
H
2
O
=
y
2
X
H
S
(3.59)
CO
= x
CO
(1X
H
)
S
(3.60)
CO
=
W
F
W
CO
y
CO
(3.61)
S
=
W
F
W
S
y
S
(3.62)
Although these formulae appear complicated, most are determined directly from the composition of the fuel
molecule. The only information expected of the modeler are the fuel composition, the soot and CO yields,
and the atomic fraction of hydrogen in the soot.
Heat Release Rate
The discussion of the various multi-step reactions above is essentially book-keeping, the accounting of the
gas molecules formed in the combustion process. But what of the heat released?
When tracking species with two mixture fraction parameters, there is a single step reaction, the conver-
sion of fuel into a xed, predened set of combustion products. Fuel and oxygen react within a grid cell and
release energy according to the rate of fuel consumption:
q
= m
F
H
F
(3.63)
The amount of fuel in the grid cell is obtained directly from the denition of Z
1
whereas the amount of
O
2
is obtained using the state relationships discussed in the prior section. Combustion is either allowed
or disallowed
2
using the relation shown in Fig. 3.2. If combustion is allowed to occur in a grid cell, the
single step combustion is assumed to be innitely fast. Combustion will consume all of either the fuel or the
oxygen in the grid cell.
In the case of the two-step, three parameter mixture fraction model, the rst step converts the fuel to CO
and other combustion products, and the second step oxidizes the CO into CO
2
. The rst step is determined
as it is for the single step reaction. The second step, however, is assumed to be rate dependent. Determining
if the second step can take place in a grid cell is done in one of two ways. For most LES simulations, ame
temperatures will not be resolved on the grid because the heat released is smeared out over a grid cell much
larger than the ame sheet thickness. Since a grid cell with heat released from the rst step has a ame,
CO conversion is always allowed in these grid cells. In all other grid cells, the local temperature is used to
perform a nite-rate computation that determines the rate of CO conversion.
On-Going Research
One can envision extending the three parameter mixture fraction model to a fourth parameter for soot pro-
duction:
Z
4
=
W
F
xW
S
Y
S
(3.64)
This extension would allow for combustion models that predict soot formation along with extinction and
CO production.
2
Note that the user has control over the parameters associated with local gas phase extinction.
28
Additional partitionings of the mixture fraction could be added as well; however, there is a point of
diminishing returns. The present advantage of using a mixture fraction approach is to track the movement
of a number of gas species using a smaller number of scalar variables at the expense of having to use state
relationships to obtain species mass fractions. As more mixture fraction variables are added, the net savings
in fewer species transport equations vs. the expense of state relationships will diminish.
29
3.2.2 Finite-Rate Reaction (DNS)
In a DNS calculation, the ne grid resolution enables the direct modeling of the diffusion of chemical species
(fuel, oxygen, and combustion products). Since the ame is being resolved in a DNS calculation, the local
gas temperatures can be used to determine the reaction kinetics. Thus, it is possible to implement a relatively
simple set of one or more chemical reactions to model the combustion. Consider the reaction of oxygen and
a hydrocarbon fuel
C
x
H
y
+
O
2
O
2
CO
2
CO
2
+
H
2
O
H
2
O (3.65)
If this were modeled as a single-step reaction, the reaction rate would be given by the expression
d[C
x
H
y
]
dt
=B[C
x
H
y
]
a
[O
2
]
b
e
E/RT
(3.66)
Suggested values of B, E, a and b for various hydrocarbon fuels are given in Refs. [33, 34]. It should be
understood that the implementation of any of these one-step reaction schemes is still very much a research
exercise because it is not universally accepted that combustion phenomena can be represented by such a
simple mechanism. Improved predictions of the heat release rate may be possible by considering a multi-
step set of reactions. However, each additional gas species dened in the computation incurs a roughly 5 %
increase in the CPU time.
30
3.3 Thermal Radiation
Energy transport consists of convection, conduction and radiation. Convection of heat is accomplished via
the solution of the basic conservation equations. Gains and losses of heat via conduction and radiation are
represented by the divergence of the heat ux vector in the energy equation, q
r
.
The Radiative Transport Equation (RTE) for an absorbing/emitting and scattering medium is
s I
(x, s) =
_
(x, ) +
s
(x, )
_
I
(x, s) +B(x, ) +
s
(x, )
4
4
(s, s
) I
(x, s
) ds
(3.67)
where I
(x, s) is the radiation intensity at wavelength , s is the direction vector of the intensity, (x, ) and
s
(x, ) are the local absorption and scattering coefcients, respectively, and B(x, ) is the emission source
term. The integral on the right hand side describes the in-scattering from other directions. In the case of a
non-scattering gas the RTE becomes
s I
(x, s) = (x, )
_
I
b
(x) I
(x, s)
_
(3.68)
where I
b
(x) is the source term given by the Planck function (see below).
In practical simulations the spectral () dependence cannot be solved accurately. Instead, the radiation
spectrum is divided into a relatively small number of bands and a separate RTE is derived for each band.
The band specic RTE is
s I
n
(x, s) =
n
(x) [I
b,n
(x) I
n
(x, s)] , n = 1...N (3.69)
where I
n
is the intensity integrated over the band n, and
n
is the appropriate mean absorption coefcient
inside the band. The source term can be written as a fraction of the blackbody radiation
I
b,n
= F
n
(
min
,
max
) T
4
/ (3.70)
where is the Stefan-Boltzmann constant. The calculation of factors F
n
is explained in Ref. [20]. When the
intensities corresponding to the bands are known, the total intensity is calculated by summing over all the
bands
I(x, s) =
N
n=1
I
n
(x, s) (3.71)
Even with a reasonably small number of bands, solving multiple RTEs is very time consuming. For-
tunately, in most large-scale re scenarios soot is the most important combustion product controlling the
thermal radiation from the re and hot smoke. As the radiation spectrum of soot is continuous, it is pos-
sible to assume that the gas behaves as a gray medium. The spectral dependence is then lumped into one
absorption coefcient (N = 1) and the source term is given by the blackbody radiation intensity
I
b
(x) = T(x)
4
/ (3.72)
This is the default mode of FDS and appropriate for most problems of re engineering. In optically thin
ames, where the amount of soot is small compared to the amount of CO
2
and water, the gray gas assumption
may produce signicant overpredictions of the emitted radiation. From a series of numerical experiments it
has been found that six bands (N = 6) are usually enough to improve the accuracy in these cases. The limits
of the bands are selected to give an accurate representation of the most important radiation bands of CO
2
and water. If the absorption of the fuel is known to be important, separate bands can be reserved for fuel,
31
Table 3.1: Limits of the spectral bands.
9 Band Model 1 2 3 4 5 6 7 8 9
Soot CO
2
CH
4
Soot CO
2
H
2
O H
2
O Soot Soot
Major Species
H
2
O, Soot Soot Soot Soot CH
4
, Soot
(1/cm) 10000 3800 3400 2800 2400 2174 1429 1160 1000 50
(m) 1.00 2.63 2.94 3.57 4.17 4.70 7.00 8.62 10.0 200
6 Band Model 1 2 3 4 5 6
Soot CO
2
CH
4
CO
2
H
2
O, CH
4
, Soot Soot
Major Species
H
2
O, Soot Soot Soot
and the total number of bands is increased to nine (N = 9). For simplicity, the fuel is assumed to be CH
4
.
The limits of the bands are shown in Table 3.1.
For the calculation of the gray or band-mean absorption coefcients,
n
, a narrow-band model, Rad-
Cal [35], has been implemented in FDS. At the start of a simulation the absorption coefcient(s) are tabulated
as a function of mixture fraction and temperature. During the simulation the local absorption coefcient is
found by table-lookup.
In calculations of limited spatial resolution, the source term, I
b
, in the RTE requires special treatment in
the neighborhood of the ame sheet because the temperatures are smeared out over a grid cell and are thus
considerably lower than one would expect in a diffusion ame. Because of its fourth-power dependence
on the temperature, the source term must be modeled in those grid cells cut by the ame sheet. Elsewhere,
there is greater condence in the computed temperature, and the source term can be computed directly
I
b
=
_
T
4
/ Outside ame zone
max(
r
q
/4 , T
4
/) Inside ame zone
(3.73)
Here, q
+
1
n
w
<0
I
w
(s
) |s
n
w
| ds
(3.74)
where I
w
(s) is the intensity at the wall, is the emissivity, and T
w
is the wall surface temperature. The radiant
heat ux vector q
r
is dened
q
r
(x) =
4
s
I(x, s
) ds
(3.75)
The gas phase contribution to the radiative loss term in the energy equation is
q
r
(x)(gas) = (x) [U(x) 4I
b
(x)] ; U(x) =
4
I(x, s
)ds
(3.76)
In words, the net radiant energy gained by a grid cell is the difference between that which is absorbed and
that which is emitted.
3
The radiative fraction,
r
, is a useful quantity in re science. Usually, it is understood to be the fraction of the total heat release
rate that takes the form of thermal radiation. For most combustibles,
r
is between 0.3 and 0.4. However, in Eq. (3.73),
r
is
interpreted as the fraction of energy radiated from the combustion region. For a small re (D < 1 m), the local
r
is approximately
equal to its global counterpart. However, as the re increases in size, the global value will typically decrease due to a net re-
absorption of the thermal radiation by the increasing smoke mantle.
32
3.4 Solid Phase Model
FDS assumes that solid obstructions consist of multiple layers, with each layer composed of multiple mate-
rial components that can undergo multiple thermal degradation reactions. Each reaction forms a combina-
tion of solid residue (i.e. another material component), water vapor, and/or fuel vapor. Heat conduction is
assumed only in the direction normal to the surface. This section describes the single mass and energy con-
servation equation for solid materials, plus the various coefcients, source terms, and boundary conditions.
3.4.1 The Heat Conduction Equation for a Solid
A one-dimensional heat conduction equation for the solid phase temperature T
s
(x, t) is applied in the direc-
tion x pointing into the solid (the point x = 0 represents the surface)
4
s
c
s
T
s
t
=
x
k
s
T
s
x
+ q
s
(3.78)
Section 3.4.2 describes the component-averaged material properties, k
s
and
s
c
s
. The source term, q
s
,
consists of chemical reactions and radiative absorption:
q
s
= q
s,c
+ q
s,r
(3.79)
Section 3.4.3 describes the term q
s,c
, which is essentially the heat production (loss) rate given by the pyrol-
ysis models for different types of solid and liquid fuels. Section 3.4.4 describes the term q
s,r
, the radiative
absorption and emission in depth. Section 3.4.5 describes the convective heat transfer to the solid surface.
Finally, Section 3.4.6 describes how the radiative and convective heat uxes are applied as boundary condi-
tions for Eq. (3.78).
3.4.2 Component-Averaged Thermal Properties
The conductivity and volumetric heat capacity of the solid are dened
k
s
=
N
m
=1
X
k
s,
;
s
c
s
=
N
m
=1
s,
c
s,
(3.80)
N
m
is the number of material components forming the solid.
s,
is the component density
s,
=
s
Y
(3.81)
where
s
is the density of the composite material and Y
s
=
N
m
=1
s,
(3.82)
4
In cylindrical and spherical coordinates, the heat conduction equation is written
s
c
s
T
s
t
=
1
r
r
_
rk
s
T
s
r
_
+ q
s
;
s
c
s
T
s
t
=
1
r
2
r
_
r
2
k
s
T
s
r
_
+ q
s
(3.77)
FDS offers the user these options, with the assumption that the obstruction is not actually recti-linear, but rather cylindrical or
spherical in shape. This option is useful in describing the behavior of small, complicated targets like cables or heat detection
devices.
33
X
=
s,
_
N
m
=1
s,
(3.83)
where
is the density of material in its pure form. Multi-component solids are dened by specifying the
mass fractions, Y
, and densities,
s,c
, are modeled. This is what is commonly referred to as the pyrolysis model, but
it actually can represent any number of reactive processes, including evaporation, charring, and internal
heating.
Specied Heat Release Rate
Often the intent of a re simulation is merely to predict the transport of smoke and heat from a specied re.
In other words, the heat release rate is a user input, not something the model predicts. In these instances, the
desired HRR is translated into a mass ux for fuel at a given solid surface, which can be thought of as the
surface of a burner:
m
f
=
f (t) q
user
H
(3.84)
Usually, the user species a desired heat release rate per unit area, q
user
, plus a time ramp, f (t), and the
mass loss rate is computed accordingly.
Solid Fuels
Solids can undergo simultaneous reactions under the following assumptions:
instantaneous release of volatiles from solid to the gas phase,
local thermal equilibrium between the solid and the volatiles,
no condensation of gaseous products, and
no porosity effects
5
Each material component may undergo several competing reactions, and each of these reactions may pro-
duce some other solid component (residue), gaseous fuel, and/or water vapor according to the yield coef-
cients
s
,
f
and
w
, respectively. These coefcients should usually satisfy
s
+
f
+
w
= 1, but smaller
yields may also be used to take into account the gaseous products that are not explicitly included in the
simulation.
Consider material component that undergoes N
r,
separate reactions. We will use the index to
represent one of these reactions:
Material
s,
Residue
+
w,
H
2
O+
f ,
HC (3.85)
5
Although porosity effects are not explicitly included in the model, it is possible to account for it because the volume fractions
dened by Eq. (3.83) need not sum to unity, in which case the thermal conductivity and absorption coefcient are effectively
reduced.
34
The local density of material component evolves in time according to the solid phase species conservation
equation
t
_
s,
s0
_
=
N
r,
=1
r
+S
(3.86)
which says that the mass of conponent is consumed by the solid phase reactions r
and produced by
other reactions. r
is the production rate of material component as a result of the reactions of the other components. The
reaction rates are functions of local mass concentration and temperature, and calculated as a combination of
Arrhenius and power functions:
r
=
_
s,
s0
_
n
s,
A
exp
_
RT
s
_
max
_
0, T
s
T
thr,
n
t,
(3.87)
where T
thr,
is a threshold temperature that can be used to dictate that the reaction must not occur below a
user-specied temperature. By default, the term is deactivated (T
thr,
= 0 K).
The production term S
is the sum over all the reactions where the solid residue is material
S
=
N
m
=1
N
r,
=1
s,
(where Residue
= Material
) (3.88)
The volumetric production rate of fuel gases and water vapor are
m
f
=
s0
N
m
=1
N
r,
=1
f ,
r
(3.89)
m
w
=
s0
N
m
=1
N
r,
=1
w,
r
(3.90)
It is assumed that the fuel gases and water vapor are transported instantaneously to the surface, where the
mass uxes are given by:
m
f
=
L
0
m
f
(x)dx ; m
w
=
L
0
m
w
(x)dx (3.91)
where L is the thickness of the surface. The chemical source term of the heat conduction equation consists
of the heats of reaction and the differences in latent heats between the original material and the products
q
s,c
(x) =
s0
N
m
=1
N
r,
=1
r
(x)
_
H
r,
T(x)
T
0
c()d
_
(3.92)
It is possible to dene a reaction where the sum of the residue, fuel and water yields is less than one. FDS
then assumes that the solid is converted to a gas that is not tracked by the gas phase calculation. In the
computation of c(T), the specic heat of this dummy species is assumed to be equal to the specic heat of
the original material c
s,
c(T) = c
s,
(T)
s,
c
s,
(T)
f ,
c
f
(T)
w,
c
w
(T) (1
=s, f ,w
,
) c
s,
(T) (3.93)
c
s,
is the specic heat of the residue of the th reaction of material . The temperature integrals of the
latent heat contents are computed in advance for all the reactions of each material and the values of q
s,c
are tabulated between T
0
and T
0
+ 2000 K. During the pyrolysis calculation, the values are found from a
look-up table.
