Download as pdf or txt
Download as pdf or txt
You are on page 1of 41

Ann. Rev. Fluid Mech. 1989.

21 : 345�5
ANNUAL
REVIEWS Further
Quick links to online content

NEW DIRECTIONS IN
COMPUTATIONAL FLUID
DYNAMICSl

Jay P. Boris
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Laboratory for Computational Physics and Fluid Dynamics,


US Naval Research Laboratory, Washington, DC 20375-5000

1. INTRODUCTION: CURRENT CAPABILITIES

1.1 Computational Fluid Dynamics in Perspective


Computational Fluid Dynamics (CFD) deals with the solution of fluid­
dynamic equations on digital computers and the related use of digital
computers in fluid-dynamic research. It is used for basic studies of fluid
dynamics, for engineering design of complex flow configurations, for
understanding and predicting the interactions of chemistry with fluid flow
for combustion and propulsion, for basic and applied research into the
nature and properties of turbulence, for interpreting and analyzing experi­
mental data, and for extrapolation into parameter regimes that are rela­
tively inaccessible or very costly to study experimentally.
Recent successes in computational fluid dynamics are based on inno­
vative numerical algorithms that discretize the continuum equations of fluid
dynamics into a large but finite number of algebraic or ordinary differential
equations. These equations are then solved using currently available, large­
memory, high-speed digital computers. A number of books, conference
proceedings, and review articles have recently been written on this subject.
Among them, CFD in aerodynamics is considered in National Academy
of Sciences/National Research Council ( 1 986), reactive flow in Oran &
Boris (1987), fluid mechanics and heat transfer in Anderson et al. (1984),
strongly shocked flows in Woodward & Colella (1984), and spectral
methods for subsonic flows in Canuto et al. ( 1 988). Books treating opti-

1 The US Government has the right to retain a nonexclusive, royalty-free license in and to
any copyright covering this paper.

345
346 BORIS

mization of CFD methods for vector and parallel processing on super­


computers have been edited by Book ( 1 98 1 ) and Rodrigue (1982), with a
somewhat more recent overview report by Ortega & Voigt ( 1 985). Numer­
ous proceedings of recent conferences on numerical methods in fluid
dynamics have come into print. Among these, those edited by Fritts et al.
( 1 9 85) and Morton & Baines ( 1 986) are noteworthy. Journals substantially
or cntirely dcvoted to CFD, such as the Journal of Computational Physics
and the International Journalfor Numerical Methods in Fluids, will contain
the most up-to-date ideas and simulations.
Fluid-dynamic systems ran be investigated in three ways: (a) analytic
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

solutions of the Navier-S)' .�es equations or other fluid-dynamic models


in tractable limits, (b) use of CFD to solve the mathematical models of
fluids on large digital computers, and (c) experimental observations of real
fluids as they occur in nature or in idealized laboratory systems where
specific processes and parameter regimes can be isolated more easily. Since
each of these approaches has its own inherent strengths and weaknesses,
all three approaches are necessary for a complete predictive understanding
of the dynamics of a fluid system.
Compared with analytic approaches, CFD requires relatively few restric­
tive assumptions and gives a complete description of the flow field for all
variables. Quite complex configurations can be treated, and the methods
are relatively easy to apply. For Newtonian flows accelerated by a scalar
pressure field, fluid-dynamic processes are well described by the Navier­
Stokes equations. Frequently, as in the case of reactive flow, additional
terms or equations are necessary to complete the description of the par­
ticular fluid system under study. Even though the mathematical models of
continuum fluids are compact and reasonably complete, their intrinsic
nonlinearity and the tendency for singularities to develop means that
closed-form analytic solutions describing the dynamics exist for only a
very few special systems. Therefore, approximations are the best that can
be expected from analytic treatmcnts of practical fluid systems.
Useful analytic approximations include steady-flow analyses, where time
derivatives can be neglected, and the use of these idealized steady flows
as a basis for linear perturbation analyses of periodic oscillations and
exponential instability. Another useful approximation employs simplified
geometry. Reducing problems to one or two dimensions allows application
of complex-variable transformation techniques in the potential-flow limit
and the use of similarity techniques in compressible and strong-shock
limits. Closed-form solutions are highly prized because the variation of
flow properties with changes in the controlling parameters is explicitly
displayed. Thus, approximate system optimization can be accomplished
dircctly with a minimum of computer time <:>r experimental trial and error.
COMPUTATIONAL FLUID DYNAMICS 347
Unfortunately, only idealized geometries can be treated this way, and then
usually with rather restrictive simplifying assumptions.
Useful approximate analytic solutions are not available for most prob­
lems, even in relatively idealized fluid-dynamic configurations. To treat
most fluid nonlinearities, time dependence, and realistic geometries in
designing appliances, automobiles, airplanes, ships, and engines, recourse
to numerics and experiment is usually required. Relative to experiments,
CFD has few Mach-number and scale limitations and is cost effective.
Simulations have an added advantage over experiments in that diagnostic
probing of a computer simulation does not disturb the flow and obscure
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

