ENERGY Science with

monkeyresultΜηχανική

22 Φεβ 2014 (πριν από 3 χρόνια και 3 μήνες)

88 εμφανίσεις

C O MB U S T I O N S C I E N C E
Introduction
About 80% of the total energy used each year is
consumed by combustion in machines such as gaso-
line- and diesel-powered vehicles, jet planes, and
electric power plants. Worldwide, fossil fuels add
more than twenty five billion tons of carbon dioxide
to the atmosphere every year, along with vast quan-
tities of other pollutants. Even modest gains in the
efficiency of combustion translate into significant
energy savings, reduced pollution, and decreased
reliance on foreign energy sources.
Scientists working on Office of Science–funded
“Terascale High-Fidelity Simulations of Turbulent
Combustion with Detailed Chemistry (TSTC)”
and other projects are exploring new ways to
burn fuels that will improve combustion effi-
ciency while controlling noxious emissions.
Results derived from computer simulations (such
as the scalar dissipation rate map indicating the
local mixing intensity, depicted in figure 1) enable
researchers to digitally evaluate new combustion
technology without resorting to costly and time-
consuming prototyping.
Large-scale simulations have proven useful in the
d
e
v
e
l
opment of engines for burning lean mixtures,
in which the ratio of fuel to air is lower than the
c
hemically optimal ratio. Reducing the fuel-to-air
r
a
ti
o r
e
sults in lower maximum temperatures and
d
ecr
e
a
s
e
s e
mis
sion of nitrogen oxides, hydrocar-
bons, and carbon monoxide. The homogeneous-
c
h
ar
g
e compression ignition (HCCI) process is an
e
x
am
p
l
e o
f a
d
vanced lean-burn technology, prom-
ising both high efficiency and low emissions by
c
o
m
p
ressing a lean, premixed fuel-air mixture
un
ti
l it i
gnit
e
s s
po
n
taneously in many separate
locations (figure 2). One challenge posed by HCCI
c
o
mb
u
stion is the prevention of engine knock, a
r
a
p
i
d r
is
e o
f combustion pressures that may occur
at high loads in engines and can often lead to
e
n
gin
e damage. One potential method of eliminat-
in
g kn
ock is t
o in
tr
odu
ce th
ermal or mixture vari-
ations in the engine cylinder to produce the desired
h
e
a
t release rate. To model and improve the per-
f
o
r
m
an
ce o
f su
ch an engine, researchers must be
able to understand the propagation of autoignition
through a thermally stratified mixture.
Combustion involves complex interactions
between chemistry and turbulent fluid flow (side-
bar “Physics of Combustion,” p45). State-of-the-art
research draws on three complementary parts: the-
ory, computer simulation, and experiment. Theo-
retical analysis provides the basic framework for
understanding what is happening at a microscopic
level. Computer simulations enable researchers to
explore how the microscopic relations that govern
the delicate balance between molecular diffusion
and reaction lead to often surprising large-scale
(macroscopic) behavior. In the case of combustion
for example, the underlying phenomena include tur-
bulent fluid flow and potentially thousands of chem-
ical reactions among hundreds of chemical species.
Understanding turbulent flow is critical to proper
design. Efficient combustion occurs when there is
sufficient turbulence to thoroughly mix the air and
fuel, but not so much that the combustion flame is
prematurely extinguished. Computer simulations
provide a way of studying the effect of turbulence
inside the combustion chamber on chemical reac-
ti
o
ns
, an
d vi
ce versa.
Despite the power of modern computers, they
c
annot yet completely simulate a combustion
c
h
ambe
r o
v
er the entire relevant span of length
sc
al
e
s
, f
r
o
m c
hemical reactions in the smallest,
submillimeter whorls of turbulent flow to the
m
an
y
-ce
ntimeter dimensions of the chamber itself.
T
h
e c
al
c
ul
a
ti
ons must span similarly wide ranges
in time. Recent advances in supercomputing
po
w
e
r t
ogether with high-order accurate scalable
al
g
o
r
ithms h
a
v
e e
nabled researchers to address
fundamental combustion questions with unprece-
d
e
n
t
ed realism using Direct Numerical Simulation
(D
N
S). D
N
S is a hi
gh
-fi
delity simulation approach
that numerically resolves all of the relevant fluid
an
d c
h
emical scales and addresses fundamental
c
o
mb
u
s
ti
o
n questions with unprecedented real-
ism. DNS is the gold standard in computational
flui
d d
yn
amics (CFD) and can provide complete
s
p
a
ti
al an
d t
e
mporal information for simple
42
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
ENERGY Science with
SciDAC combustion researchers are using digital combustors to probe the complex
chemical and physical processes that affect performance in practical applications.
These detailed simulations provide insight into novel designs, such as clean and
efficient homogeneous-charge compression ignition (HCCI) engines.
Even modest gains in the
efficiency of combustion
translate into significant
energy savings, reduced
pollution, and decreased
reliance on foreign energy
sources.