35
Liquid Fuels
The rate at which liquid fuel evaporates when burning is a function of the liquid temperature and the con-
centration of fuel vapor above the pool surface. According to the Clausius-Clapeyron relation, the volume
fraction of the fuel vapor above the surface is a function of the liquid boiling temperature
X
f
= exp
_
h
v
W
f
R
_
1
T
s
1
T
b
__
(3.94)
where h
v
is the heat of vaporization, W
f
is the molecular weight, T
s
is the surface temperature, and T
b
is the
boiling temperature of the fuel [36].
For simplicity, the liquid fuel itself is treated like a thermally-thick solid for the purpose of computing
the heat conduction. There is no computation of the convection of the liquid within the pool.
3.4.4 Radiation Heat Transfer to Solids
If it is assumed that the thermal radiation from the surrounding gases is absorbed within an innitely thin
layer at the surface of the solid obstruction, then the net radiative heat ux is the sum of incoming and
outgoing components, q
r
= q
r,in
q
r,out
, where:
q
r,in
=
n
w
<0
I
w
(s
) |s
n
w
| d (3.95)
q
r,out
= T
4
w
(3.96)
However, many common materials do not have innite optical thickness. Rather, the radiation penetrates
the material to some nite depth. The radiative transport within the solid (or liquid) can be described as
a source term in Eq. (3.78). A two-ux model based on the Schuster-Schwarzschild approximation [20]
assumes the radiative intensity is constant inside the forward and backward hemispheres. The transport
equation for the intensity in the forward direction is
1
2
dI
+
(x)
dx
=
s
_
I
b
I
+
(x)
_
(3.97)
where x is the distance from the material surface and
s
is the absorption coefcient
s
=
N
m
=1
X
s,
(3.98)
A corresponding formula can be given for the backward direction. Multiplying Eq. 3.97 by gives us the
forward radiative heat ux into the solid
1
2
d q
+
r
(x)
dx
=
s
_
T
4
s
q
+
r
(x)
_
(3.99)
The radiative source term in the heat conduction equation is a sum of the forward and backward ux
gradients
q
s,r
(x) =
d q
+
r
(x)
dx
+
d q
r
(x)
dx
(3.100)
The boundary condition for Eq. 3.99 at the solid (or liquid) surface is given by
q
+
r
(0) = q
r,in
+(1) q
r
(0) (3.101)
where q
r
(0) is the backward radiative heat ux at the surface. In this formulation, the surface emissivity
and the internal absorption are assumed to be independent properties of the material.
36
3.4.5 Convective Heat Transfer to Solids
The calculation of the convective heat ux depends on whether one is performing a Direct Numerical Sim-
ulation (DNS) or a Large Eddy Simulation (LES). In a DNS calculation, the convective heat ux to a solid
surface q
c
is obtained directly from the gas temperature gradient at the boundary
q
c
=k
T
n
(3.102)
where n is the spatial coordinate pointing into the solid. In an LES calculation, the convective heat ux to
the surface is obtained from a combination of natural and forced convection correlations
q
c
= h T W/m
2
; h = max
_
C|T|
1
3
,
k
L
0.037 Re
4
5
Pr
1
3
_
W/m
2
/K (3.103)
where T is the difference between the wall and the gas temperature (taken at the center of the grid cell
abutting the wall), C is the coefcient for natural convection (1.52 for a horizontal surface and 1.31 for
a vertical surface) [18], L is a characteristic length related to the size of the physical obstruction, k is the
thermal conductivity of the gas, and the Reynolds Re and Prandtl Pr numbers are based on the gas owing
past the obstruction. Since the Reynolds number is proportional to the characteristic length, L, the heat
transfer coefcient is weakly related to L. For this reason, L is taken to be 1 m for most calculations.
3.4.6 Boundary Conditions
The boundary condition on the front surface is
k
s
T
s
x
(0, t) = q
c
+ q
r
(3.104)
If the internal radiation is solved for a solid, the radiation boundary condition q
r
is not used.
On the back surface, two possible boundary condition types may be prescribed by the user. (1) If the
back surface is assumed to be open either to an ambient void or to another part of the computational domain,
the back side boundary condition is similar to that of the front side. (2) If the back side is assumed to be
perfectly insulated, an adiabatic boundary condition is used
k
s
T
s
x
= 0 (3.105)
37
3.5 Fire Detection Devices
FDS predicts the thermal environment resulting from a re, but it relies on various empirical models that
describe the activation of various re detection devices. These models are described in this section.
3.5.1 Sprinkler Activation
The temperature of the sensing element (or link) of an automatic re sprinkler is estimated from the
differential equation put forth by Heskestad and Bill [37], with the addition of a term to account for the
cooling of the link by water droplets in the gas stream from previously activated sprinklers
dT
l
dt
=
_
|u|
RTI
(T
g
T
l
)
C
RTI
(T
l
T
m
)
C
2
RTI
|u| (3.106)
Here T
l
is the link temperature, T
g
is the gas temperature in the neighborhood of the link, T
m
is the temper-
ature of the sprinkler mount (assumed ambient), and is the volume fraction of (liquid) water in the gas
stream. The sensitivity of the detector is characterized by the value of RTI. The amount of heat conducted
away from the link by the mount is indicated by the C-Factor, C. The RTI and C-Factor are determined
experimentally. The constant C
2
has been empirically determined by DiMarzo and co-workers [38, 39, 40]
to be 610
6
K/(m/s)
1
2
, and its value is relatively constant for different types of sprinklers.
The algorithm for heat detector activation is exactly the same as for sprinkler activation, except there
is no accounting for conductive losses or droplet cooling. Note that neither the sprinkler nor heat detector
models account for thermal radiation.
3.5.2 Heat Detectors
As far as FDS is concerned, a heat detector is just a sprinkler with no water spray. In other words, the
activation of a heat detector is governed by Eq. (3.106), but with just the rst term on the right hand side:
dT
l
dt
=
_
|u|
RTI
(T
g
T
l
) (3.107)
Both the RTI and activation temperature are determined empirically.
3.5.3 Smoke Detectors
An informative discussion of the issues associated with smoke detection can be found in the SFPE Handbook
chapter Design of Detection Systems, by Schiliti, Meacham and Custer [21]. The authors point out that
the difculty in modeling smoke detector activation stems from a number of issues: (1) the production and
transport of smoke in the early stage of a re are not well-understood, (2) detectors often use complex
response algorithms rather than simple threshold or rate-of-change criteria, (3) detectors can be sensitive to
smoke particle number density, size distribution, refractive index, composition, etc., and (4) most computer
models, including FDS, do not provide detailed descriptions of the smoke besides its bulk transport. This
last point is the most important. At best, in its present form, FDS can only provide to an activation algorithm
the velocity and smoke concentration of the ceiling jet owing past the detector. Regardless of the detailed
mechanism within the device, any activation model included within FDS can only account for the entry
resistance of the smoke due to the geometry of the detector. Issues related to the effectiveness of ionization
or photoelectric detectors cannot be addressed by FDS.
Consider the simple idealization of a spot-type smoke detector. A disk-shaped cover lined with a ne
mesh screen forms the external housing of the device, which is usually mounted to the ceiling. Somewhere
38
within the device is a relatively small sensing chamber where the smoke is actually detected in some way.
A simple model of this device has been proposed by Heskestad [21]. He suggested that the mass fraction of
smoke in the sensing chamber of the detector Y
c
lags behind the mass fraction in the external free stream Y
e
by a time period t = L/u, where u is the free stream velocity and L is a length characteristic of the detector
geometry. The change in the mass fraction of smoke in the sensing chamber can be found by solving the
following equation:
dY
c
dt
=
Y
e
(t) Y
c
(t)
L/u
(3.108)
The detector activates when Y
c
rises above a detector-specic threshold.
A more detailed model of smoke detection involving two lling times rather than one has also been
proposed. Smoke passing into the sensing chamber must rst pass through the exterior housing, then it must
pass through a series of bafes before arriving at the sensing chamber. There is a time lag associated with
the passing of the smoke through the housing and also the entry of the smoke into the sensing chamber. Let
t
e
be the characteristic lling time of the entire volume enclosed by the external housing. Let t
c
be the
characteristic lling time of the sensing chamber. Cleary et al. [41] suggested that each characteristic lling
time is a function of the free-stream velocity u outside the detector
t
e
=
e
u
e
; t
c
=
c
u
c
(3.109)
The s and s are empirical constants related to the specic detector geometry The change in the mass
fraction of smoke in the sensing chamber Y
c
can be found by solving the following equation:
dY
c
dt
=
Y
e
(t t
e
) Y
c
(t)
t
c
(3.110)
where Y
e
is the mass fraction of smoke outside of the detector in the free-stream. A simple interpretation of
the equation is that the concentration of the smoke that enters the sensing chamber at time t is that of the
free-stream at time t t
e
.
An analytical solution for Eq. (3.110) can be found, but it is more convenient to simply integrate it
numerically as is done for sprinklers and heat detectors. Then, the predicted mass fraction of smoke in the
sensing chamber, Y
c
(t), can be converted into an expression for the percent obscuration per unit length by
computing:
_
1e
Y
c
l
_
100 (3.111)
where is the specic extinction coefcient, is the density of the external gases in the ceiling jet, and
l is the preferred unit of length (typically 1 m or 1 ft). For most aming fuels, a suggested value for is
8700 m
2
/kg 1100 m
2
/kg at a wavelength of 633 nm [42].
The SFPE Handbook has references to various works on smoke detection and suggested values for the
characteristic length L. FDS includes the one parameter Heskestad model as a special case of the four param-
eter Cleary model. For the Cleary model, one must set
e
,
e
,
c
, and
c
, whereas for the Heskestad model
only L =
c
needs to be prescribed. Eq. (3.110) is still used, with
e
= 0 and
e
=
c
= 1. Proponents
of the four-parameter model claim that the two lling times are needed to better capture the behavior of
detectors in a very slow free-stream (u < 0.5 m/s). Rather than declaring one model better than another, the
algorithm included in FDS allows the user to pick these various parameters, and in so doing, pick whichever
model the user feels is appropriate [43].
Additionally, FDS can model the behavior of beam and aspiration smoke detectors. For a beam detector
the user species the emitter and receiver positions and the total obscuration at which the detector will alarm.
FDS will then integrate the obscuration over the path length using the predicted soot concentration in each
grid cell along the path. For an aspiration detector the user species the sampling locations, the ow rate at
39
each location, the transport time from each sampling point to the detector, the ow rate of any bypass ow,
and the total obscuration at which the detector will alarm. FDS will compute that soot concentration at the
detector by weighting the predicted soot concentrations at the sampling locations with their ow rates after
applying the appropriate time delay.
40
3.6 Liquid Sprays
There are several uses for liquid sprays within FDS, most notably for sprinklers, but also for liquid fuel
sprays. The basic physics is the same in each case, but the liquid properties are different. Also, simulating
the effects of a sprinkler spray involves a number of elements beyond just activation: computing the droplet
trajectories and tracking the water as it drips onto the burning surface.
3.6.1 Droplet Size Distribution
A spray consists of a sampled set of spherical droplets. The size distribution is expressed in terms of its Cu-
mulative Volume Fraction (CVF), a function that relates the fraction of the liquid volume (mass) transported
by droplets less than a given diameter. Researchers at Factory Mutual have suggested that the CVF for a
sprinkler may be represented by a combination of log-normal and Rosin-Rammler distributions [44]
F(d) =
_
_
_
1
d
0
1
d
[ln(d
/dm)]
2
2
2
dd
(d d
m
)
1e
0.693(
d
dm
)
(d
m
< d)
(3.112)
where d
m
is the median droplet diameter (i.e. half the mass is carried by droplets with diameters of d
m
or
less), and and are empirical constants equal to about 2.4 and 0.6, respectively.
6
The median droplet
diameter is a function of the sprinkler orice diameter, operating pressure, and geometry. Research at
Factory Mutual has yielded a correlation for the median droplet diameter [45]
d
m
D
We
1
3
(3.113)
where D is the orice diameter of the sprinkler. The Weber number, the ratio of inertial forces to surface
tension forces, is given by
We =
d
u
2
d
D
d
(3.114)
where
d
is the density of liquid, u
d
is the discharge velocity, and
d
is the liquid surface tension (72.810
3
N/m at 20
C for water). The discharge velocity can be computed from the mass ow rate, a function of the
sprinklers operating pressure and orice coefcient known as the K-Factor. FM reports that the constant
of proportionality in Eq. (3.113) appears to be independent of ow rate and operating pressure. Three
different sprinklers were tested in their study with orice diameters of 16.3 mm, 13.5 mm, 12.7 mm and the
constants were approximately 4.3, 2.9, 2.3, respectively. The strike plates of the two smaller sprinklers were
notched, while that of the largest sprinkler was not [45].
In the numerical algorithm, the size of the droplets are chosen to mimic the Rosin-Rammler/log-normal
distribution. A Probability Density Function (PDF) for the droplet diameter is dened
f (d) =
F
(d)
d
3
_
0
F
(d
)
d
3
dd
; F
dF
dd
(3.115)
Droplet diameters are randomly selected by equating the Cumulative Number Fraction of the droplet distri-
bution with a uniformly distributed random variable U
U(d) =
d
0
f (d
)dd
(3.116)
Figure 3.3 displays typical Cumulative Volume Fraction and Cumulative Number Fraction functions.
6
The Rosin-Rammler and log-normal distributions are smoothly joined if = 2/(
2(ln 2) ) = 1.15/ .
41
Figure 3.3: Cumulative Volume Fraction and Cumulative Number Fraction functions of the droplet size
distribution from a typical industrial-scale sprinkler. The median diameter d
m
is 1 mm, = 0.6 and = 2.4.
3.6.2 Droplet Transport in the Gas Phase
Every droplet from a given sprinkler or nozzle is not tracked. Instead, a sampled set of the droplets is
tracked. The procedure for selecting droplet sizes is as follows: Suppose the mass ow rate of the liquid
is m. Suppose also that the time interval for droplet insertion into the numerical simulation is t, and
the number of droplets inserted each time interval is N. Choose N uniformly distributed random numbers
between 0 and 1, call them U
i
, obtain N droplet diameters d
i
based on the given droplet size distribution,
Eq. (3.116), and then compute a weighting constant C from the mass balance
m t =C
N
i=1
4
3
w
_
d
i
2
_
3
(3.117)
The mass and heat transferred from each droplet will be multiplied by the weighting factor C.