" "

the phenomena under investigation.


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Today, CFD solutions are often quite comparable to, or even exceed,
the accuracy and resolution of laboratory experiments. Nevertheless, since
engineers are often dissatisfied with existing experimental results and the
scalability of small experiments to full-scale systems, their reluctance to
trust the computers completely is quite understandable. Furthermore, no
matter how accurate a simulation is, it is not, and cannot be, the real
world. Indeed, it is through the differences between simulations, which
reflect our current mathematical theory of fluid behavior, and the real
world as seen through experiments that we learn new things about fluid
dynamics. Finally, although simulations are cost effective relative to lab­
oratory, field, and wind-tunnel tests, the costs and limitations encountered
in developing and applying simulations to complex fluid dynamic prob­
-

lems are still substantial.

1.2 The Current Status of CFD


Each advance in the CFD community's capability to perform a particular
class of computations has engendered a corresponding increase in the
engineer's expectations. These expectations, in turn, are quickly translated
into research and developmental requirements. To satisfy these needs,
CFD developers have provided improved numerical algorithms and other
diagnostic software advances. At the same time, hardware developers have
been providing a sequence oflarger, faster, more accurate, and "friendlier"
computers. Over the last few decades, improved computational repre­
sentations and algorithms have contributed at least as much to improve­
ments in the overall CFD capability as hardware development. Together,
hardware, software, and algorithm advances have increased the composite
capability by orders of magnitude. Early one-dimensional gas-dynamic
calculations were resolved with 40 or 50 spatial cells. Such problems can
now be performed with roughly the same resolution in each of three
dimensions in the same time as the old one-dimensional calculations,
requiring about four orders of magnitude more computing.
348 BORIS

There is still a long way to go. Solving for steady-state flow, really only
a mathematical idealization, over a realistic representation of a complete
automobile, airplane, or ship is still a formidable task. The turbulence
models used (for example, additional coupled Reynolds-averaged stress
equations) are at best limited approximations and often computationally
expensive. In the real world almost all of these flows are unsteady in the
important aspects of vorticity shedding and acoustic-vortex or wave-vortex
interactions. These unsteady phenomena, in turn, invalidate the procedures
used to derive most statistical turbulence phenomenologies.
Finite-difference meshes of a few hundred thousand grid points to
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

resolve the flow variables are expensive when carried for even ten or twenty
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