43
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
DIGITAL
Combustors
Figure 1.
A volume rendering of scalar dissipation rates indicative of the local mixing intensity in a temporal plane jet flame. High values in
red, medium in yellow, and lower values in blue. The scalar dissipation rate shows a complex, turbulent structure and exhibits a wide dynamic
range. The highest scalar dissipation rates exist in thin, highly intermittent layers.
K. L. MA, H. AKIBA, AND
H. YU
, UNIVERSITYOF
CALIFORNIA–DAVISAND
E. HAWKES, SANDIA
N
ATIONAL
L
ABORATORIES
Figure 2.
SciDAC researchers are actively exploring the science underlying novel engine concepts like homogeneous-charge compression ignition
(HCCI), shown here in contrast to conventional gasoline and diesel engine designs. In this process, a lean fuel-air mixture is heated by compression
and ignites spontaneously in many places. Researchers hope that such engines will provide the high efficiency of diesel engines without the
associated emissions.
C O MB U S T I O N S C I E N C E
canonical configurations, such as the model sys-
tems encountered in laboratory-scale experiments.
Scientists can use this information to determine
the causal relationships between the chemical and
physical events occurring during combustion.
While experimentation provides partial informa-
ti
o
n un
d
e
r realistic conditions, DNS uses the power
of the world’s largest supercomputers to provide
c
omplete information beginning to approach real-
is
ti
c c
o
n
ditions. DNS can test how accurately var-
i
o
u
s a
p
p
r
o
ximations capture the average behavior
of the combustion process for engineering mod-
e
ls
. T
h
ese engineering models can then be used in
m
o
r
e c
o
n
v
e
ntional CFD codes to calculate the
behavior of the complete combustion chamber
un
d
e
r v
arious conditions. Finally, researchers can
cr
o
s
s
-c
h
eck th
e r
esults of these calculations with
the behavior of experimental chambers, to ensure
th
a
t all n
ece
ssary features have been included in
th
e m
od
e
lin
g
. In a
ddition to validating these
approximate models, the detailed microscopic
d
e
scr
iption of the combustion process yielded by
D
N
S p
r
o
vi
d
es a stringent test of theoretical mod-
els in a way experiments cannot.
T
h
e S
ciDAC program has funded two projects
t
o s
t
u
d
y th
e in
teractions between the fluid flow and
chemical reactions that constitute combustion.
The thrust of this article concerns the TSTC proj-
ect, which is led by investigators Dr. Hong Im of
the University of Michigan, Dr. Jacqueline Chen of
Sandia National Laboratories (SNL), Dr. Christo-
pher Rutland of the University of Wisconsin–
M
a
d
is
o
n, and Dr. Arnaud Trouvé of the University
of Maryland. Involving multiple institutions and
r
esearchers, the TSTC team has adapted existing
c
od
e
, d
e
veloped at SNL, to incorporate improved
al
g
o
r
ithms f
o
r tr
e
atment of numerical boundary
conditions. TSTC researchers have also included
im
po
r
t
ant new physical models to simulate spray
c
o
mb
u
s
ti
o
n
, as well as soot and thermal radiation
interactions.
T
S
T
C w
orkers cooperate with scientists from
o
th
e
r S
c
iD
A
C p
rojects including the Computa-
tional Facility for Reacting Flow Science (CFRFS;
s
ee si
d
e
bar “Computational Facility for Reacting
F
l
o
w S
c
i
e
nce,” p48), as well as other SciDAC-
supported researchers like Dr. John Bell of
L
a
wr
ence Berkeley National Laboratory (LBNL;
s
ee “
Sim
ul
a
tin
g T
urbulent Flames,” p25). Other
DOE programs support exploration of engineer-
in
g m
od
els such as the Large Eddy Simulation
(LE
S; s
ee si
d
e
bar “
En
gineering Models,” p50).
44
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
Gaso
line Engine
(Spark Ignition)
Diesel Engine
(Compre
ssion Ignition)
HCCI Engine
(Homogeneous Charge
Compre
ssion Ignition)
Fuel Injector
Spark Plug
Ignition Point
Many
Ignition Points
Air
Fuel-Air Mixture
(Homogeneous)
Fuel-Air Mixture
(Homogeneous)
Fuel-Air Mixture
(Heterogeneous)
An important challenge in
combustion simulation is
the wide range of time
scales over which
important chemical
reactions occur.
I
LLUSTRATION: A. TOVEY
45
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
Advanced Computing
Only in recent years has DNS begun to approach
the goal of simulating full-scale, three-dimen-
sional combustion. One of the most powerful
examples of this effort is the S3D solver developed
at SNL. S3D solves the full compressible Navier-
Stokes equations that describe the conservation
of mass, momentum, and energy, and laws of gas
behavior, while simultaneously tracking the evo-
lution of reactive species on a rectangular mesh.
It uses the message passing interface (MPI) to effi-
ciently distribute the calculation among parallel
processors, and is built on a hierarchical, modu-
lar structure.