For a spray, the force term f
b
in Eq. (3.2) represents the momentum transferred from the droplets to the
gas. It is obtained by summing the force transferred from each droplet in a grid cell and dividing by the cell
volume
f
b
=
1
2
C
D
r
2
d
(u
d
u)|u
d
u|
xyz
(3.118)
where C
D
is the drag coefcient, r
d
is the droplet radius, u
d
is the velocity of the droplet, u is the velocity of
the gas, is the density of the gas, and xyz is the volume of the grid cell. The trajectory of an individual
droplet is governed by the equation
d
dt
(m
d
u
d
) = m
d
g
1
2
C
D
r
2
d
(u
d
u)|u
d
u| (3.119)
where m
d
is the mass of the droplet. The drag coefcient is a function of the local Reynolds number (based
on droplet diameter)
C
D
=
_
_
_
24/Re
D
Re
D
< 1
24
_
1+0.15Re
0.687
D
_
/Re
D
1 < Re
D
< 1000
0.44 1000 < Re
D
(3.120)
42
Re
D
=
|u
d
u| 2r
d
(T)
(3.121)
where (T) is the dynamic viscosity of air at temperature T.
3.6.3 Droplet Transport on a Surface
When a liquid droplet hits a solid horizontal surface, it is assigned a random horizontal direction and moves
at a xed velocity until it reaches the edge, at which point it drops straight down at the same xed velocity.
This dripping velocity has been measured for water to be on the order of 0.5 m/s [46, 47]. While attached
to a surface, the droplet is assumed to form a thin lm of liquid that transfers heat to the solid, and heat
and mass to the gas. Details are included in the next section.
3.6.4 Absorption and Scattering of Thermal Radiation by Droplets
The attenuation of thermal radiation by liquid droplets is an important consideration, especially for water
mist systems [48]. Liquid droplets attenuate thermal radiation through a combination of scattering and
absorption [49]. The radiation-droplet interaction must therefore be solved for both the accurate prediction
of the radiation eld and for the droplet energy balance.
If the gas phase absorption and emission in Eq. (3.67) are temporarily neglected for simplicity, the
radiative transport equation becomes
s I
(x, s) =[
d
(x, ) +
d
(x, )] I(x, s) +
d
(x, ) I
b,d
(x, ) +
d
(x, )
4
4
(s, s
) I
(x, s
) ds
(3.122)
where
d
is the droplet absorption coefcient,
d
is the droplet scattering coefcient and I
b,d
is the emission
term of the droplets. (s, s
) is a scattering phase function that gives the scattered intensity from direction
s
to s. The local absorption and scattering coefcients are calculated from the local droplet number density
N(x) and mean diameter d
m
(x) as
d
(x, ) = N(x)
0
f (r, d
m
(x)) C
a
(r, ) dr
d
(x, ) = N(x)
0
f (r, d
m
(x)) C
s
(r, ) dr
(3.123)
where r is the droplet radius and C
a
and C
s
are absorption and scattering cross sections, respectively, given
by Mie theory. The droplet number density function f (r, d
m
) is assumed to have the same form as the initial
droplet size distribution, but a mean diameter depending on the location x. For the numerical implementa-
tion, the above equations are written in the form
d
(x, ) = A
d
(x)
0
f (r,d
m
(x)) C
a
(r,)
(d
m
(x)/2)
2
dr
d
(x, ) = A
d
(x)
0
f (r,d
m
(x)) C
s
(r,)
(d
m
(x)/2)
2
dr
(3.124)
where A
d
is the total cross sectional area per unit volume of the droplets. A
d
is approximated as
A
d
d
(x)
2
w
d
m
(x)/3
(3.125)
An accurate computation of the in-scattering integral on the right hand side of Eq. (3.122) would be
extremely time consuming. It is here approximated by dividing the total 4 solid angle to a forward angle
l
and ambient angle
= 4
l
. For compatibility with the FVM solver,
l
is set equal to the
43
control angle given by the angular discretization. However, it is assumed to be symmetric around the center
of the control angle. Within
l
the intensity is I
(x, ) =
U(x, )
l
I
(x, s)
(3.126)
where U(x) is the total integrated intensity. The in-scattering integral can now be written as
d
(x, )
4
4
(s, s
) I
(x, s
) d
=
d
(x, )[
f
I
(x, s) +(1
f
)U
(x)] (3.127)
where
f
=
f
(r, ) is a fraction of the total intensity originally within the solid angle
l
that is scattered
into the same angle
l
. Dening an effective scattering coefcient section
d
(x, ) =
4N(x)
4
l
0
(1
f
) C
s
(r, ) dr (3.128)
the spray RTE becomes
s I
(x, s) =[
d
(x, ) +
d
(x, )] I(x, s) +
d
(x, ) I
b,d
(x, ) +
d
(x, )
4
U(x, ) (3.129)
This equation can be integrated over the spectrum to get the band specic RTEs. The procedure is exactly
the same as that used for the gas phase RTE. After the band integrations, the spray RTE for band n becomes
s I
n
(x, s) =[
d,n
(x) +
d,n
(x)] I
n
(x, s) +
d,n
(x) I
b,d,n
(x) +
d
(x, )
4
U
n
(x) (3.130)
where the source function is based on the average droplet temperature within a cell. The droplet contribution
to the radiative loss term in the energy equation is
q
r
(x)(droplets) =
d
(x) [U(x) 4I
b,d
(x)] (3.131)
For each individual droplet, the radiative heating/cooling power is computed as
q
r
=
m
d
d
(x)
d
(x) [U(x) 4I
b,d
(x)] (3.132)
where m
d
is the mass of the droplet and
d
(x) is the total density of droplets in the cell.
The absorption and scattering cross sections and the scattering phase function are calculated using the
MieV code developed by Wiscombe [50]. Both
d
and
d
are averaged over the possible droplet radii and
wavelength before the actual simulation. A constant radiation temperature, T
rad
, is used in the wavelength
averaging. T
rad
should be selected to represent a typical radiating ame temperature. A value of 1173 K
is used by default. The averaged quantities, now functions of the droplet mean diameter only, are saved in
one-dimensional arrays. During the simulation, the local properties are calculated as a table look-up using
the local mean droplet diameter [51]. Details of the computation are given in Section 4.5.1.
3.6.5 Heating and Evaporation of Liquid Droplets
Liquid droplets are represented either as discrete airborne spheres propelled through the gas, or as rectan-
gular blocks that collectively form a thin liquid lm on solid objects. These droplets are still individually
tracked as lagrangian particles, but the heat and mass transfer coefcients are different. In the discussion to
follow, the term droplets will be used to describe either form.
44
Over the course of a time step of the gas phase solver, the droplets in a given grid cell evaporate as a
function of the liquid equilibrium vapor mass fraction, Y
l
, the local gas phase vapor mass fraction, Y
g
, the
(assumed uniform) liquid temperature, T
l
, and the local gas temperature, T
g
. If the droplet is attached to a
surface, T
s
is the solid temperature. The mass and energy transfer between the gas and the liquid can be
described by the following set of equations [52]
dm
l
dt
= A h
m
(Y
l
Y
g
) (3.133)
m
l
c
l
dT
l
dt
= Ah(T
g
T
l
) +Ah
s
(T
s
T
l
) + q
r
+
dm
l
dt
h
v
(3.134)
Here, m
l
is the mass of the liquid droplet (or that fraction of the surface lm associated with the formerly
airborne droplet), A is the surface area of the liquid droplet (or that fraction of the lm exposed to the gas
and to the wall), h
m
is the mass transfer coefcient to be discussed below, is the gas density, c
l
is the
liquid specic heat, h is the heat transfer coefcient between the liquid and the gas, h
s
is the heat transfer
coefcient between the liquid and the solid surface, q
r
is the rate of radiative heating of the droplet, and h
v
is the latent heat of vaporization of the liquid. The vapor mass fraction of the gas, Y
g
, is obtained from the
gas phase mass conservation equations, and the liquid equilibrium vapor mass fraction is obtained from the
Clausius-Clapeyron equation
X
l
= exp
_
h
v
W
l
R
_
1
T
b
1
T
l
__
; Y
l
=
X
l
X
l
(1W
a
/W
l
) +W
a
/W
l
(3.135)
where X
d
is the equilibrium vapor volume fraction, W
l
is the molecular weight of the evaporated liquid, W
a
is
the molecular weight of air, R is the universal gas constant, and T
b
is the boiling temperature of the liquid.
Mass and heat transfer between liquid and gas are described with analogous empirical correlations. The
mass transfer coefcient, h
m
, is described by the empirical relationships [19]:
h
m
=
Sh D
lg
L
; Sh =
_
_
_
2+0.6 Re
1
2
D
Sc
1
3
droplet
0.037 Re
4
5
L
Sc
1
3
lm
(3.136)
Sh is the Sherwood number, D
lg
is the binary diffusion coefcient between the liquid vapor and the sur-
rounding gas (usually assumed air), L is a length scale equal to either the droplet diameter or 1 m for a
surface lm, Re
D
is the Reynolds number of the droplet (based on the diameter, D, and the relative air-
droplet velocity), Re
L
is the Reynolds number based on the length scale L, and Sc is the Schmidt number
(/D
lg
, assumed 0.6 for all cases).
An analogous relationship exists for the heat transfer coefcient:
h =
Nu k
L
; Nu =
_
_
_
2+0.6 Re
1
2
D
Pr
1
3
droplet
0.037 Re
4
5
L
Pr
1
3
lm
(3.137)
Nu is the Nusselt number, k is the thermal conductivity of the gas, and Pr is the Prandtl number (assumed
0.7 for all cases).
3.6.6 Fire Suppression by Water
The previous two sections describe heat transfer from a droplet of water to a hot gas, a hot solid, or both.
Although there is some uncertainty in the values of the respective heat transfer coefcients, the fundamental
45
physics are fairly well understood. However, when the water droplets encounter burning surfaces, simple
heat transfer correlations become more difcult to apply. The reason for this is that the water is not only
cooling the surface and the surrounding gas, but it is also changing the pyrolysis rate of the fuel. If the
surface of the fuel is planar, it is possible to characterize the decrease in the pyrolysis rate as a function of
the decrease in the total heat feedback to the surface. Unfortunately, most fuels of interest in re applications
are multi-component solids with complex geometry at scales unresolvable by the computational grid.
To date, most of the work in this area has been performed at Factory Mutual. An important paper on
the subject is by Yu et al. [53]. The authors consider dozens of rack storage commodity res of different
geometries and water application rates, and characterize the suppression rates in terms of a few global
parameters. Their analysis yields an expression for the total heat release rate from a rack storage re after
sprinkler activation
Q =
Q
0
e
k(tt
0
)
(3.138)
where
Q
0
is the total heat release rate at the time of application t
0
, and k is a fuel-dependent constant. For
the FMRC Standard Plastic commodity k is given as
k = 0.716 m
w
0.0131 s
1
(3.139)
where m
w
is the ow rate of water impinging on the box tops, divided by the area of exposed surface (top
and sides). It is expressed in units of kg/m
2
/s. For the Class II commodity, k is given as
k = 0.536 m
w
0.0040 s
1
(3.140)
Unfortunately, this analysis is based on global water ow and burning rates. Equation (3.138) accounts
for both the cooling of non-burning surfaces as well as the decrease in heat release rate of burning surfaces.
In the FDS model, the cooling of unburned surfaces and the reduction in the heat release rate are computed
locally. Thus, it is awkward to apply a global suppression rule. However, the exponential nature of suppres-
sion by water is observed both locally and globally, thus it is assumed that the local burning rate of the fuel
can be expressed in the form [46, 47]
m
f
(t) = m
f ,0
(t) e
k(t)dt
(3.141)
Here m
f ,0
(t) is the burning rate per unit area of the fuel when no water is applied and k(t) is a linear function
of the local water mass per unit area, m
w
, expressed in units of kg/m
2
,
k(t) = a m
w
(t) s
1
(3.142)
Note that a is an empirical constant.
46
Chapter 4
Numerical Method
This chapter presents a description of the numerical methods used to solve the equations of the previous
chapter, with sections in roughly the same order.
4.1 Hydrodynamic Model
Each of the conservation equations emphasizes the importance of the velocity divergence and vorticity elds,
as well as the close relationship between the thermally expandable uid equations [10] and the Boussinesq
equations for which the authors have developed highly efcient solution procedures [54, 55]. All spatial
derivatives are approximated by second-order nite differences and the ow variables are updated in time
using an explicit second-order predictor-corrector scheme.
4.1.1 Simplied Equations
Conservation of Mass
t
+u = u (4.1)
Conservation of Species
Y
t
+u Y
=Y
u+ D
+ m
(4.2)
Conservation of Momentum
u
t
+u+H =
1
((
0
)g+f
b
+
i j
) (4.3)
Pressure Equation
2
H =
( u)
t
F ; F = u
1
((
m
)gf
b
i j
) (4.4)
Equation of State
p
m
(z, t) = TR
/W
(4.5)
Notice that the source terms from the energy conservation equation have been incorporated into the diver-
gence and ultimately are involved in the mass conservation equation. The temperature is found from the
density and background pressure via the equation of state.
47
Temporal Discretization
All calculations start with ambient initial conditions. At the beginning of each time step, the density,
n
,
species mass fractions, Y
n
, velocity, u
n
, modied pressure, H
n
, and background pressure in zone m, p
n
m
are known. All other quantities can be derived from them. Note that the superscript (n +1)
e
refers to an
estimate of the value of the quantities at the (n+1)st time step.
1. The thermodynamic quantities , Y
, and p
m
are estimated at the next time step with an explicit Euler
step. For example, the density is estimated by
(n+1)
e
=
n
t(u
n
n
+
n
u
n
) (4.6)
2. The divergence ( u)
(n+1)
e
is formed from these estimated thermodynamic quantities. The normal
velocity components at boundaries that are needed to form the divergence are assumed known.
3. A Poisson equation for the pressure is solved with a direct solver
2
H
n
=
( u)
(n+1)
e
( u)
n
t
F
n
(4.7)
Note that the vector F contains the convective, diffusive and force terms of the momentum equation.
These will be described in detail below.
4. The velocity is estimated at the next time step
u
(n+1)
e
= u
n
t (F
n
+H
n
) (4.8)
Note that the divergence of the estimated velocity eld is identically equal to the estimated divergence
( u)
(n+1)
e
that was derived from the estimated thermodynamic quantities. The time step is checked
at this point to ensure that
t max
_
|u|
x
,
|v|
y
,
|w|
z
_
< 1 ; 2 t
_
1
x
2
+
1
y
2
+
1
z
2
_
< 1 (4.9)
If the time step is too large, it is reduced so that it satises both constraints and the procedure returns
to the beginning of the time step. If the time step satises the stability criteria, the procedure continues
to the corrector step. See Section 4.1.3 for more details on stability.
5. The thermodynamic quantities , Y
, and p
m
are corrected at the next time step. For example, the
density is corrected
n+1
=
1
2
_
n
+
(n+1)
e
t(u
(n+1)
e
(n+1)
e
+
(n+1)
e
u
(n+1)
e
)
_
(4.10)
6. The corrected divergence ( u)
(n+1)
is computed from the corrected thermodynamic quantities.