thousand time steps, so computations with a million cells or more are rarely
attempted. Finite-element algorithms are receiving increased attention for
CFD because of their more general geometric capabilities, but the cost of
inverting the global matrices that they introduce in time-dependent flows
is quite prohibitive. Spectral algorithms for CFD problems, based on
global expansions, are argued to be more accurate than comparably
resolved finite-difference and finite-volume algorithms, but they are corre­
spondingly more difficult to use in realistic geometries and generally
require more operations per mesh point. Full simulations of the Navier­
Stokes equations are giving way to the application of Large-Eddy Simu­
lations (LES) using monotone finite-difference or finite-volume algorithms
to treat fluid convection. The models can have viscous terms, but they
do not resolve the small scales of turbulence in most regions, replacing
statistically based turbulence models with locally time-dependent subgrid
turbulence models.
Problems in CFD today are tackled primarily on supercomputers such
as the Cray, Cyber, NEC, Fujitsu, and Hitachi machines, which rely on
vector registers and pipelines to obtain their speed and which have rela­
tively few processors. These computers are adequate for some steady three­
dimensional problems and some unsteady two-dimensional problems, but
they are generally too small, too slow, and too expensive to bring solutions
of unsteady three-dimensional problems within the range of most fluid
dynamicists.
Current attempts at unsteady three-dimensional (3-D) problems take
dozens to hundreds of hours on the biggest computers available and even
then have generally only limited spatial resolution. Extended simulations
of 3-D incompressible turbulence in a periodic box and in partly periodic,
wall-bounded channels (relatively idealized configurations) have been con­
ducted on meshes of up to 432 x 80 x 320 cells by, for example, Spalart
(1986, 1 988), Kim et al. (1987), Moser & Moin (1987), and Metcalfe et al.
(1987) but at relatively great cost. Because only a few installations can
COMPUTATIONAL FLUID DYNAMICS 349
perform such calculations for more than a few thousand time steps, snap­
shots of these computations are being archived in national data bases so
that a significant component of the CFD community can have access to
them.
While such idealized simulations have great value, there is an under­
standable requirement for calculations of more complex configurations
with more complex physics carried to much longer times. To relate the
simulated behavior of a reasonable fluid system to the new mathematics
of chaos and nonlinear system dynamics requires at least a few thousand
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

vortex sheddings to "flesh out" the system's computed return map. This
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

in turn requires millions of time steps, considerably farther than any of


the "high"-resolution 3-D models have ever been carried. We are only
now approaching runs of the required length in two dimensions. Winkler
et al. (1987) and Boris & Oran (1988) have conducted high-resolution
(typically 600 x 400 cells) two-dimensional (2-D) simulations of com­
pressible flows with shocks and complex vortex interactions. Though these
can be carried beyond a million time steps (e.g. Boris & Oran 1 987), to do
so requires the equivalent of hundreds of Cray hours of computer time.
There a re other prices to pay for these computations besides the long
wait for results. There tends to be a trade-off between supercomputer
power and ease of use-the so-called "user friendliness" of the composite
software/hardware system. Special programming techniques are invariably
required to get the most out of the available hardware. The choice of
problem representation, data structure, and algorithms is dictated inevi­
tably by the optimal implementation. New languages and new language
compilers, designed to help the CFD practitioner achieve a higher level of
program performance, do so at the cost of converting him to a software
guinea pig. Because the field is still advancing very rapidly, any CFD model
that is working reliably, flexibly, and accurately in today's supercomputer
world will be argued to be slow, inefficient, or obsolete by some academic,
laboratory scientist, or computer salesman.

1.3 Research Directions Using CFD


Unfortunately, the rapid advances in both algorithms and hardware
mentioned above appear to be approaching some intrinsic limits. Current
representations of continuous fluids are getting about as mu�h out of each
degree of freedom in the representation as can be expected. Current finite­
difference, finite-volume, finite-element, and spectral algorithms for solv­
ing partial differential equations have reached the point of diminishing
returns in terms of trading off computational cost for accuracy. It is now
more effective to increase the number of grid points to improve spatial
resolution and hence accuracy than to seek greater accuracy through
3 50 BORIS

higher-order algorithms. Even increasing the complexity of turbulence and


physical submodels is now less important than resolution improvements.
Current supercomputers, limited by the speed of light and the cost and
reliability oflarge-scale integration, are getting about as much out of single
processors as can be expected. Finding work for more pipelines in a single
processor with shared memory has also reached the point of diminishing
returns. In addition to the obvious difficulty of programming these com­
puters efficiently, there are hardware communications bottlenecks and
bottlenecks in scheduling the access of more and more arithmetic units to
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

a shared memory or to each other's stored data.