With support from the SciDAC Scientific
Application Program (SAP), combustion re-
searchers at SNL—together with computer sci-
entists at the National Energy Research Scientific
Computing (NERSC) Center and the National
C
e
n
t
e
r for Computational Sciences (NCCS) at
Oak Ridge National Laboratory (ORNL)—opti-
mi
zed and rewrote several important modules
in S3D
, im
p
r
o
ving performance by about 50% on
sc
al
ar arc
hit
ec
t
ur
e
s. The adaptations also permit
large simulations on vector machines like the
Cr
a
y X1E
. T
hese efforts have dramatically
im
p
r
o
v
ed th
e a
b
ility of the S3D software to run
efficiently on a variety of terascale computing
p
l
a
tf
orms (figure 3). Dr. Chen’s group is currently
w
o
r
kin
g c
l
o
s
ely with computer scientists at
NCCS/ORNL on scaling S3D to petascale
m
a
c
hin
es.
T
S
T
C r
e
s
e
archers have developed special code
modules to include additional important physi-
cal effects. For example, simulating the injection
of fuel into a combustion chamber as a spray
requires accounting for the evaporation of the
fuel droplets, as well as the entrainment of the gas
by the droplets. Dr. Rutland’s team has developed
tools for incorporating these effects into simu-
l
a
ti
o
ns (sidebar “Simulating Sprays,” p46).
In a
d
d
iti
o
n t
o th
e tr
aditional mechanisms of
conduction and convection, researchers are rec-
o
gni
zin
g th
e increasing importance of including
h
e
a
t tr
ans
f
e
r thr
ough radiation. Dr. Im, Dr.
Trouvé, and their collaborators are addressing
s
oo
t
, r
adiation, and the interaction between the
t
w
o (si
d
e
bar “
Soo
t and Radiation,” p47).
An important challenge in combustion simula-
ti
o
n is th
e wide range of time scales over which
im
po
r
t
an
t c
h
emical reactions occur, from nanosec-
onds to milliseconds. The large discrepancy in time
sc
al
e
s, known as stiffness, can drastically degrade
th
e e
ffi
c
i
e
n
cy of calculations, since a simulation that
What makes the simulation of combustion so
difficult? For one thing, the gas flow in any
modern engine is highly turbulent. This means
that the gas, instead of flowing along smooth,
predictable trajectories, forms into swirling
eddies, which in turn spawn smaller eddies
down to very small length scales. This swirling
action is important to mix the fuel and air for
efficient burning. The chaotic nature of this
fluid motion is one of the most poorly
understood problems of classical physics.
Within the turbulent flow, other complicated
events occur. In each small volume of the fluid
there are numerous chemical species, and
each can potentially undergo many possible
chemical reactions. Computer modeling must
track the changing concentrations of each of
these species as combustion proceeds.
Researchers frequently simplify the chemistry,
replacing the detailed chemistry with more
general reaction systems.
Individually, both chemistry and turbulence
are challenging problems. Adding to the
complexity, in a burning mixture the chemistry
and the turbulence are constantly interacting.
The turbulent flow stirs the mixture, creating new
conditions. The molecules also diffuse rapidly
within the gas, with lighter species like hydrogen
moving faster than heavier ones. At the same
time, the energy released by combustion heats
the gas, changing the flow further.
To accurately model these complex
turbulence-chemistry interactions requires high
resolution in space and time, as well as wide
dynamic range. These simulations must also be
advanced over long periods of time to achieve
the statistical stationarity required to test
models. A quite tractable simulation can very
quickly grow into a problem requiring terascale
or petascale computers. Developing petascale
computing capabilities for open science is a
major goal of the SciDAC program.
Physi cs of Combust i on
Figur
e 3.
S3D software has been modified to run efficiently on a variety of
processing platforms, and to maintain its performance when thousands of processors
are run in parallel.
An important method for
focusing computational
resources where they are
needed is adaptive mesh
refinement (AMR).
ILLUSTRATION: A. TOVEY
1,000
10
1
10
100
1,000
10,000
Number of Processors
100
Cost per Grid Point per Time-Step (µs)
SP3 (NERSC)
SP5 (NERSC)
Itanium (PNNL)
X1E (ORNL)
XT3(ORNL)
Opteron (SNL)
C O MB U S T I O N S C I E N C E
can accurately capture the chemical activity will
require a great number of steps to describe the
slowly evolving flow. Researchers at Princeton Uni-
versity are developing non-stiff, accurate, efficient
models for chemical reactions using an automated
mechanism generation software library. Another
approach is to modify the algorithms for calculat-
ing the time evolution by integrating the chemical
s
o
urce
s im
p
licitly. The CFRFS has also independ-
ently developed methods for reducing stiffness
(si
debar “Computational Facility for Reacting Flow
S
c
i
e
n
ce,” p48).