7. The pressure is recomputed using estimated quantities
2
H
(n+1)
e
=
2( u)
n+1
( u)
(n+1)
e
( u)
n
t
F
(n+1)
e
(4.11)
8. The velocity is corrected
u
n+1
=
1
2
_
u
n
+u
(n+1)
e
t
_
F
(n+1)
e
+H
(n+1)
e
__
(4.12)
Note again that the divergence of the corrected velocity eld is identically equal to the corrected
divergence.
48
Spatial Discretization
Spatial derivatives in the governing equations are written as second-order accurate nite differences on a
rectilinear grid. The overall domain is a rectangular box that is divided into rectangular grid cells. Each cell
is assigned indices i, j and k representing the position of the cell in the x, y and z directions, respectively.
Scalar quantities are assigned in the center of each grid cell; thus,
n
i jk
is the density at the nth time step
in the center of the cell whose indices are i, j and k. Vector quantities like velocity are assigned at their
appropriate cell faces. For example, u
n
i jk
is the x-component of velocity at the positive-oriented face of the
i jkth cell; u
n
i1, jk
is dened at the negative-oriented face of the same cell.
49
4.1.2 The Mass Transport Equations and the Divergence
Due to the use of the low Mach number approximation, the mass and energy equations are combined through
the divergence. The divergence of the ow eld contains many of the re-specic source terms described
above.
Discretizing the Convective and Diffusive Transport Terms
The density at the center of the i jkth cell is updated in time with the following predictor-corrector scheme.
In the predictor step, the density at the (n+1)st time level is estimated based on information at the nth level
(n+1)
e
i jk
n
i jk
t
+(u )
n
i jk
=
n
i jk
( u)
n
i jk
(4.13)
Following the prediction of the velocity and background pressure at the (n +1)st time level, the density is
corrected
(n+1)
i jk
1
2
_
n
i jk
+
(n+1)
e
i jk
_
1
2
t
+(u )
(n+1)
e
i jk
=
(n+1)
e
i jk
( u)
(n+1)
e
i jk
(4.14)
The species conservation equations are differenced the same way, with the addition of diffusion and source
terms:
(Y
)
(n+1)
e
i jk
(Y
)
n
i jk
t
+ = +( D
)
n
i jk
+ m
,i jk
(4.15)
at the predictor step, and
(Y
)
(n+1)
i jk
1
2
_
(Y
)
n
i jk
+(Y
)
(n+1)
e
i jk
_
1
2
t
+ = +( D
)
(n+1)
e
i jk
+ m
,i jk
(4.16)
at the corrector step.
The convective terms are written as upwind-biased differences in the predictor step and downwind-
biased differences in the corrector step [56]. In the denition to follow, the symbol means + in the
predictor step and in the corrector step. The opposite is true for .
(u )
i jk
1
u
2
u
i jk
i+1, jk
i jk
x
+
1
u
2
u
i1, jk
i jk
i1, jk
x
+
1
v
2
v
i jk
i, j+1,k
i jk
y
+
1
v
2
v
i, j1,k
i jk
i, j1,k
y
+
1
w
2
w
i jk
i j,k+1
i jk
z
+
1
w
2
w
i j,k1
i jk
i j,k1
z
(4.17)
The convective term in the species transport equation, (u Y
)
i jk
is differenced the exact same way. Note
that without the inclusion of the s, these are simple central difference approximations. The s are local
Courant numbers,
u
= ut/x,
v
= vt/y, and
w
= wt/z, where the velocity components are those
that immediately follow. Their role is to bias the differencing upwind at the predictor step. Where the
local Courant number is near unity, the difference becomes nearly fully upwinded. Where the local Courant
number is much less than unity, the differencing is more centralized.
50
Flux Correction
The second-order nite-differencing scheme used in FDS cannot fully resolve sharp gradients on a relatively
coarse grid. Instead, steep gradients cause local over-shoots and under-shoots of quantities like temperature,
density and species mass fraction. For mass fraction in particular, this can result in a solution where the mass
fraction exceeds its permissible limits (i.e. the numerical method transports into or out of a grid cell more
mass than is physically possible). The overall numerical scheme is still mass conserving, but non-physical in
regions of high gradients. This problem can be reduced, but not eliminated, by using higher order numerical
methods. These methods are, however, more expensive. Another solution is to perform a ux transport
correction. This involves examining the solution and locating regions where a non-physical solution exists
and then redistributing mass to correct it. Typically this results in some increased numerical diffusion;
however, this is partially mitigated since at any one time step, the correction is applied to a small number of
cells.
The ux correction scheme is performed in both the predictor and corrector steps after updating the
species mass fractions, Y
. For each species, two loops are performed over each computational mesh. The
rst loop searches for and corrects under-shoots, and the second loop searches for and corrects over-shoots.
An under-shoot occurs if the mass fraction of a species is less than its permissible minimum, typically 0, or
if there was an outux of that species in the prior time step, Y
,i jk
< 0 , and the new species mass fraction
is less than all of its surrounding cells. An over-shoot occurs if the mass fraction of a species is greater than
its permissible maximum, or if there was an inux of that species in the prior time step, Y
,i jk
> 0 , and
the new species mass fraction is greater than all of its surrounding cells. In each loop a temporary array is
used to store the corrected values which are then applied globally at the end of each loop. Using a temporary
array rather than a cell by cell immediate correction ensures that cells requiring correction are not bypassed
due to the sweep direction of the loop. Under-shoots are corrected rst because in a typical mixture fraction
computation it is more likely to have an under-shoot with a value less than the absolute minimum for the
species than it is for an over-shoot to exceed the absolute maximum for the species.
Discretizing the Divergence
The divergence (see Eq. (3.17)) in the mth pressure zone in both the predictor and corrector step is discretized
( u)
i jk
=
R
Wc
p
p
m
_
q
i jk
+( kT)
i jk
+. . .
_
+
1
p
n
_
R
Wc
p
1
__
p
m
t
w
i jk
0,k
g
_
(4.18)
The thermal and material diffusion terms are pure central differences, with no upwind or downwind bias,
thus they are differenced the same way in both the predictor and corrector steps. For example, the thermal
conduction term is differenced as follows:
( kT)
i jk
=
1
x
_
k
i+
1
2
, jk
T
i+1, jk
T
i jk
x
k
i
1
2
, jk
T
i jk
T
i1, jk
x
_
+
1
y
_
k
i, j+
1
2
,k
T
i, j+1,k
T
i jk
y
k
i, j
1
2
,k
T
i jk
T
i, j1,k
y
_
+
1
z
_
k
i j,k+
1
2
T
i j,k+1
T
i jk
z
k
i j,k
1
2
T
i jk
T
i j,k1
z
_
(4.19)
The temperature is extracted from the density via the equation of state
T
i jk
=
p
m
i jk
R
N
s
l=0
(Y
,i jk
/W
)
(4.20)
51
Because only species 1 through N
s
are explicitly computed, the summation is rewritten
W
N
s
=0
Y
,i jk
W
=
1
W
0
+
N
s
=1
_
1
W
1
W
0
_
Y
(4.21)
In isothermal calculations involving multiple species, the density can be extracted from the average molec-
ular weight
i jk
=
p
m
T
RW
=
W
0
p
m
T
R
+
N
s
=1
_
1
W
0
W
_
(Y
)
i jk
(4.22)
52
4.1.3 The Momentum Equation
The three components of the momentum equation are
u
t
+F
x
+
H
x
= 0 ; F
x
= w
y
v
z
_
f
x
+
xx
x
+
xy
y
+
xz
z
_
(4.23)
v
t
+F
y
+
H
y
= 0 ; F
y
= u
z
w
x
_
f
y
+
yx
x
+
yy
y
+
yz
z
_
(4.24)
w
t
+F
z
+
H
z
= 0 ; F
z
= v
x
u
y
_
f
z
+
zx
x
+
zy
y
+
zz
z
_
(4.25)
The spatial discretization of the momentum equations takes the form
u
t
+F
x,i jk
+
H
i+1, jk
H
i jk
x
= 0 (4.26)
v
t
+F
y,i jk
+
H
i, j+1,k
H
i jk
y
= 0 (4.27)
w
t
+F
z,i jk
+
H
i j,k+1
H
i jk
z
= 0 (4.28)
where H
i jk
is taken at center of cell i jk, u
i jk
and F
x,i jk
are taken at the side of the cell facing in the forward x
direction, v
i jk
and F
y,i jk
at the side facing in the forward y direction, and w
i jk
and F
z,i jk
at the side facing in
the forward z (vertical) direction. In the denitions to follow, the components of the vorticity (
x
,
y
,
z
) are
located at cell edges pointing in the x, y and z directions, respectively. The same is true for the off-diagonal
terms of the viscous stress tensor:
zy
=
yz
,
xz
=
zx
, and
xy
=
yx
. The diagonal components of the stress
tensor
xx
,
yy
, and
zz
; the external force components ( f
x
, f
y
, f
z
); and the Courant numbers
u
,
v
, and
w
are
located at their respective cell faces.
F
x,i jk
=
_
1
w
2
w
i+
1
2
, jk
y,i jk
+
1
w
2
w
i+
1
2
, j,k1
y,i j,k1
_
_
1
v
2
v
i+
1
2
, jk
z,i jk
+
1
v
2
v
i+
1
2
, j1,k
z,i, j1,k
_
i+
1
2
, jk
_
f
x,i jk
+
xx,i+1, jk
xx,i jk
x
+
xy,i jk
xy,i, j1,k
y
+
xz,i jk
xz,i, j,k1
z
_
(4.29)
F
y,i jk
=
_
1
u
2
u
i, j+
1
2
,k
z,i jk
+
1
u
2
u
i1, j+
1
2
,k
z,i1, jk
_
_
1
w
2
w
i, j+
1
2
,k
x,i jk
+
1
w
2
w
i, j+
1
2
,k1
x,i j,k1
_
i, j+
1
2
,k
_
f
y,i jk
+
yx,i jk
yx,i1, jk
x
+
yy,i, j+1,k
yy,i jk
y
+
yz,i jk
yz,i, j,k1
z
_
(4.30)
F
z,i jk
=
_
1
v
2
v
i j,k+
1
2
x,i jk
+
1
v
2
v
i, j1,k+
1
2
x,i, j1,k
_
_
1
u
2
u
i j,k+
1
2
y,i jk
+
1
u
2
u
i1, j,k+
1
2
y,i1, jk
_
i j,k+
1
2
_
f
z,i jk
+
zx,i jk
zx,i1, jk
x
+
zy,i jk
zy,i, j1,k
y
+
zz,i j,k+1
zz,i jk
z
_
(4.31)
x,i jk
=
w
i, j+1,k
w
i jk
y
v
i j,k+1
v
i jk
z
(4.32)
53
y,i jk
=
u
i j,k+1
u
i jk
z
w
i+1, jk
w
i jk
x
(4.33)
z,i jk
=
v
i+1, jk
v
i jk
x
u
i, j+1,k
u
i jk
y
(4.34)
xx,i jk
=
i jk
_
4
3
( u)
i jk
2
v
i jk
v
i, j1,k
y
2
w
i jk
w
i j,k1
z
_
(4.35)
yy,i jk
=
i jk
_
4
3
( u)
i jk
2
u
i jk
u
i1, jk
x
2
w
i jk
w
i j,k1
z
_
(4.36)
zz,i jk
=
i jk
_
4
3
( u)
i jk
2
u
i jk
u
i1, jk
x
2
v
i jk
v
i, j1,k
y
_
(4.37)
xy,i jk
=
yx,i jk
=
i+
1
2
, j+
1
2
,k
_
u
i, j+1,k
u
i jk
y
+
v
i+1, jk
v
i jk
x
_
(4.38)
xz,i jk
=
zx,i jk
=
i+
1
2
, j,k+
1
2
_
u
i j,k+1
u
i jk
z
+
w
i+1, jk
w
i jk
x
_
(4.39)
yz,i jk
=
zy,i jk
=
i, j+
1
2
,k+
1
2
_
v
i j,k+1
v
i jk
z
+
w
i, j+1,k
w
i jk
y
_
(4.40)
u
=
ut
x
(4.41)
v
=
vt
y
(4.42)
w
=
wt
z
(4.43)
The variables
u
,
v
and
w
are local Courant numbers evaluated at the same locations as the velocity com-
ponent immediately following them, and serve to bias the differencing of the convective terms in the upwind
direction. The subscript i +
1
2
indicates that a variable is an average of its values at the ith and the (i +1)th
cell. By construction, the divergence dened in Eq. (4.18) is identically equal to the divergence dened by
( u)
i jk
=
u
i jk
u
i1, jk
x
+
v
i jk
v
i, j1,k
y
+
w
i jk
w
i j,k1
z
(4.44)
The equivalence of the two denitions of the divergence is a result of the form of the discretized equations,
the time-stepping scheme, and the direct solution of the Poisson equation for the pressure.
Viscous Terms (LES)
The major difference between an LES and a DNS calculation is the form of the viscosity, and the thermal
and material diffusivities. For a Large Eddy Simulation, the dynamic viscosity is dened at cell centers
i jk
=
i jk
(C
s
)
2
|S| (4.45)
where C
s
is an empirical constant, = (xyz)
1
3
, and
|S|
2
=2
_
u
x
_
2
+2
_
v
y
_
2
+2
_
w
z
_
2
+
_
u
y
+
v
x
_
2
+
_
u
z
+
w
x
_
2
+
_
v
z
+
w
y
_
2
2
3
( u)
2
(4.46)
The quantity |S| consists of second order spatial differences averaged at cell centers. For example
u
x
u
i jk
u
i1, jk
x
i
(4.47)
u
y
1
2
_
u
i, j+1,k
u
i jk
y
j+
1
2
+
u
i jk
u
i, j1,k
y
j
1
2
_
(4.48)
54
The divergence is described in Section 4.1.2.
The thermal conductivity and material diffusivity of the uid are related to the viscosity by
k
i jk
=
c
p,0
i jk
Pr
t
; (D)
i jk
=
i jk
Sc
t
(4.49)
where Pr
t
is the turbulent Prandtl number and Sc
t
is the turbulent Schmidt number, both assumed constant.
Note that the specic heat c
p,0
is that of the dominant species of the mixture. Based on simulations of
smoke plumes, C
s
is 0.20, Pr
t
and Sc
t
are 0.5. There are no rigorous justications for these choices other
than through comparison with experimental data [57].
Viscous Terms (DNS)
The dynamic viscosity, thermal conductivity and diffusion coefcients for a DNS calculation are dened at
cell centers
i jk
=
l
Y
l,i jk
l
(T
i jk
) (4.50)
k
i jk
=
l
Y
l,i jk
k
l
(T
i jk
) (4.51)
D
l,i jk
= D
l0
(T
i jk
) (4.52)
where the values for each individual species are approximated from kinetic theory [32]. The term D
l0
is
the binary diffusion coefcient for species l diffusing into the predominant species 0, usually nitrogen. It
is often the case that the numerical grid is too coarse to resolve steep gradients in ow quantities when the
temperature is near ambient. However, as the temperature increases and the diffusion coefcients increase
in value, the situation improves. As a consequence, there is a provision in the numerical algorithm to place
a lower bound on the viscous coefcients to avoid numerical instabilities at temperatures close to ambient.