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Sections 2-5 of this review consider some of the new directions and
novel approaches currently on the horizon that attempt to circumvent
these limits. Four levels of new directions are identified and the strong
interactions between them discussed. New directions in representational
models of the fluid state are discussed in Section 2. These include cellular­
automata, hybrid-cellular-automata, and molecular-dynamics approaches
as well as novel approaches to decimating the fluid-dynamic equations to
a few dynamically significant degrees of freedom. New approaches for
discretizing (representing) continuous functions within the context of the
standard continuum-fluid models such as the Euler and Navier-Stokes
equations are also considered.
Algorithmic extensions and new directions are discussed in Section 3.
One must recognize and adapt to the existence of limits on how accurately
a continuous function can be represented and convected through a discrete
grid. Imperfect but optimal solutions to the simple convection equation
provide an intrinsic upper bound on the accuracy that conventional
methods can obtain. Useful, quite general existing algorithms that
approach this performance already exist. New algorithmic directions
include the extension of these near-optimal monotone methods to fully
adaptive and unstructured gridding to simulate general flow geometries,
the use of spectral elements to obtain better accuracy in moderately com­
plex geometries, and the use of fully Lagrangian algorithms with the goal
of avoiding numerical diffusion errors.
Section 4 focuses on evolving hardware and the related subject of par­
allelism in computational fluid dynamics. There is a very close connection
between the representation and algorithms chosen for a computational
model and its implementation on a particular hardware system. Some
mathematical models of the fluid state are better suited to one type of
hardware than another, and novel numerical algorithms will increasingly
be aimed at "parallelizing" the model's implementation. Indeed, the model
choices are often dictated by the most efficient implementation of the
model on a particular computer.
COMPUTATIONAL FLUID DYNAMICS 351
Future directions in terms of problem complexity and novel applications
of CFD to experimental fl u id systems and observations are considered in
Section 5. These are the connections between CFD and the real world. They
involve more complex problems, the use of phenomenological models, the
potential use of artificial-intelligence programs in organizing, writing, and
running CFD models, and the increasing use of CFD technology in experi­
ments and realtime systems. Since these four levels of new directions all
interact, the present is a very exciting time for CFD.
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

2. NOVEL REPRESENTATIONS
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

2. 1 Discretization in Computational Fluid Dynamics


The goal of CFO is to compute the correct time evolution of continuous
fluid variables such as p(x, t), where x represents the spatial variables
and t represents time. The evolution in time of the function p(x, t) is
approximated by updating a set of discrete cell values at a sequence of
times separated by discrete time steps. This is a fundamental idea in
computational fluid dynamics; spatial dimensions are divided into discrete
contiguous cells, usually called finite volumes or finite elements, and time
is discretized in short intervals called time steps. While such discrete
marching techniques are developed with an eye to time-dependent prob­
lems, they are also useful for seeking steady-state solutions. This discret­
ization of space and time is forced by conventional computers having
finite-sized memories segmented into floating-point words of data.
There are a number of ways to interpret the discretized variables in terms
of the conserved continuous quantities being approximated. Consider Xj'
the location of cell j. Depending on how one chooses to interpret Xj' it
could be

1 . the location of the cell center,


2. the location of the cell center of mass, or
3. a particular point in the cell where Pj happens to be known.

Consider pj, the cell value of p and Xi associated with some time t". This
can be

1. the value of p(x, t) at a particular position Xj at time t",


2. a characteristic value of p(x, t) near the specific location Xj at time t",
3. an average of the continuous solution in the volume enclosed by the
computational cell interfaces at a specific time t",
4. the average value in time of P between t"-1/2 and t"+ 1/2 in the cell Xi'
5. the average in space and time, or
352 BORIS

6. the coefficients of a set of basis functions in a linear expansion of the


dependent variable p(x, t).

Each discretization consists of both a set of values and a rule for inter­
preting these values.
Each of these interpretations has different properties, each approximates
some situations better than others, and each has different associated solu­
tion algorithms. The number of possible CFD representations and cor­
responding algorithms is enormous. The correct choice depends on many
properties of the problem being solved and the computer resources avail­
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

able. In all cases, however, there is only a finite number of discrete values
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

in the representation, and each value is only specified to finite precision.