An im
po
r
t
an
t m
e
th
od for focusing computa-
tional resources where they are needed is adaptive
m
e
sh r
e
finement (AMR). In this technique, the
c
o
m
p
u
t
a
ti
onal mesh is made finer in the small
regions where the gas properties are varying rap-
i
dl
y with po
siti
on. CFRFS researchers have imple-
m
e
n
t
ed AMR o
n th
e
ir p
latform as well.
AMR is most effective when combustion occurs
in fl
am
e
l
ets, thin localized sheets, as in the simu-
l
a
ti
o
ns o
f Dr
. B
ell and his coworkers (“Simulat-
ing Turbulent Flames,” p25). In contrast, Dr.
C
h
e
n’s group uses simulations to probe the “thin
r
e
a
c
ti
o
n z
ones” regime, where the turbulence and
flame scales are comparable and occupy a signif-
i
c
an
t fraction of the computational domain, mak-
in
g AMR l
e
s
s e
ff
ec
tive.
The large, complex datasets that emerge from
simulations of turbulent combustion make data
management and extraction of useful informa-
tion a challenge. Combustion researchers have
worked closely with computer scientists and
applied mathematicians to explore new ways to
manage, mine, and visualize the terabytes of data
(“From Data to Discovery,” p28). In collaboration
with r
e
s
e
arc
hers from the University of Califor-
nia–Davis led by Dr. Kwan-Liu Ma, TSTC
r
esearchers have been exploring new ways to
sim
ult
an
eo
u
sly visualize multiple variables from
l
ar
g
e d
a
t
a
s
ets. The continuing efforts toward
improved visualization draw on technical illus-
tr
a
ti
o
n techniques and methods such as anima-
ti
o
n
, sli
c
in
g
, an
d user control of data selection to
provide a real-time, interactive platform with
w
hi
c
h r
esearchers can explore their results. Vol-
um
e r
e
n
d
e
r
in
g, like that shown in figure 1,
enables researchers to easily identify regions of
hi
gh sc
al
ar d
issipation rate.
O
n
e c
h
all
e
n
ge in analyzing HCCI lies in track-
ing the ignition, propagation, and extinction of
l
oc
al i
gnition fronts. Researchers have adapted
th
e F
a
s
tBit in
d
e
xing techniques (sidebar “FastBit:
Indexing for Fast Searches,” p34) developed by the
S
c
i
entific Data Management (SDM) Integrated
So
f
t
w
ar
e In
f
rastructure Center (ISIC) to detect
46
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
In diesel engines, droplets of liquid fuel are
sprayed directly into the hot combustion
chamber, where they evaporate and form a
flame. Automobile engines are “heading towards
direct injection” as well, says Dr. Chris Rutland
of the University of Wisconsin (UW) at Madison.
With support from several combustion
research projects, the UW researchers have
simulated the injection of as many as one
hundred thousand droplets in a Direct Numerical
Simulation (DNS). They track the shrinking size
of the droplets so that they can quantify the
evaporation rate as well as the transfer of
momentum to the fluid. However, they model
the droplets as simple point particles moving
through the computational grid, simplifying the
complex fluid flow around each droplet.
The researchers designed simulation models of
the injection of n-heptane. This seven-carbon
molecule is a primary component of diesel fuel,
Dr. Rutland says, but “the complexity in chemistry
[grows] roughly with the number of carbon
atoms” because of the larger number of species
and reactions in the flame. Still, capturing the
detailed evaporation of droplets is critical to
describing the cooling of the gas near the jet tip.
Because of this cooling, ignition tends to occur in
a sheath surrounding the jet (figure 4).
Incorporating the physics of sprays into
large, three-dimensional DNS codes will enable
researchers to better understand the behavior
of diesel engines, as well as direct-injection
gasoline engines. Engineers believe this
understanding should help in designing cleaner
and more efficient engines.
Si mul at i ng Spr ays
Figure 4.
Two-dimensional simulations of a jet spray show the cooling caused by evaporation
near the tip of the jet (indicated by cooler colors), causing ignition to start near the edges of
the jet. The black line shows the position where the fuel and air are in stoichiometric ratio.
The expanded capability
of the S3D platform has
allowed combustion
researchers to simulate
more realistic problems.
C. RUTLAND, UNIVERSITYOF
W
ISCONSIN–MADISON
and track these local features. They do this by pro-
viding a definition of an ignition front based on
local conditions such as temperature and chem-
ical composition, which then allows them to
tr
ans
f
o
r
m the challenge of describing the com-
plex autoigniting mixture into one of tracking
in
dividual ignition features (figure 3, p32). Sophis-
ti
c
a
t
ed r
ules identify critical points that provide
t
o
po
l
o
gi
c
al in
formation about the turbulent
eddies and how they relate to the temperature and
r
e
a
c
tive species. By extracting these critical data
po
in
ts f
r
o
m o
th
e
rwise continuous data, combus-
tion researchers will be able to skeletonize the
d
a
t
a
, retaining only the salient features in an intel-
li
gib
l
e an
d un
amb
i
gu
ous form. By incorporating
information about the persistence of structures
in tim
e an
d b
y l
ooking over multiple length
sc
al
e
s
, th
e t
ec
hnique should identify and track
intermittent events of extinction and reignition.