Velocity Boundary Conditions at a Solid Surface
At solid boundaries, the tangential components of velocity need to be set inside the solid in order to establish
a velocity gradient for the numerical velocity eld. As an example, consider a solid surface forming the
lower boundary of the computational domain (z = 0). The horizontal velocity components u
i j,1
and v
i j,1
are
computed at the vertical faces of the rst layer of gas phase grid cells above the solid surface. What should
the values of u
i j,0
and v
i j,0
be set to such that the ow near the wall is impeded by the same amount as a real
turbulent ow? Assuming that the numerical grid is too coarse to resolve the boundary layer (of nominal
thickness d), and assuming a turbulent velocity prole
U(z) =
_
U
0
(z/d)
1/7
z < d
U
0
z > d
(4.53)
a simple slip boundary condition can be worked out such that
u
i j,0
= u
i j,1
_
1
d
z
_
; v
i j,0
= v
i j,1
_
1
d
z
_
(4.54)
where z is the height of the grid cell. For coarse grids, the boundary condition is very similar to a free-slip
condition because d << z.
Past and present versions of FDS have employed a default boundary condition for LES calculations
where u
i j,0
and v
i j,0
are set to one-half the values of u
i j,1
and v
i j,1
, respectively. Although from the discussion
55
above this implies that the thickness of the boundary layer d is assumed to be half that of the grid cell z, in
practice, it really implies that real walls are covered with small obstructions that are half a cell thick or less
in size, in which case they do not appear explicitly in the simulation. The half-slip condition is an attempt
to account for this wall clutter.
In a DNS calculation, this discussion is moot. The inner-wall ghost cell velocities are set so that the
velocity at the wall surface is zero.
Force Terms
The external force term components are dened at cell faces, just like the velocity components. For example,
the drag force components from sprinkler droplets are:
f
x,i jk
=
1
2
C
D
r
2
d
(u
d
u
i jk
)|u
d
u|
xyz
(4.55)
f
y,i jk
=
1
2
C
D
r
2
d
(v
d
v
i jk
)|u
d
u|
xyz
(4.56)
f
z,i jk
=
1
2
C
D
r
2
d
(w
d
w
i jk
)|u
d
u|
xyz
(4.57)
where r
d
is the radius of a droplet, u = (u
d
, v
d
, w
d
) the velocity of a droplet, C
D
the drag coefcient, and
xyz the volume of the i jkth cell. The summations represent all droplets within a grid cell centered about
the x, y and z faces of a grid cell respectively.
Time Step
The time step is constrained by the convective and diffusive transport speeds via two conditions. The rst is
known as the Courant-Friedrichs-Lewy (CFL) condition:
t max
_
|u
i jk
|
x
,
|v
i jk
|
y
,
|w
i jk
|
z
_
< 1 (4.58)
The estimated velocities u
(n+1)
e
, v
(n+1)
e
and w
(n+1)
e
are tested at each time step to ensure that the CFL
condition is satised. If it is not, then the time step is set to 0.8 of its allowed maximum value and the
estimated velocities are recomputed (and checked again). The CFL condition asserts that the solution of
the equations cannot be updated with a time step larger than that allowing a parcel of uid to cross a grid
cell. For most large-scale calculations where convective transport dominates diffusive, the CFL condition
restricts the time step.
However, in small, nely-gridded domains, a second condition often dominates:
2 max
_
, D,
k
c
p
_
t
_
1
x
2
+
1
dy
2
+
1
z
2
_
< 1 (4.59)
Note that this constraint is applied to the momentum, mass and energy equations via the relevant diffusion
parameter viscosity, material diffusivity or thermal conductivity. This constraint on the time step, often
referred to as the Von Neumann criterion, is typical of any explicit, second-order numerical scheme for
solving a parabolic partial differential equation. To save CPU time, the Von Neumann criterion is only
invoked for DNS calculations or for LES calculations with grid cells smaller than 5 mm.
56
4.1.4 The Pressure Equation
The momentum equation is
u
t
+F+H = 0 ; H =
|u|
2
2
+
p
(4.60)
The divergence operator applied to the three components of the discretized momentum equation (4.26
4.28) yields a single elliptic partial differential equation for the modied pressure, H, known as the Poisson
equation:
H
i+1, jk
2H
i jk
+H
i1, jk
x
2
+
H
i, j+1,k
2H
i jk
+H
i, j1,k
y
2
+
H
i j,k+1
2H
i jk
+H
i j,k1
z
2
=
F
x,i jk
F
x,i1, jk
x
F
y,i jk
F
y,i, j1,k
y
F
z,i jk
F
z,i j,k1
z
t
( u)
i jk
(4.61)
The lack of a superscript implies that all quantities are to be evaluated at the same time level. This elliptic
partial differential equation is solved using a direct (non-iterative) FFT-based solver [58] that is part of a
library of routines for solving elliptic PDEs called CRAYFISHPAK
1
. To ensure that the divergence of the
uid is consistent with the denition given in Eq. (3.17), the time derivative of the divergence is dened
t
( u)
i jk
=
( u)
(n+1)
e
i jk
( u)
n
i jk
t
(4.62)
at the predictor step, and then
t
( u)
i jk
=
2( u)
n+1
i jk
( u)
(n+1)
e
i jk
( u)
n
i jk
t
(4.63)
at the corrector step. The discretization of the divergence was given in Eq. (4.18).
Direct Poisson solvers are most efcient if the domain is a rectangular region, although other geometries
such as cylinders and spheres can be handled almost as easily. For these solvers, a no-ux condition is simple
to prescribe at external boundaries. For example, at the oor, z = 0, the Poisson solver is supplied with the
Neumann boundary condition
H
i j,1
H
i j,0
z
=F
z,i j,0
(4.64)
However, many practical problems involve more complicated geometries. For building res, doors and
windows within multi-room enclosures are very important features of the simulations. These elements may
be included in the overall domain as masked grid cells, but the no-ux condition (4.64) cannot be directly
prescribed at the boundaries of these blocked cells. Fortunately, it is possible to exploit the relatively small
changes in the pressure from one time step to the next to enforce the no-ux condition. At the start of a time
step, the components of the convection/diffusion term F are computed at all cell faces that do not correspond
to walls. At those cell faces that do correspond to solid walls, we prescribe
F
n
=
H
n
+
u
n
t
(4.65)
where F
n
is the normal component of F at the wall.
The asterisk indicates the most recent value of the pressure. The pressure at this particular time step
is not known until the Poisson equation is solved. Equation (4.65) asserts that following the solution of
1
CRAYFISHPAK, a vectorized form of the elliptic equation solver FISHPAK, was originally developed at the National Center
for Atmospheric Research (NCAR) in Boulder, Colorado.
57
the Poisson equation for the pressure, the normal component of velocity u
n
will be driven nearly (but not
exactly) to zero. This is approximate because the true value of the velocity time derivative depends on the
solution of the pressure equation, but since the most recent estimate of pressure is used, the approximation
is very good. Also, even though there are small errors in normal velocity at solid surfaces, the divergence
of each blocked cell remains exactly zero for the duration of the calculation. In other words, the total
ux into a given obstruction is always identically zero, and the error in normal velocity is usually at least
several orders of magnitude smaller than the characteristic ow velocity. When implemented as part of a
predictor-corrector updating scheme, the no-ux condition at solid surfaces is maintained remarkably well.
At open boundaries (say i = I), H is prescribed depending on whether the ow is incoming or outgoing
H
I+
1
2
, jk
= (u
2
I, jk
+v
2
I, j
1
2
,k
+w
2
I, j
1
2
,k
)/2 u
I, jk
> 0
H
I+
1
2
, jk
= 0 u
I, jk
< 0
(4.66)
where I is the index of the last gas phase cell in the x direction and u
I, jk
is the x component of velocity at the
boundary. The value of H in the ghost cell is
H
I+1, jk
= 2H
I+
1
2
, jk
H
I, jk
(4.67)
Multiple Mesh Considerations (On-Going Research)
Solving the Poisson equation on muliple meshes is considerably more difcult than on a single mesh. On a
single mesh, the linear system of equations involves a simple block tri-diagonal matrix for which efcient
solvers have been developed. For multiple mesh calculations, the same solver is used for each mesh as
described above, with the following (Dirichlet) boundary conditions applied at mesh interfaces (in this case,
the interface x = x
max
= x
+
):
H
I+
1
2
, jk
=H
I, jk
x
2
_
u
+
0, jk
u
I, jk
t
+F
x,I jk
_
u
I, jk
> 0
H
I+
1
2
, jk
=H
+
1
2
, jk
u
I, jk
< 0
(4.68)
Here, the term u
+
indicates that the value is from the other mesh. In short, if the ow is incoming, then H
is taken from the other mesh, if the ow is outgoing, then H is chosen so that the value of u at the next time
step is driven towards the value of its counterpart (u
+
) in the other mesh. For abutting meshes of the same
cell size, u
I, jk
and u
+
0, jk
overlap each other, and the goal is to keep their values as close as possible.
This procedure is fairly stable, and provides a reasonable exchange of momentum mesh to mesh. How-
ever, there is no guarantee that the individual solutions of the Poisson equation on each mesh lead to con-
sistent ows into and out of neighboring meshes. Using similar boundary conditions as described above
for open boundaries, it is possible to ensure a continuous solution of the pressure term, H, across mesh
boundaries, but it is not possible to ensure a continuous gradient. Thus, it is not possible to ensure that the
ow out of one mesh is the same as the ow into the next.
The strategy of approximating the solution of the Poisson equation on multiple meshes is as follows.
First, decompose the pressure term H =H
0
+H
. Solve
2
H
0
= F
t
( u) (4.69)
on each mesh, the same way as described above. At this point, each mesh has a pressure eld, but when
the velocity is updated in each mesh, the normal components will not match at mesh interfaces. At the very
least, we need ensure that the volume ux,
V =
u dS, at each mesh interface is consistent mesh to mesh.
What should the volume ux be? There is no way to get it locally. Instead, re-discretize the original Poisson
58
equation over a coarse mesh spanning the entire domain, and apply the divergence theorem to the coarse
grid cells:
m
H dS =
m
F dS
m
( u)
t
dV (4.70)
The subscript m refers to individual cells of the coarse mesh. Introduce a single scalar H
m
for each coarse
grid cell that satises:
H
m+
H
m
x
A
x
+
H
m
H
m
x
A
x
+ =
x
+
F
x
dydz +
F
x
dydz +
m
( u)
t
dV (4.71)
The subscripts m
+
and m
x
+
=
x
+
F
x
dydz
H
m+
H
m
x
A
x
+ (4.72)
We want the ne grid solution of the Poisson equation on the individual abutting meshes to satisfy this
constraint at their interface. To force this, solve
2
H
= 0 (4.73)
with Neumann boundary conditions on each of the abutting meshes such that at the interface the gradient of
H
x
+
=
x
+
_
F
x,I jk
+
H
0
I+1, jk
H
0
I jk
x
_
dydz +
x
+
H
x
dydz (4.74)
A similar constraint must be satised by the mesh to the right. In this way, the resulting volume uxes at the
next time step
V =
u dS into and out of each mesh match. It does not guarantee, however, that the normal
component of velocity at each cell matches its counterpart in the other mesh.
The question now is what values should be assigned to H
, such that the sum of the components equals the mixture fraction. Each component is
tracked via a transport equation, and the conversion of mass from one component to another represents a
reaction step and an associated release of energy.
Two-Parameter Mixture Fraction
When the mixture fraction is divided into two components, Z
1
and Z
2
, there is one chemical reaction that
converts Z
1
to Z
2
. Recall from Section 3.2.1 that this represents single-step combustion with the possibility
of local extinction. From the mixture fraction variables one can determine the amount of fuel and oxygen
present in a grid cell. Ideally, we could use these values in a nite-rate computation to determine the heat
release rate; however, for most computations, the grid resolution is too coarse to resolve the ame. Thus,
ame temperatures will not be realized and a nite-rate computation will not succeed.
Instead, if any grid cell surrounding one containing both fuel and oxygen satises the Burn criteria
depicted in Fig. 3.2, combustion is assumed to occur. The heat release rate is computed as
q
= min
_
max(Y
F
, sY
O
2
)
t
H , q
max
_
; s =
W
F
O
2
W
O
2
(4.75)
where q
max
provides an upper bound of the volumetric heat release rate. It is an empirical parameter that is
tied to the resolution of the underlying grid:
q
max
=
q
max
(xyz)
1
3
; q
max
= 200 kW/m
2
(4.76)
The 200 kW/m
2
value is a user-controlled parameter. It is an estimate of the maximum heat release rate per
unit area of ame sheet.
Once the heat release rate is computed, the mixture fraction variables are updated:
Z
1
n+1
= Z
1
n
t
H
; Z
2
n+1
= Z
2
n
+
q
t
H
(4.77)
Note that the total fuel mass is conserved in this process; Z = Z
1
+Z
2
is still a conserved quantity.
Three-Parameter Mixture Fraction
When the mixture fraction is divided into three components, Z
1
, Z
2
, and Z
3
, there are two chemical reactions
that convert Z
1
to Z
2
and Z
2
to Z
3
. Recall from Section 3.2.1 that this represents two-step combustion
(fuel to CO and CO to CO
2
). The rst step occurs as it does for the two-parameter mixture fraction with
a modied heat of combustion that accounts for the conversion of fuel to CO rather than CO
2
. The second
step is performed for all grid cells that contain CO and O
2
. If q
CO
= min
_
max(Z
2
, sY
O
2
)
t
H
CO
, q
max
q
_
(4.78)
60
If q
=0 after the rst step, then it is presumed that the cell is out of the combustion region (say in the upper
layer of smoke-lled compartment), and a nite-rate reaction computation is performed to convert CO to
CO
2
(see the next section for a discussion of the algorithm for computing a nite-rate reaction). The q
CO
computed using the nite-rate reaction is still limited by q
max
. Once q
CO
is computed the mixture fraction
variables are updated:
Z
2
n+1
= Z
2
n
CO
t
H
CO
; Z
3
n+1
= Z
3
n
+
q
CO
t
H
CO
(4.79)
4.2.2 Heat Release Rate (Finite-Rate Reaction)
For simulations that do not employ a mixture fraction-based combustion model, multi-step, nite-rate reac-
tions are assumed, taking the form:
C
x
H
y
C
x
H
y
+
O
2
O
2
CO
2
CO
2
+
H
2
O
H
2
O (4.80)
It is assumed that the chemical reaction time scale is much shorter than any convective or diffusive transport
time scale. Thus, it makes sense to calculate the consequences of the reaction assuming all other processes
are frozen in a state corresponding to the beginning of the time step. For each grid cell, at the start of a time
step where t = t
n
and Y
n
C
x
H
y
,i jk
/
i jk
W
F
X
F
(t
n
) and Y
n
O
2
,i jk
/
i jk
W
O
2
X
O
2
(t
n
), the following set of ODEs is
solved numerically with a second-order Runge-Kutta scheme
dX
F
dt
= B X
F
(t)
a
X
O
2
(t)
b
e
E/RT
i jk
(4.81)
dX
O
2
dt
=
O
2
F
dX
F
dt
(4.82)
The temperature T
i jk
and density
i jk
are xed at their values at time t
n
and the ODEs are iterated from t
n
to
t
n+1
in about 20 time steps. The pre-exponential factor B, the activation energy E, and the exponents a and
b are input parameters which are in typically assigned the values of
F
and
O
2
. At the end of each sub-time
step the values of X
F
and X
O
2
are updated.