Combustion Insights
Combustion occurs under many different condi-
tions, such as with or without premixing of the
air and fuel, under various concentrations and
chemical compositions, and using continuous or
repetitive ignition. The expanded capability of the
S3D platform has allowed TSTC scientists and
other researchers to simulate more realistic prob-
l
e
ms
, an
d h
as thereby provided insight into a vari-
ety of combustion conditions.
In th
e context of the HCCI process, researchers
h
a
v
e l
e
arned important details about how the igni-
ti
o
n f
r
o
n
t p
r
opagates under lean conditions. Dur-
ing traditional fuel-rich operation, an ignited flame
p
r
o
p
agates by deflagration, the process in which
th
e h
e
a
t o
f c
o
mb
ustion ignites nearby regions in a
steadily propagating front. The motion of the front
in
v
o
l
ves significant molecular transport. A differ-
e
n
t s
e
t o
f c
o
n
ditions could produce a flame front
that appears similar, but on a microscopic scale
in
v
o
l
ves many sequential, independent ignition
e
v
e
n
ts
. U
sin
g DNS, researchers at SNL discovered
that the local speed of the front indicates which
o
f th
e
se phenomena is occurring. When the tem-
pe
r
a
t
ur
e is m
a
de to vary with position to spread
the ignition out over time (figure 6) to avoid exces-
si
v
e p
ressure rise rates, they showed that both
t
y
pe
s o
f f
r
o
nts are present. Sequential ignition is
47
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
Dr. H. Im, Dr. A Trouvé, and collaborators have
included two additional physical phenomena in
flame simulation with the S3D code: thermal
radiation and soot formation.
Anyone can see that the hot gases of a flame
emit radiation in the form of visible light. More
important for combustion researchers, however,
is the infrared radiation they emit, which can
transfer energy over long distances. In contrast,
simulations traditionally include local heat
transfer by convection and conduction.
The researchers developed two different
models for radiation. The discrete ordinate
method (DOM) is easily implemented because
it directly exploits the grid structure. The
discrete transfer method (DTM) uses a ray-
tracing algorithm, and then determines the
radiation power at each grid point by a local
projection operation. Although the models are
established, they have not been widely
incorporated into DNS calculations.
The team also included a model for the
formation of soot, which occurs due to the
aggregation of incompletely burned fuel. This
aggregation begins with a few molecules, and
can grow to the large, black particles often
observed in truck exhaust. However, small,
invisible soot particles are of increasing concern
for their potential effects on both health and
climate. The virtually endless variety of sizes of
soot particles precludes any direct tracking as is
done for detailed chemical interactions. Instead,
researchers track the total mass fraction of soot
and the density of soot particles. The ready
absorption of infrared radiation by soot particles
causes local heating. This interdependence of
radiation and soot behavior is the reason the
two phenomena must be considered together in
models. Studies of turbulent non-premixed
flames showed that tracking of the local flow
and chemistry is critical for accurate prediction
of soot formation.
Soot and Radi at i on
Figure 5.
TSTC researchers have simulated turbulent sooting non-premixed ethylene-air flames in
an opposing jet geometry. Images from left to right correspond to vorticity magnitude, temperature,
and soot volume fraction.
Researchers recently
performed an
unprecedented three-
dimensional simulation of
a flame in the thin
reaction zones regime.
H. G. IM
, UNIVERSITYOF
MICHIGAN
C O MB U S T I O N S C I E N C E
prominent when the temperature is relatively uni-
form, while deflagration predominates when the
temperature variation is greater. This sort of micro-
scopic information is virtually impossible to glean
from experimental studies, and can guide engine
researchers to devise optimal strategies for mixture
preparation to control the rate of pressure rise.
Researchers at Stanford University used these
data to validate a flamelet model for combustion.
Their model relied on the results for the diffusion
of scalar quantities in the turbulent system.
Because of this diffusion, detailed comparison
between the results of their model and the DNS
simulation showed a much better agreement than
a traditional multi-zone model which includes
pressure coupling between zones but neglects dif-
fusive transport (figure 7). In this case, the DNS
calculations provide the detailed information
n
eed
ed t
o t
e
st a specific microscopic theory of the
combustion process.
L
ean premixed combustion is also found in tur-
b
in
e
s f
o
r power generation. The laboratory tur-
b
ul
e
n
t B
uns
e
n b
urner configuration (figure 8) is
useful to study the effect of turbulence on flame
s
tr
u
c
ture and propagation in parameter regimes
r
e
l
e
v
an
t t
o s
tationary gas turbines. Lean premixed
combustion has the advantage of higher thermal
e
ffi
c
i
ency and low NO
x
e
mis
si
o
ns. Lean flames
t
e
n
d t
o be b
r
o
a
der and propagate more slowly.
Hence, these devices operate at higher turbulence
in
t
e
nsit
y relative to the flame propagation speed.
T
his s
y
s
t
e
m o
f combustion is known as the thin
reaction zones regime, where small scale turbu-
l
e
n
ce can penetrate the leading edge of the flame
kn
o
wn a
s th
e p
r
e
heat zone, but not the reaction
zone. Thin reaction zones combustion has not
bee
n w
e
ll understood or modeled effectively in
th
e p
a
s
t
. R
e
searchers only recently performed a
48
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
Figure 6.
In these snapshots from two-dimensional simulations of HCCI combustion,
the initial temperature of the hydrogen-air charge varies from place to place,
encouraging earlier ignition in hotter regions. Heat release is shown on a rainbow
scale, varying from blue (in the regions of lowest heat release) through red to white. In
the upper left panel, turbulence is suppressed, so combustion propagates steadily
away from the initial ignition sites. In the upper right panel, turbulent flow speeds
combustion by mixing the fluid, but also disrupts the propagation of the combustion
front. The effect is less pronounced in the lower panels, where both the length scale
of the temperature fluctuation is smaller by a factor of four, and the size of the
simulated volume is reduced by the same factor.
Under the direction of Dr. Habib Najm of
Sandia National Laboratories, the
Computational Facility for Reacting Flow
Science (CFRFS) project has been building a
flexible modular platform for simulating
combustion and the problems of reactive
fluids. In particular, researchers have built tools
for high-order adaptive mesh refinement (AMR)
and chemical reduction within the framework of
the Common Component Architecture (CCA),
which was developed by SciDAC’s Center for
Component Technology for Terascale
Simulation Software (CCTTSS). This modular
platform supports various types of reacting
flows, including combustion, climate studies,
and other fields.
The primary work of this group is to provide a
modular structure for combustion compatible
with the CCA. One aspect of this is a library
that allows parallel implementation of adaptive
mesh refinement (a complementary use of
adaptive mesh refinement is discussed in
“Simulating Turbulent Flames,” p25).
Another CFRFS thrust is the automated
reduction of chemical complexity in reacting
flows. Researchers are using computational
singular perturbation (CSP) theory to
eliminate unimportant reactions. “Outside of
the flame front, most of the chemistry is
dormant or exhausted,” Dr. Najm says,
allowing researchers to simplify the chemical
modeling almost everywhere in the
simulation.
Achieving this goal requires an automated,
adaptive chemical reduction scheme, crudely
analogous to the adaptive mesh refinement in
the spatial domain. The idea is to render the
chemical system with “as much complexity as
needed and no more.” The research team has
explored the practicality of this idea in
simplified systems, and is evaluating its use in
full combustion simulations.
Comput at i onal Faci l i t y f or React i ng Fl ow Sci ence
E. R. HAWKESETAL., 2006, COMBUST. FLAME, 145, 145–159.
R. SANKARAN, E. R. HAWKES, J. H. CHEN, T. LU, AND
C. K. LAW
three-dimensional simulation of a flame in this
regime for the first time. These simulations clar-
ify the effect of the small-scale turbulence on the
flame structure that can affect the turbulent burn-
ing rate and flame stability near the lean limit. A
new mathematical formulation to describe the
upstream boundary conditions developed by
TSTC researchers was used to prevent spurious
pressure waves as turbulence is introduced into
the computational domain.
To simplify the chemistry of the methane-air sys-
tem, combustion scientists worked closely with
engineers at Princeton University to systematically
eliminate chemical reactions that are unimportant
under lean conditions and to reduce the chemical
stiffness. CFRFS researchers have also been devel-
oping chemical reduction software (sidebar “Com-
putational Facility for Reacting Flow Science,” p48).
They found that turbulent mixing increased the
thickness of the flame with important consequences
for the overall turbulent propagation speed.
In some cases fuel and air are introduced sepa-
rately into the combustion chamber, and turbu-
lent flow is needed to quickly mix the reactants.
Jet aircraft employ such “non-premixed” combus-
tion for safety reasons, and it occurs in direct injec-
tion diesel and gasoline engines as well. In
recognition of the growing power of DNS and the
S3D platform, combustion chemistry researchers
Dr. Chen and Dr. Evatt Hawkes received an award
of computing resources under the 2005 Innova-
tive and Novel Computational Impact on Theory
and Experiment (INCITE) program to explore this
regime. This work is the largest ever simulation of
turbulent combustion using detailed chemistry,
with over half a billion grid points used to trans-
port fifteen variables. The results showed that, in
contrast to the assumptions of current models, the
mixin
g p
r
oce
s
s of reactive species is not analo-
gous to the dissipation of kinetic energy from the
t
urbulence field. This is due to the strong interplay
be
t
w
ee
n r
eaction and molecular diffusion and to
th
e l
ar
g
e d
i
ff
e
rences in transport properties of the
species. The results also gave hints of new mech-
anisms f
o
r r
e
ignition subsequent to local extinc-
ti
o
n (si
d
e
bar “
I
N
CITE Project,” p51).