The average heat release rate over the entire time step is given by
q
n
i jk
= H
n
i jk
Y
F
(t
n
) Y
F
(t
n+1
)
t
(4.83)
where t = t
n+1
t
n
. The species mass fractions are adjusted at this point in the calculation (before the
convection and diffusion update)
Y
n
,i jk
=Y
(t
n
)
F
W
F
_
Y
F
(t
n
) Y
F
(t
n+1
)
_
(4.84)
If multiple chemical reactions have been specied, equations 4.81 and 4.82 are evaluated for each reaction
during each of the 100 time steps. The reactions are evaluated in the order that they are entered in the input
le.
61
4.3 Thermal Radiation
This section describes how q
r
(the radiative loss term) is computed at all gas-phase cells. Plus, it describes
how the radiative and convective heat uxes, q
r
and q
c
, are computed at solid boundaries.
4.3.1 Discretization of the Radiative Heat Transfer Equation
The radiative transport equation (3.69) is solved using techniques similar to those for convective transport
in nite volume methods for uid ow [59], thus the name given to it is the Finite Volume Method (FVM).
To obtain the discretized form of the RTE, the unit sphere is divided into a nite number of solid angles. In
each grid cell a discretized equation is derived by integrating Eq. (3.69) over the volume of cell i jk and the
control angle
l
, to obtain
V
i jk
s
I(x
, s
)dx
ds
V
i jk
(x
)
_
I
b
(x
) I(x
, s
dx
ds
(4.85)
The volume integral on the left hand side is replaced by a surface integral over the cell faces using the
divergence theorem. Note that the procedure outlined below is appropriate for each band of a wide band
model, thus the subscript n has been removed for clarity.
Assuming that the radiation intensity I(x, s) is constant on each of the cell faces, the surface integral can
be approximated by a sum over the cell faces. Assuming further that I(x, s) is constant within the volume
V
i jk
and over the angle
l
we obtain
6
m=1
A
m
I
l
m
l
(s
n
m
)ds
=
i jk
_
I
b,i jk
I
l
i jk
_
V
i jk
l
(4.86)
where
I
l
i jk
radiant intensity in direction l
I
l
m
radiant intensity at cell face m
I
b,i jk
radiant blackbody Intensity in cell
l
solid angle corresponding to direction l
V
i jk
volume of cell i jk
A
m
area of cell face m
n
m
unit normal vector of the cell face m
Note that while the intensity is assumed constant within the angle
l
, its direction covers the angle
l
exactly.
In Cartesian coordinates
2
, the normal vectors n
m
are the base vectors of the coordinate system and the
integrals over the solid angle do not depend on the physical coordinate, but the direction only. The intensities
on the cell boundaries, I
l
m
, are calculated using a rst-order upwind scheme. If the physical space is swept
in the direction s
l
, the intensity I
l
i jk
can be directly obtained from an algebraic equation. This makes the
numerical solution of the FVM very fast. Iterations are needed only to account for the reective boundaries.
However, this is seldom necessary in practice, because the time step set by the ow solver is small.
The grid used for the RTE solver is the same as for the uid solver. The coordinate system used to
discretize the solid angle is shown in Figure 4.1. The discretization of the solid angle is done by dividing
rst the polar angle, , into N
bands, where N
()
2
In the axisymmetric case equation (4.86) becomes a little bit more complicated, as the cell face normal vectors n
m
are not
always constant. However, the computational efciency can still be retained.
62
s
x
z
y
and N
() are chosen to
give the total number of angles N
is calculated as
N
=
N
i=1
N
(
i
) (4.87)
The distribution of the angles is based on empirical rules that try to produce equal solid angles
l
=4/N
.
The number of -bands is
N
= 1.17 N
1/2.26
(4.88)
rounded to the nearest even integer. The number of -angles on each band is
N
() = max
_
4, 0.5N
_
cos(
) cos(
+
)
_
(4.89)
rounded to the nearest integer that is divisible by 4.
and
+
are the lower and upper bounds of the -
band, respectively. The discretization is symmetric with respect to the planes x = 0, y = 0, and z = 0. This
symmetry has three important benets: First, it avoids the problems caused by the fact that the rst-order
upwind scheme, used to calculate intensities on the cell boundaries, is more diffusive in non-axial directions
than axial. Second, the treatment of the mirror boundaries becomes very simple, as will be shown later.
Third, it avoids so called overhang situations, where s i, s j or s k changes sign inside the control angle.
These overhangs would make the resulting system of linear equations more complicated.
In the axially symmetric case these overhangs can not be avoided, and a special treatment, developed
by Murthy and Mathur [60], is applied. In these cases N
(
i
) is kept constant, and the total number of angles
is N
= N
. In addition, the angle of the vertical slice of the cylinder is chosen to be same as .
The cell face intensities, I
l
m
appearing on the left hand side of (4.86) are calculated using a rst order
upwind scheme. Consider, for example, a control angle having a direction vector s. If the radiation is
traveling in the positive x-direction, i.e. s i 0, the intensity on the upwind side, I
l
xu
is assumed to be the
intensity in the neighboring cell, I
l
i1 jk
, and the intensity on the downwind side is the intensity in the cell
itself I
l
i jk
.
On a rectilinear grid, the normal vectors n
m
are the base vectors of the coordinate system and the
integrals over the solid angle can be calculated analytically. Equation (4.86) can be simplied
a
l
i jk
I
l
i jk
= a
l
x
I
l
xu
+a
l
y
I
l
yu
+a
l
z
I
l
zu
+b
l
i jk
(4.90)
63
where
a
l
i jk
= A
x
|D
l
x
| +A
y
|D
l
y
| +A
z
|D
l
z
| +
i jk
V
i jk
l
(4.91)
a
l
x
= A
x
|D
l
x
| (4.92)
a
l
y
= A
y
|D
l
y
| (4.93)
a
l
z
= A
z
|D
l
z
| (4.94)
b
l
i jk
=
i jk
I
b,i jk
V
i jk
l
(4.95)
l
=
l
d =
sin d d (4.96)
D
l
x
=
l
(s
l
i)d (4.97)
=
(s
l
i)sin dd
=
cossinsin dd
=
1
2
_
sin
+
sin
__
_
cos
+
sin
+
cos
sin
_
D
l
y
=
l
(s
l
j)d (4.98)
=
sinsinsin dd
=
1
2
_
cos
cos
+
__
_
cos
+
sin
+
cos
sin
_
D
l
z
=
l
(s
l
k)d (4.99)
=
cossin dd
=
1
2
_
_
sin
+
_
2
_
sin
_
2
_
Here i, j and k are the base vectors of the Cartesian coordinate system.
+
,
,
+
and
+
are the upper and
lower boundaries of the control angle in the polar and azimuthal directions, respectively, and =
+
and =
+
. The solution method of (4.90) is based on an explicit marching sequence [61]. The
marching direction depends on the propagation direction of the radiation intensity. As the marching is done
in the downwind direction, the upwind intensities in all three spatial directions are known, and the
intensity I
l
i jk
can be solved directly. Iterations may be needed only with the reective walls and optically
thick situations. Currently, no iterations are made.
4.3.2 Radiation Heat Transfer to Solid Objects
The boundary condition on a solid wall is given as
I
l
w
=
T
4
w
+
1
D
l
w
<0
I
l
w
|D
l
w
| (4.100)
64
where D
l
w
=
l
(s n
w
)d. The constraint D
l
w
< 0 means that only the incoming directions are taken into
account when calculating the reection. The net radiative heat ux on the wall is
q
r
=
N
l=1
I
l
w
l
(s
n
w
)ds
=
N
l=1
I
l
w
D
l
n
(4.101)
where the coefcients D
l
n
are equal to D
l
x
, D
l
y
or D
l
z
, and can be calculated for each wall element at the
start of the calculation.
The open boundaries are treated as black walls, where the incoming intensity is the black body intensity
of the ambient temperature. On mirror boundaries the intensities leaving the wall are calculated from the
incoming intensities using a predened connection matrix:
I
l
w,i jk
= I
l
(4.102)
Computationally intensive integration over all the incoming directions is avoided by keeping the solid angle
discretization symmetric on the x, y and z planes. The connection matrix associates one incoming direction
l
l=1
I
l
i jk
l
(4.103)
65
4.4 Solid Phase Model
4.4.1 Solid Surface and Interior Temperatures
A one dimensional heat transfer calculation is performed at each solid boundary cell for which the user has
prescribed thermal properties. The solid can consist of multiple layers of materials. Each layer is partitioned
into non-uniform cells, clustered near the front and back faces. The smallest cells are chosen based on the
criteria
x < S
s
k
s
s
c
s
(4.104)
where S
s
is a cell size factor dened by the user. By default, F
s
is 1.0. Interior cells increase in size by a
user-dened stretch factor when moving inwards from the surfaces. By default, the stretch factor is 2.0. The
cell boundaries are located at points x
i
. The temperature at the center of the ith cell is denoted T
s,i
. The
(temperature-dependent) thermal conductivity of the solid at the center of the ith cell is denoted k
s,i
. The
temperatures are updated in time using an implicit Crank-Nicolson scheme
T
n+1
s,i
T
n
s,i
t
=
1
2(
s
c
s
)
i
x
i
_
k
s,i+
1
2
T
n
s,i+1
T
n
s,i
x
s,i+
1
2
k
s,i
1
2
T
n
s,i
T
n
s,i1
x
s,i
1
2
+
k
s,i+
1
2
T
n+1
s,i+1
T
n+1
s,i
x
i+
1
2
k
s,i
1
2
T
n+1
s,i
T
n+1
s,i1
x
i
1
2
_
(4.105)
for 1 i N. The width of each cell is x
i
. The distance from the center of cell i to the center of cell i +1
is x
i+
1
2
.
The boundary condition is discretized
k
s,1
T
n+1
s,1
T
n+1
s,0
x1
2
= q
c
(n+1)
+ q
r
(n+1)
(4.106)
The convective and radiative uxes at the next time step are approximated
q
c
(n+1)
q
c
n
h
_
T
n+1
s,
1
2
T
n
s,
1
2
_
(4.107)
q
r
(n+1)
q
r
n
4 T
n
3
s,
1
2
_
T
n+1
s,
1
2
T
n
s,
1
2
_
(4.108)
The wall temperature is dened T
w
T
s,
1
2
= (T
s,0
+T
s,1
)/2.
4.4.2 Convective Heat Transfer to Solid Objects
In a DNS calculation where the boundary layer is resolved, the convective ux to the wall is given by
q
c
=k
T
gas
T
w
n/2
(4.109)
where T
gas
is the gas temperature in the center of the rst gas-phase grid cell, T
w
is the surface or wall
temperature, n is the size of a grid cell in the normal direction to the wall, and k is the temperature-
dependent thermal conductivity of the gas mixture.
In an LES calculation, where the boundary layer is not resolved, an empirical heat transfer coefcient,
h, is used:
q
c
= h(T
gas
T
w
) W/m
2
(4.110)
66
The heat transfer coefcient, h, is the maximum of its natural and forced empirical values:
h = max
_
C|T
gas
T
w
|
1
3
,
k
air
L
0.037Re
4/5
L
Pr
1/3
_
W/m
2
/K (4.111)
C is an empirical coefcient (1.31 for vertical surface; 1.52 for horizontal), k
air
is the thermal conductivity of
air (not the turbulent value used in the LES solver), L is a length scale (assumed to be 1 m), the Reynolds
number is based on the density and velocity of the gases in the middle of the rst grid cell and the length
scale L, and the Prandtl number is assumed to be 0.7.
4.4.3 Coupling the Gas and Solid Phase
Gas phase temperatures are dened at cell centers; solid surfaces lie at the interface of the bordering gas
phase cell and a ghost cell inside the solid. As far as the gas phase calculation is concerned, the normal
temperature gradient at the surface is expressed in terms of the temperature difference between the gas
cell and the ghost cell. The solid surface temperature is not used directly in the gas phase calculation.
Rather, the ghost cell temperature is used to couple the gas and solid phases. The ghost cell temperature has
no physical meaning on its own. Only the difference between ghost and gas cell temperatures matters, for
this denes the heat transfer to the wall.
In a DNS calculation, the wall temperature is assumed to be an average of the ghost cell temperature
and the temperature of the rst cell in the gas, thus the ghost cell temperature is dened
T
ghost
= 2T
w
T
gas
(4.112)
For an LES calculation, the numerical expression for the heat lost to the boundary is equated with the
empirical convective heat transfer
k
LES
T
gas
T
ghost
n
c
p
u
n
T
gas
+T
ghost
2
= h (T
gas
T
s
) c
p
u
n
T
s
(4.113)
where n is the distance between the center of the ghost cell and the center of the gas cell. This equation is
solved for T
ghost
, so that when the conservation equations are updated, the amount of heat lost to the wall is
equivalent to the empirical expression on the right hand side. Note that T
ghost
is purely a numerical construct.
It does not represent the temperature within the wall, but rather establishes a temperature gradient at the wall
consistent with the empirical correlation.
At solid walls there is no transfer of mass, thus the boundary condition for the lth species at a wall is
simply
Y
l,ghost
=Y
l,gas
(4.114)
where the subscripts ghost and gas are the same as above since the mass fraction, like temperature, is
dened at cell centers. At forced ow boundaries either the mass fraction Y
l,w
or the mass ux m
l
of species
l may be prescribed. Then the ghost cell mass fraction can be derived because, as with temperature, the
normal gradient of mass fraction is needed in the gas phase calculation. For cases where the mass fraction
is prescribed
Y
l,ghost
= 2Y
l,w
Y
l,gas
(4.115)
For cases where the mass ux is prescribed, the following equation must be solved iteratively
m
l
= u
n
ghost
Y
l,ghost
+
gas
Y
l,gas
2
D
Y
l,gas
Y
l,ghost
n
t u
2
n
2
gas
Y
l,gas
ghost
Y
l,ghost
n
(4.116)
where m
l
is the mass ux of species l per unit area, u
n
is the normal component of velocity at the wall
pointing into the ow domain, and n is the distance between the center of the ghost cell and the center of
67
the gas cell. Notice that the last term on the right hand side is subtracted at the predictor step and added at
the corrector step, consistent with the biased upwinding introduced earlier.
Once the temperature and species mass fractions have been dened in the ghost cell, the density in the
ghost cell is computed from the equation of state
ghost
=
p
0
R T
ghost
l
(Y
l,ghost
/W
l
)
(4.117)
68
4.5 Liquid Sprays
Lagrangian particles are sometimes introduced into the ow eld as a means of visualization or simulation
of droplets, aerosols, etc.. The position x
p
of each particle is governed by the equations
dx
p
dt
= u (4.118)
The particle positions are updated according to the same predictor-corrector scheme that is applied to the
other ow quantities. Briey, the position x
p
of a given particle is updated according to the two step scheme
x
(n+1)
e
p
= x
n
p
+t u
n
(4.119)
x
n+1
p
=
1
2
_
x
n
p
+x
(n+1)
e
p
+t u
(n+1)
e
_
(4.120)
where the bar over the velocity vector indicates that the velocity of the uid is linearly interpolated at the
particles position.
4.5.1 Interaction of Droplets and Radiation
The computation of
f
for a similar but simpler situation has been derived in Ref. [62]. It can be shown that
here
f
becomes
f
=
1
l
0
l
0
d,
d,0
P
0
(
d
)
(12)(1
2
) (
d
)2
d
d
dd
(4.121)
where
d
is a cosine of the scattering angle
d
and P
0
(
d
) is a single droplet scattering phase function
P
0
(
d
) =
2
_
|S
1
(
d
)|
2
+|S
2
(
d
)|
2
_
2C
s
(r, )
(4.122)
S
1
(
d
) and S
2
(
d
) are the scattering amplitudes, given by Mie-theory. The integration limit
l
is a cosine of
the polar angle dening the boundary of the symmetric control angle
l
l
= cos(
l
) = 1
2
N
(4.123)
The limits of the innermost integral are
d,0
=
+
_
1
2
_
1
2
;
d,
=
_
1
2
_
1
2
(4.124)
When
f
is integrated over the droplet size distribution to get an averaged value, it is multiplied by C
s
(r, ).
It is therefore |S
1
|
2
+|S
2
|
2
, not P
0
(
d
), that is integrated. Physically, this means that intensities are added,
not probabilities [50].
4.5.2 Heating and Evaporation of Liquid Droplets
The exchange of mass and energy between liquid droplets and the surrounding gases (or solid surfaces) is
computed droplet by droplet. After the temperature of each droplet is computed, the appropriate amount of
vaporized liquid is added to the given mesh cell, and the cell gas temperature is reduced slightly based on
the energy lost to the droplet.
69
Equation (3.134) is solved semi-implicitly over the course of a gas phase time step as follows. Note that
a few terms have been left out to make the algorithm more clear.
T
n+1
l
T
n
l
t
=
1
m
l
c
l
_
Ah
_
T
g
T
n+1
l
+T
n
l
2
_
Ah
m
_
Y
n+1
l
+Y
n
l
2
Y
g
_
h
v
_
(4.125)
The equilibrium vapor mass fraction, Y
n
l
, is a function of T
n
l
via Eq. (3.135), and its value at the next time
step is approximated via
Y
n+1
l
Y
n
l
+
_
dY
l
dT
l
_
n _
T
n+1
l
T
n
l
_
(4.126)
where the derivative of Y
l
with respect to temperature is obtained via the chain rule:
dY
l
dT
l
=
dY
l
dX
l
dX
l
dT
l
=
W
a
/W
l
(X
l
(1W
a
/W
l
) +W
a
/W
l
)
2
h
v
W
l
R T
2
l
exp
_
h
v
W
l
R
_
1
T
b
1
T
l
__
(4.127)
The amount of evaporated liquid is given by
m
l
= t Ah
m
_
Y
n
l
+
1
2
_
dY
l
dT
l
_
n _
T
n+1
l
T
n
l
_
_
(4.128)
The amount of heat extracted from the gas is
q = t Ah
_
T
g
T
n
l
+T
n+1
l
2
_
(4.129)
70
Chapter 5
Conclusion
The equations and numerical algorithm described in this document form the core of an evolving re model.
As research into specic re-related phenomena continues, the relevant parts of the model can be improved.
Because the model was originally designed to predict the transport of heat and exhaust products from res,
it can be used reliably when the re is prescribed and the numerical grid is sufciently resolved to capture
enough of the ow structure for the application at hand. It is the job of the user to determine what level of
accuracy is needed.
Any user of the numerical model must be aware of the assumptions and approximations being employed.
There are two issues for any potential user to consider before embarking on calculations. First, for both real
and simulated res, the growth of the re is very sensitive to the thermal properties (conductivity, specic
heat, density, burning rate, etc.) of the surrounding materials. Second, even if all the material properties are
known, the physical phenomena of interest may not be simulated due to limitations in the model algorithms
or numerical grid. Except for those fewmaterials that have been studied to date at NIST, the user must supply
the thermal properties of the materials, and then validate the performance of the model with experiments to
ensure that the model has the necessary physics included. Only then can the model be expected to predict
the outcome of re scenarios that are similar to those that have actually been tested.
71
72
Bibliography
[1] S. Olenick and D. Carpenter. An Updated International Survey of Computer Models for Fire and
Smoke. Journal of Fire Protection Engineering, 13:87110, May 2003.
[2] R. Friedman. Survey of Computer Models for Fire and Smoke. Technical report, Factory Mutual
Research Corporation, Norwood, Massachusetts, March 1990.
[3] American Society for Testing and Materials, West Conshohocken, Pennsylvania. ASTM E 1355-04,
Standard Guide for Evaluating the Predictive Capabilities of Deterministic Fire Models, 2004.
[4] K.B. McGrattan, B.W. Klein, S. Hostikka, and J.E. Floyd. Fire Dynamics Simulator (Version 5),
Users Guide. NIST Special Publication 1019-5, National Institute of Standards and Technology,
Gaithersburg, Maryland, October 2007.
[5] H.C. Hottel. Stimulation of Fire Research in the United States After 1940 (A Historical Account).
Combustion Science and Technology, 39:110, 1984.
[6] W.W. Jones. A Review of Compartment Fire Models. NBSIR 83-2684, National Bureau of Standards
(now NIST), Gaithersburg, Maryland, 1983.
[7] J. Quintiere. A Perspective on Compartment Fire Growth. Combustion Science and Technology,
39:1154, 1984.
[8] G.P. Forney and W.F. Moss. Analyzing and Exploiting Numerical Characteristics of Zone Fire Models.
Fire Science and Technology, 14:4960, 1994.
[9] S.V. Patankar. Numerical Heat Transfer and Fluid Flow. Hemisphere Publishing, New York, 1980.
[10] R.G. Rehm and H.R. Baum. The Equations of Motion for Thermally Driven, Buoyant Flows. Journal
of Research of the NBS, 83:297308, 1978.
[11] E.S. Oran and J.P. Boris. Numerical Simulation of Reactive Flow. Elsevier Science Publishing Com-
pany, New York, 1987.
[12] V. Babrauskas. Ignition Handbook. Fire Science Publishers, Issaquah, Washington USA, 1st edition,
2003. Co-published by the Society of Fire Protection Engineers.
[13] G.P. Forney. Users Guide for Smokeview Version 5 - A Tool for Visualizing Fire Dynamics Simulation
Data. NIST Special Publication 1017-1, National Institute of Standards and Technology, Gaithersburg,
Maryland, August 2007.
[14] D.A. Anderson, J.C. Tannehill, and R.H. Pletcher. Computational Fluid Mechanics and Heat Transfer.
Hemisphere Publishing Corporation, Philadelphia, Pennsylvania, 1984.
73
[15] R. Peyret and T.D. Taylor. Computational Methods for Fluid Flow. Springer-Verlag, New York, 1983.
[16] J.H. Ferziger and M. Peri c. Computational Methods for Fluid Dynamics. Springer-Verlag, Berlin, 2nd
edition, 1999.
[17] R.W. Bilger. Turbulent Diffusion Flames. Annual Review of Fluid Mechanics, 21:101135, 1989.
[18] J.P. Holman. Heat Transfer. McGraw-Hill, New York, 5th edition, 1989.
[19] F.P. Incropera and D.P. De Witt. Fundamentals of Heat and Mass Transfer. John Wiley and Sons, New
York, 4th edition, 1996.
[20] R. Siegel and J. R. Howell. Thermal Radiation Heat Transfer. Taylor & Francis, New York, 4th
edition, 2002.
[21] P.J. DiNenno, editor. SFPE Handbook of Fire Protection Engineering. National Fire Protection Asso-
ciation, Quincy, Massachusetts, 3rd edition, 2002.
[22] D. Drysdale. An Introduction to Fire Dynamics. John Wiley and Sons, New York, 2nd edition, 2002.
[23] J.G. Quintiere. Principles of Fire Behavior. Delmar Publishers, Albany, New York, 1998.
[24] Verication and Validation of Selected Fire Models for Nuclear Power Plant Applications. NUREG
1824, United States Nuclear Regulatory Commission, Washington, DC, 2007.
[25] H.R. Baum and K.B. McGrattan. Simulation of Large Industrial Outdoor Fires. In Fire Safety Science
Proceedings of the Sixth International Symposium. International Association for Fire Safety Science,
2000.
[26] S.B. Pope. Ten questions concerning the large-eddy simulation of turbulent ows. New Journal of
Physics, 6:124, 2004.
[27] J. Smagorinsky. General Circulation Experiments with the Primitive Equations. I. The Basic Experi-
ment. Monthly Weather Review, 91(3):99164, March 1963.
[28] J.W. Deardorff. Numerical Investigation of Neutral and Unstable Planetary Boundary Layers. Journal
of Atmospheric Sciences, 29:91115, 1972.
[29] M. Germano, U. Piomelli, P. Moin, and W.H. Cabot. A Dynamic Subgrid-Scale Eddy Viscosity Model.
Physics of Fluids A, 3(7):17601765, 1991.
[30] D.K. Lilly. A Proposed Modication of the Germano Subgrid-Scale Closure Method. Physics of Fluids
A, 4(3):633635, 1992.
[31] H.R. Baum, K.B. McGrattan, and R.G. Rehm. Three Dimensional Simulations of Fire Plume Dynam-
ics. Journal of the Heat Transfer Society of Japan, 35(139):4552, 1997.
[32] B.E. Poling, J.M. Prausnitz, and J.P. OConnell. The Properties of Gases and Liquids. McGraw-Hill,
New York, 5th edition, 2000.
[33] I.K. Puri and K. Seshadri. Extinction of Diffusion Flames Burning Diluted Methane and Diluted
Propane in Diluted Air. Combustion and Flame, 65:137150, 1986.
[34] C.K. Westbrook and F.L. Dryer. Simplied Reaction Mechanisms for the Oxidation of Hydrocarbon
Fuels in Flames. Combustion Science and Technology, 27:3143, 1981.
74
[35] W. Grosshandler. RadCal: ANarrowBand Model for Radiation Calculations in a Combustion Environ-
ment. NIST Technical Note TN 1402, National Institute of Standards and Technology, Gaithersburg,
Maryland, 1993.
[36] K. Prasad, C. Li, K. Kailasanath, C. Ndubizu, R. Ananth, and P.A. Tatem. Numerical modelling of
methanol liquid pool res. Combustion Theory and Modelling, 3:743768, 1999.
[37] G. Heskestad and R.G. Bill. Quantication of Thermal Responsiveness of Automatic Sprinklers In-
cluding Conduction Effects. Fire Safety Journal, 14:113125, 1988.
[38] P. Rufno and M. di Marzo. Temperature and Volumetric Fraction Measurements in a Hot Gas Laden
with Water Droplets. Journal of Heat Transfer, 125(2):356364, April 2003.
[39] P. Rufno and M. di Marzo. The Effect of Evaporative Cooling on the Activation Time of Fire Sprin-
klers. In Fire Safety Science Proceedings of the Seventh International Symposium, pages 481492.
International Association for Fire Safety Science, 2002.
[40] F. Gavelli, P. Rufno, G. Anderson, and M. di Marzo. Effect of Minute Water Droplets on a Simulated
Sprinkler Link Thermal Response. NIST GCR99-776, National Institute of Standards and Technology,
Gaithersburg, Maryland, July 1999.
[41] T. Cleary, A. Chernovsky, W. Grosshandler, and M. Anderson. Particulate Entry Lag in Spot-Type
Smoke Detectors. In Fire Safety Science Proceedings of the Sixth International Symposium, pages
779790. International Association for Fire Safety Science, 1999.
[42] G.W. Mulholland and C. Croarkin. Specic Extinction Coefcient of Flame Generated Smoke. Fire
and Materials, 24:227230, 2000.
[43] R.J. Roby, S.J. Olenick, W. Zhang, D.J. Carpenter, M.S. Klassen, and J.L. Torero. A Smoke Detector
Algorithm for Large Eddy Simulation Modeling. NIST GCR 07-911, National Institute of Standards
and Technology, Gaithersburg, Maryland, 2007.
[44] T.S. Chan. Measurements of Water Density and Droplet Size Distributions of Selected ESFR Sprin-
klers. Journal of Fire Protection Engineering, 6(2):7987, 1994.
[45] H.Z. Yu. Investigation of Spray Patterns of Selected Sprinklers with the FMRC Drop Size Measuring
System. In Fire Safety Science Proceedings of the First International Symposium, pages 11651176.
International Association For Fire Safety Science, 1986.
[46] A. Hamins and K.B. McGrattan. Reduced-Scale Experiments to Characterize the Suppression of Rack
Storage Commodity Fires. NISTIR 6439, National Institute of Standards and Technology, Gaithers-
burg, Maryland, 1999.
[47] A. Hamins and K.B. McGrattan. Reduced-Scale Experiments on the Water Suppression of a Rack-
Storage Commodity Fire for Calibration of a CFD Fire Model. In Fire Safety Science Proceedings
of the Seventh International Symposium, pages 457468. International Association for Fire Safety
Science, 2002.
[48] T. Ravigururajan and M. Beltran. A Model for Attenuation of Fire Radiation Through Water Droplets.
Fire Safety Journal, 15:171181, 1989.
[49] A. Tuntomo, C. Tien, and S. Park. Optical Constants of Liquid Hydrocarbon Fuels. Combustion
Science and Technology, 84:133140, 1992.
75
[50] W.J. Wiscombe. Improved Mie Scattering Algorithms. Applied Optics, 19(9):15051509, 1980.
[51] G.M. Hale and M.R. Querry. Optical constants of water in the 200 nm to 200 mum wavelength region.
Applied Optics, 12(3):555563, 1973.
[52] N. Cheremisinoff. Encyclopedia of Fluid Mechanics, Volume 3: Gas-Liquid Flows. Gulf Publishing
Company, Houston, Texas, 1986.
[53] H.Z. Yu, J.L. Lee, and H.C. Kung. Suppression of Rack-Storage Fires by Water. In Fire Safety Science
Proceedings of the Fourth International Symposium, pages 901912. International Association For
Fire Safety Science, 1994.
[54] K.B. McGrattan, R.G. Rehm, and H.R. Baum. Fire-Driven Flows in Enclosures. Journal of Computa-
tional Physics, 110(2):285291, 1994.
[55] H.R. Baum, O.A. Ezekoye, K.B. McGrattan, and R.G. Rehm. Mathematical Modeling and Computer
Simulation of Fire Phenomenon. Theoretical and Computational Fluid Dynamics, 6:125139, 1994.
[56] G. Continillo, F.M Denaro, and F.S. Marra. Accuracy and Stability Analysis of Time-Integrated
Schemes for Advection-Diffusion-Reaction Equations. In Seventh International Conference on Nu-
merical Combustion, page 99, 1998.