Cooling effects near the combustion chamber
w
all c
an s
tr
ongly influence the flame. Although
this c
oo
lin
g c
an be u
s
e
f
ul by helping to stratify
the charge in an HCCI device, under lean condi-
ti
o
ns th
e c
ooling may become excessive.
R
e
s
e
arc
h
e
rs led by the University of Maryland
group simulated a two-dimensional ethylene-air
jet, and found flame extinction near the walls.
Such extinction is a dominant mechanism for
emission of unburned hydrocarbons. The impor-
tance of walls will be even more crucial in com-
pact microcombustion chambers that are
49
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
Figure 7.
The stratification of the premixed HCCI charge into regions of different
temperature can spread the spontaneous ignition in time, making the heat release
less abrupt. When the temperature variation imposed on a simulation is large, like
that in figure 6, the spreading is significant. A flamelet-based theory successfully
described the resulting heat release much better than a competing multi-zone model.
Figure 8.
TSTC researchers have simulated the combustion of premixed gases in an
elongated Bunsen geometry, as illustrated here. The left image corresponds to
temperature, and the right image shows flame surface location of maximum heat
release.
Burned
Products
CH
4
/Air
800 K,φ = 0.7
ILLUSTRATION: A. TOVEY
SOURCE: D. J. COOK, H. PITSCH, J. H. CHEN, AND
E. R. HAWKES
0.
8
0.7
0.
6
0.5
0.4
0.3
0.2
0.1
0.0
0.3
0.4
0.5
0.6
0.7
0.8 0.9
1.0 1.1
Time/τ
0
Heat Release/HR
T′′ = 30K: Case 4
FLAMEL
ET
DNS
Mul
ti-Zone
C O MB U S T I O N S C I E N C E
c
ur
r
e
ntly under study by the SNL group and col-
l
a
bo
r
a
t
o
r
s at the University of Trondheim.
Futur
e Dir
ections & Summar
y
S
c
iD
A
C p
r
o
grams are vastly improving the power
of DNS of turbulent combustion, as part of a
b
r
o
ader program to improve the efficiency and
r
edu
ce th
e e
n
vir
o
nmental impact of these critical
technologies. Together with experiments, DNS
all
o
w
s researchers to discern the most important
f
a
c
t
o
r
s th
at affect the behavior of combustion. For
e
x
am
p
le, DNS clarifies the interactions between
c
h
e
mis
tr
y an
d t
ur
bulence, which are critical to
getting the most out of modern engine concepts
su
c
h a
s HCCI
.
B
e
n
c
hm
ar
k D
NS simulations also provide
highly detailed data to test models for engineer-
in
g sim
ul
ations. Such engineering analyses are
l
ar
g
e e
n
o
u
gh to include the details of engine
geometry, such as chamber shape, valve design,
an
d s
o f
orth. However, these calculations must
a
p
p
r
o
xim
a
te the behavior of the reacting gases
50
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
Opt i mi z i ng t he DNS S3D Code
Direct Numerical Simulations of combustion have
advanced tremendously in recent years, and in some
cases can simulate three-dimensional regions
centimeters in size (“Simulating Turbulent Flames,”
p25). However, it is still impossible to use DNS on the
scale of a complete combustor, for example to select
the best geometry for a reaction chamber. Small
details in the chamber can have an important effect on
the performance, such as by inducing additional
turbulent mixing at the inlet valves.
Instead, DNS accurately captures the finest-scale
details of the combustion over a reasonably large
volume, which can then be compared with theoretical
combustion models. In addition, the simulations are
used to validate large-scale engineering models of the
combustion process, in which finer-scale spatial or
temporal details of the flow are replaced by averages or
smoothed representations. Two important techniques
for doing this are Reynolds Averaged Navier-Stokes
(RANS) and Large Eddy Simulation (LES).
RANS creates a model for the larger problem by
averaging the local quantities entered into the full
Navier-Stokes equations describing fluid flow. In LES,
the largest eddies are directly simulated, while eddies
smaller than the grid are modeled.
Engi neer i ng Model s
Figure 9.
Large Eddy Simulation (LES) of a swirling
premixed flame in a laboratory-scale annular combustor.
The left side shows the instantaneous velocity field, while
the right side shows the time-averaged velocity field.
Direct Numerical Simulation (DNS) is an essential tool for
understanding the physics of turbulence. The S3D code
developed at Sandia National Laboratories (SNL) is a
massively parallel DNS solver for turbulent reacting flows
and includes multiple physical and chemical aspects such
as detailed chemistry and molecular transport.
In the past year, Dr. Ramanan Sankaran (SNL), Dr. Evatt
Hawkes (SNL), Dr. Mark Fahey (Oak Ridge National
Laboratory; ORNL) and Dr. David Skinner (National Energy
Research Scientific Computing Center; NERSC) optimized
several key kernels in S3D for both scalar and vector
architectures. As a result of the optimization, there was a
45% improvement in performance on NERSC’s IBM SP,
nicknamed “Seaborg”. Vectorization of S3D resulted in a
ten-fold improvement in performance on ORNL’s Cray X1E,
called “Phoenix”. Additionally, S3D has been shown to
scale remarkably well on several different platforms (figure
3, p45). Notably, S3D has 90% parallel efficiency on 5120
Cray XT3 processors at ORNL.
Writer:
Dr. Ramanan Sankaran, Post-Doc, Combustion Research
Facility, Sandia National Laboratories
Together with
experiments, Direct
Numerical Simulation
(DNS) allows researchers
to discern the most
important factors that
affect the behavior of
combustion.
J. C. OEFELEIN, SANDIA
NATIONAL
LABORATORIES
51
S
C I
D A C R
E V I E W
F
A L L
2 0 0 6
W W W
.
S C I D A C R E V I E W
.
O R G
(si
d
e
bar “
En
gineering Models,” p50). DNS allows
researchers to validate these approximations so
th
at they can best design the entire system and
c
o
n
fir
m mi
croscopic theoretical pictures for the
k
e
y p
r
oce
s
s
e
s that govern combustion.
As computers become more powerful and
r
e
s
e
archers continually enhance their code,
n
um
e
r
i
c
al sim
ul
ations of combustion will
become ever more detailed and physically realis-
ti
c
. Alth
o
ugh the dream of simulating an entire
c
o
mb
u
s
ti
o
n c
hamber is still elusive, current sim-
ulations give critical validation to the understand-
in
g o
f w
h
at governs combustion behavior and
h
o
w be
s
t t
o d
e
scribe it in engineering simula-
tions. In light of the economic and environmen-
tal impact of combustion, the importance of this
research is enormous. To put it into perspective,
a 50% increase in efficiency of automobiles would
result in over 20% savings in the nation’s petro-
leum consumption for transportation. Combus-
ti
o
n r
e
s
earch supported by SciDAC and other
programs will help to realize such rewarding
a
ccomplishments in the future.

Writer:
Don Monroe, Ph.D.
Further Reading
TSTC Project
https://cmcs.ca.sandia.gov:9443/sam/files/public/tstc/index.html
CFRFS Project
http://www.ca.sandia.gov/cfrfs/
E. R. Hawkes, et al. 2006. Direct Numerical Simulation of
ignition front propagation in a constant volume with temperature
inhomogeneities, part II: parametric study.
Combust. Flame,
145
: 145-159.
R. Sankaran et al. (in press). Structure of a spatially-developing
lean methane-air turbulent Bunsen flame.
Proc. Combustion
Institute
.
The Innovative and Novel Computational Impact on
Theory and Experiment (INCITE) program encourages
researchers to address some of the grandest challenges
in simulation. In 2005, researchers built on the success
of the TSTC and other Office of Science programs to
directly simulate the first 3D turbulent non-premixed
H
2
-CO-N
2
flame with detailed chemistry aimed at studying
the mechanisms of extinction and reignition. Such non-
premixed combustion occurs in aircraft engines and in
direct-injection internal combustion engines. The flow
must be highly turbulent to ensure adequate mixing of the
fuel and the oxidizer. However, this same turbulence also
increases the chances for local extinction of the flame
when turbulent strain produces mixing rates so large that
the reactions cannot keep up.
The INCITE-supported simulations clarified the process
of mixing, which is central to the combustion process. In
most large-scale engineering models of combustion,
scalar quantities such as gas temperature and
composition are presumed to mix at the same rate as
vector properties of the fluid flow, such as momentum.
For non-reactive scalars (those that do not participate in
the chemical reactions of combustion) the simulations
validate this assumption. However, the researchers
found that for light species such as molecular and
atomic hydrogen, as well as for scalar properties affected
by the evolving chemical environment, the mixing time
differed by as much as a factor of three.
Turbulence is needed to properly mix reactants for
combustion. However, if the turbulence is too great it
can separate two adjacent volumes of fluid, one
burning, the other containing unreacted material, before
the flame can propagate between them. In a steady
flame like those found in turbines, this can lead to blow-
out. Under the right conditions, however, the mixture
can reignite. Understanding the detailed processes of
extinction and reignition is critical to modeling
combustion in aircraft engines. The researchers have
identified important new features of the reignition
process, and continue to explore the 30 terabytes of
data that emerged from these simulations with help
from computer scientists.
I NCI TE Pr oj ect
Figure 10.
Using a large computer allocation provided
by the Innovative and Novel Computational Impact on
Theory and Experiment (INCITE) program, researchers
performed the largest Direct Numerical Simulation (DNS)
of non-premixed turbulent combustion to date. This
volume from the simulation of a H
2
-CO flame shows the
vorticity from high to low as red, yellow, and blue.
H. AKIBAAND
K. L. MA
, UNIVERSITYOF
CALIFORNIA–DAVIS