[57] W. Zhang, N. Ryder, R.J. Roby, and D. Carpenter. Modeling of the Combustion in Compartment Fires
Using Large Eddy Simulation Approach. In Proceedings of the 2001 Fall Technical Meeting, Eastern
States Section. Combustion Institute, Pittsburgh, Pennsylvania, December 2001.
[58] R.A. Sweet. Direct Methods for the Solution of Poissons Equation on a Staggered Grid. Journal of
Computational Physics, 12:422428, 1973.
[59] G.D. Raithby and E.H. Chui. A Finite-Volume Method for Predicting Radiant Heat Transfer in Enclo-
sures with Participating Media. Journal of Heat Transfer, 112(2):415423, 1990.
[60] J.Y. Murthy and S.R. Mathur. Radiative Heat Transfer in Axisymmetric Geometries Using an Un-
structured Finite-Volume Method. Numerical Heat Transfer, Part B, 33:397416, 1998.
[61] S.H. Kim and K.Y. Huh. Assessment of The Finite-Volume Method and The Discrete Ordinate Method
for Radiative Heat Transfer In a Three-Dimensional Rectangular Enclosure. Numerical Heat Transfer,
Part B, 35:85112, 1999.
[62] W. Yang, T. Parker, H.D. Ladouceur, and R.J. Kee. The Interaction of Thermal Radiation and Water
Mist in Fire Suppression. Fire Safety Journal, 39:4166, 2004.
76
Appendix A
Nomenclature
A
s
droplet surface area
A
enthalpy of species
h
0
f
fuel mass ux
m
w
water mass ux
m
w
water mass per unit area
Nu Nusselt number
Pr Prandtl number
p pressure
p
0
atmospheric pressure prole
p
m
background pressure of mth pressure zone
p pressure perturbation
q
heat ux vector
q
r
radiative ux to a solid surface
q
c
convective ux to a solid surface
characteristic re size
R universal gas constant
Re Reynolds number
r
d
droplet radius
r
O
2
mass fraction of oxygen in the ambient
Y
I
F
mass fraction of fuel in the fuel stream
y
s
soot yield
Z mixture fraction
Z
f
stoichiometric value of the mixture fraction
ratio of specic heats; Rosin-Rammler exponent
H heat of combustion
H
O
2
energy released per unit mass oxygen consumed
wall thickness
dissipation rate
absorption coefcient
dynamic viscosity
s
yield of solid residue in solid phase reaction
f
yield of gaseous fuel in solid phase reaction
w
yield of gaseous water in solid phase reaction
density
i j
viscous stress tensor
r
radiative loss fraction
Stefan-Boltzmann constant; constant in droplet size distribution; surface tension
d
droplet scattering coefcient
s
scattering coefcient
= (
x
,
y
,
z
) vorticity vector
78
Appendix B
Derivation of the Velocity Divergence
Randall McDermott, NIST Postdoctoral Fellow
In this appendix we derive the divergence of the velocity eld as presented in Eq. (3.17). Note that the
constitutive relationships presented here for the mass diffusion and thermal heat uxes are valid for direct
numerical simulations (i.e., well-resolved calculations). The minor modications required of the transport
coefcients for large-eddy simulation are presented in Section 3.1.6. We start the derivation by rearranging
the continuity equation. Next, we differentiate the equation of state to reveal the relationship between
transport equations for mass and energy. We then show how the transport equations may be combined to
yield the velocity divergence constraint. In the last section we present the nal result in FDS notation.
Continuity Equation
Let denote the uid mass density; let u = [u, v, w]
T
denote the uid mass-average velocity; and let m
b
denote a bulk source of mass per unit volume (which may come from the evaporation of water droplets, for
example). The continuity equation may be rearranged to yield the following divergence constraint on the
velocity
u =
1
_
m
b
D
Dt
_
(B.1)
where D( )/Dt ( )/t +u ( ) is the material derivative.
Equation of State
We consider the transport of n
s
species mass fractions Y
for ={1, . . . , n
s
}, n
s
1 of which are independent.
The molecular weight of a given species is denoted W
_
1
(B.2)
where as a shorthand notation, which is used throughout this document, we write
for
n
s
=1
. Let p
i
(x, t)
denote the hydrostatic pressure in the ith zone of the domain, which in general we take to be a function
of space and time. In practice, however, p
i
= p
i
(t) for closed (i.e., sealed or pressurized) domains and
p
i
= p
i
(z), where z represents the coordinate aligned with the gravity vector, for large, open domains (e.g.,
forest res large enough to interact with the stratied atmosphere). The divergence constraint derived below
is based on the ideal gas equation of state (EOS), which, for low-Mach ows, we write as
p
i
=
R T
W
(B.3)
79
where R = 8.3145 kJ/(kmol K) is the gas law constant.
Differentiating the EOS (B.3) we obtain
Dp
i
Dt
= R T
D
Dt
_
1
W
_
+
R
W
DT
Dt
+
R T
W
D
Dt
(B.4)
which rearranges to
D
Dt
=
W
R T
Dp
i
Dt
W
D
Dt
_
1
W
_
T
DT
Dt
(B.5)
Species Transport Equation
The species transport equation plays a role in both the second and third terms on the RHS of (B.5). Including
the bulk mass source, the evolution of species mass fractions is governed by
(Y
)
t
+ (Y
u) = J
+ m
+ m
b,
(B.6)
where J
is the diffusive mass ux vector for species (relative to the mass-average velocity), m
is the
chemical mass production rate of per unit volume [kg- produced /(sec m
3
)], and m
b,
is the bulk mass
source of per unit volume [kg- introduced /(sec m
3
)]. Note that
b,
= m
b
(B.7)
and
= 0 (B.8)
Additionally, by construction, the ith component of the species diffusive uxes sum to zero,
J
,i
= 0 (B.9)
Thus, as must be the case, summing (B.6) over yields the continuity equation.
It is convenient to work in terms of the material derivative of the mass fraction. Care must be exercised
in obtaining this expression because the continuity equation is of a non-standard form. Expanding (B.6) we
obtain
t
+Y
t
+u Y
+Y
(u) = J
+ m
+ m
b,
,
DY
Dt
+Y
t
+ (u)
_
. .
m
b
= (B.10)
Thus, the material derivative of the mass fraction can be written as
DY
Dt
=
1
_
m
+ m
b,
Y
b
J
_
=
1
_
m
+ m
b
[Y
b,
Y
] J
_
(B.11)
where in the second step we use the identity m
b,
=Y
b
with Y
b,
being the mass fraction of in the bulk
prior to its introduction into the uid mixture.
80
Utilizing (B.2) and (B.11) we obtain
D
Dt
_
1
W
_
=
D
Dt
_
_
,
=
1
W
DY
Dt
,
=
1
1
W
_
m
+ m
b
[Y
b,
Y
] J
_
, (B.12)
which is needed in the second term on the RHS of (B.5).
Enthalpy Transport Equation
The specic sensible enthalpy of species relative to reference temperature T
0
is
h
(T) =
T
T
0
c
p,
(T
)dT
, (B.13)
where the specic heat of is
c
p,
T
. (B.14)
The specic sensible enthalpy of the mixture is then given by
h
s
(Y, T) =
(T). (B.15)
Neglecting viscous heating and the effect of the uctuating pressure on dilation work (both assumptions
are valid for low-Mach ows), the transport equation for the sensible enthalpy is
Dh
s
Dt
=
h
0
+
Dp
i
Dt
q
+ m
b
(h
s,b
h
s
) (B.16)
where h
0
=kT +
h
s,
J
+ q
r
(B.17)
Here k is the thermal conductivity of the mixture and q
r
is the radiant heat ux.
Relating Enthalpy, Temperature, and Species
Using the chain rule of calculus, we may expand the derivative of the sensible enthalpy h
s
(Y, T) to obtain
Dh
s
Dt
=
_
h
s
T
_
DT
Dt
+
_
h
s
Y
_
DY
Dt
(B.18)
Note that since h
s
=
we have
h
s
Y
=
Y
(Y
) =
= h
(B.19)
81
where
_
h
T
_
=
c
p,
c
p
(B.20)
dening the specic heat of the mixture. Thus, by rearranging (B.18) and utilizing (B.19) and (B.20) we
obtain
DT
Dt
=
1
c
p
_
Dh
s
Dt
DY
Dt
_
(B.21)
Utilizing (B.11) and (B.16) in (B.21) yields
DT
Dt
=
1
c
p
_
h
0
+
Dp
i
Dt
q
b
+ m
b
(h
s,b
h
s
)
h
s,
_
m
+ m
b
[Y
b,
Y
] J
_
_
(B.22)
Note that
h
s,
m
b
[Y
b,
Y
] = m
b
(h
s,b
h
s
) and hence these terms cancel, leaving
DT
Dt
=
1
c
p
_
h
0
+
Dp
i
Dt
q
_
m
_
_
(B.23)
Assembling Terms
We now have all the pieces we need to construct the divergence constraint which we introduced in Eq. (B.1).
Using (B.5) in (B.1) we obtain
u =
1
_
m
b
_
W
R T
Dp
i
Dt
W
D
Dt
_
1
W
_
T
DT
Dt
__
=
1
b
1
p
i
Dp
i
Dt
+W
D
Dt
_
1
W
_
+
1
T
DT
Dt
(B.24)
where in the second step the EOS (B.3) is used to simplify the second term on the RHS. Using (B.12) and
(B.23) in (B.24) yields
u =
1
b
1
p
i
Dp
i
Dt
+ W
_
1
1
W
_
m
+ m
b
[Y
b,
Y
] J
_
_
+
1
T
_
1
c
p
_
h
0
+
Dp
i
Dt
q
_
m
_
__
(B.25)
Note that W
(Y
/W
) = 1 and also W
(Y
b,
/W
) =W/W
b
, where W
b
is the molecular weight of
the bulk mixture prior to its introduction into the uid mixture. Equation (B.25) thus simplies to
u =
_
1
c
p
T
1
p
i
_
Dp
i
Dt
+
1
_
m
b
W
W
b
+W
1
W
_
m
_
_
+
1
c
p
T
_
h
0
_
m
_
_
(B.26)
82
FDS Notation
The following relationships are used to rearrange (B.26) into the formshown in the FDS Technical Reference
Guide.
We employ the binary form of Ficks law using mixture-averaged diffusivities D
as a constitutive
relation for the diffusive ux,
J
=D
(B.27)
Note that summation is not implied over repeated sufxes. The heat release rate per unit volume is dened
by
q
h
0
(B.28)
Taking the z direction to be aligned with the gravity vector we have p
i
/z =
i
g, where g = 9.8 m/s
2
and
i
is a specied background density for the ith zone. Thus, the material derivative of the background
pressure may be written as
Dp
i
Dt
=
p
i
t
w
i
g. (B.29)
Hence, utilizing (B.27), (B.28), and (B.29), and noting 1/(T) =R /(Wp
i
) from the EOS, for the ith zone
we may write the divergence (B.26) as
u =D+P
p
i
t
(B.30)
where
P =
1
p
i
_
p
i
c
p
T
1
_
=
1
p
i
_
R
Wc
p
1
_
(B.31)
and
D =
m
W
W
b
+
R
Wc
p
p
i
_
q
b
q
_
Pw
i
g
+
W
(D
[Y
/W
])
R
Wc
p
p
i
(D
) +
1
_
W
W
c
p
T
_
m
(B.32)
83
84
Appendix C
A Simple Model of Flame Extinction
Frederick W. Mowrer, Department of Fire Protection Engineering, University of Maryland
A diffusion ame immersed in a vitiated atmosphere will extinguish before consuming all the available
oxygen from the atmosphere. The classic example of this behavior is a candle burning within an inverted jar.
This same concept has been applied within FDS to determine the conditions under which the local ambient
oxygen concentration will no longer support a diffusion ame. In this appendix, the critical adiabatic ame
temperature concept is used to estimate the local ambient oxygen concentration at which extinction will
occur.
Consider a control volume characterized by a bulk temperature, T
m
, a mass, m, an average specic heat,
c
p
, and an oxygen mass fraction, Y
O
2
. Complete combustion of the oxygen within the control volume would
release a quantity of energy given by:
Q = mY
O
2
_
H
r
O
2
_
(C.1)
where H/r
O
2
has a relatively constant value of approximately 13100 kJ/kg for most fuels of interest for
re applications.
1
Under adiabatic conditions, the energy released by combustion of the available oxygen
within the control volume would raise the bulk temperature of the gases within the control volume by an
amount equal to:
Q = mc
p
(T
f
T
m
) (C.2)
The average specic heat of the gases within the control volume can be calculated based on the composition
of the combustion products as:
c
p
=
1
(T
f
T
m
)
T
f
T
m
c
p,
(T)dT (C.3)
To simplify the analysis, the combustion products are assumed to have an average specic heat of 1.1 kJ/kg/K
over the temperature range of interest, a value similar to that of air, the primary component of the products.
The relationship between the oxygen mass fraction within the control volume and the adiabatic temperature
rise of the control volume is evaluated by equating Eqs. (C.1) and (C.2):
Y
O
2
=
c
p
(T
f
T
m
)
H/r
O
2
(C.4)
If the critical adiabatic ame temperature is assumed to have a constant value of approximately 1700 K for
hydrocarbon diffusion ames, as suggested by Beyler,
2
then the relationship between the limiting oxygen
1
C. Huggett, Estimation of the Rate of Heat Release by Means of Oxygen Consumption, Fire and Materials, Vol. 12, pp. 61-
65, 1980.
2
C. Beyler, Flammability Limits of Premixed and Diffusion Flames, SFPE Handbook of Fire Protection Engineering (3rd
Ed.), National Fire Protection Association, Quincy, MA, 2003.
85
mass fraction and the bulk temperature of a control volume is given by:
Y
O
2
,lim
=
c
p
(T
f ,lim
T
m
)
H/r
O
2
1.1(1700T
m
)
13100
(C.5)
The relationship represented by Eq. (C.5) is shown in Fig. 3.2. For a control volume at a temperature of
300 K, i.e., near room temperature, the limiting oxygen mass fraction would evaluate to Y
O
2
,lim
= 0.118.
This value is consistent with the measurements of Morehart, Zukoski and Kubota,
3
although it is about 10
percent lower than their measured values and the limiting oxygen index values determined by Simmons and
Wolfhard.
4
This difference would tend to be conservative in terms of indicating ammability of gas mixtures
at oxygen concentrations lower than might occur in practice. Comparisons with additional experimental data
are needed to further validate the concept. Some improvement might result from a more exact analysis of the
average specic heat of the combustion products or from the use of a slightly higher critical adiabatic ame
temperature in the range of 1700 K to 1800 K, as suggested by the limiting oxygen index data of Simmons
and Wolfhard that is summarized in the SFPE Handbook Chapter of Beyler.
3
Morehart, J., Zukoski, E., and Kubota, T., Characteristics of Large Diffusion Flames Burning in a Vitiated Atmosphere, Third
International Symposium on Fire Safety Science, Elsevier Science Publishers, pp. 575-583, 1991.
4
R.F. Simmons and H.G. Wolfhard, Combustion and Flame, Vol. 1, p. 155, 1957.
86