This means that information is inevitably lost in the computational solu­
tion relative to the continuous problem being approximated. Errors in the
computation arise from uncertainty in the discretization. The source of
the uncertainty is the missing information about the detailed solution
structure within the discrete spatial cells and time steps. All approaches
that use a finite number of values to represent a continuous profile have
this problem.

2.2 Alternative Approaches to Discretizing Continua


Digital computers demand a discrete representation for continuous media,
but there are many ways the numbers in such a discretization can be used.
Alternate representations for the fluid state, each based in a different
discretization, can be conceived at four different levels: (a) the hardware
level, (b) the algorithmic level, (c) tbe mathematical level, and (d) the
physical-paradigm level. These four levels are not really easy to untangle
and discuss separately because choosing a paradigm or model at any one
level strongly influences the representational choices made subsequently
at all the lower levels. These alternate representations of the fluid are the
most fundamental of the new directions being manipulated in the hope of
bringing qualitative improvements to CFD. The possible payoffs are large.
The risks are correspondingly high, however, because major changes in
viewpoint are required and existing computer hardware is not optimized
to show the more radical of these alternatives off to best advantage.
The highest representational level is that of the physical paradigm.
Fluids are not really continuous on the atomic level, so it is natural to ask
whether computational discretizations based on the atomically discrete
nature of fluids might not be a useful approach. Cellular-automata and
closely related molecular-dynamics models for fluid dynamics are examples
of alternate physical representations based on this idea and are currently
receiving much attention. Section 2.3 briefly reviews lattice-gas cellular-
COMPUTATIONAL FLUID DYNAMICS 353
automata models. Section 2.4 considers molecular-dynamics discretiza­
tions founded even more directly on the atomistic basis of fluids.
Just beneath the physical level is the mathematical level-that is, how
the physical paradigm is expressed and manipulated mathematically. On
this level, moving Lagrangian representations of the usual continuum fluid
model are being studied to overcome numerical diffusion that occurs in
CFD models on a fixed Eulerian grid and to represent interfaces between
fluids and materials of different type. Section 3.5 below considers this
alternate representation.
Alternate algorithmic interpretations for a set of Eulerian cell densities
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

were used above to show how a discrete representation depends as much


on the rules of interpretation accompanying the cell values as on the actual
numbers themselves. Sections 3 .2-3.4 below look at extrapolations of
current trends and new directions on the algorithmic level. Monotone
algorithms are coming into widespread use. Designed to ensure that posi­
tive-definite physical variables such as mass density stay positive when
convected across a computational grid, these algorithms are approaching
optimal performance. Thus, variants of these monotone algorithms for
adaptive and unstructured gridding are being designed to extend their use
to complex, time-dependent geometries.
This algorithmic level of representation is not the lowest. At the even
more basic hardware level there is the possibility of representing the same
continuous function in alternate ways using essentially the same cell-value
definitions and interpretations. Cellular-automata models replace floating­
point numbers with single bits, macroscopic fluid variables being defined
as averages over many such bits. While it is questioned whether these
cellular automata are as accurate as the usual floating-point, grid-based
discretizations on a bit-for-bit basis, other hardware numerical paradigms
might also prove to be more accurate than a sequence of floating-point
cell values on a discrete grid.
Consider a scalar density p(x, t) varying between zero and some
maximum value Pmax on a finite interval (x" X2) . If the cells have uniform
width, only the density values associated with the cells need to be defined.
Twice as much computer memory is used if the cells sizes vary arbitrarily,
because the locations of the cell interfaces (or cell centers) become degrees
of freedom along with the cell densities. Because the cell size can be quite
large where the density varies slowly, however, the apparent additional
expense of arbitrary cell spacing almost always pays off.
Each of the cell-density values and cell locations in the usual repre­
sentations is usually specified relative to a single chosen origin. Thus,
many of the bits in each cell density and cell location are used to respecify
this offset from zero. If only the differences in density and the differences
354 BORIS

in cell interface location from one cell to the next were specified, numbers
of considerably lower precision could be used and many more of them
could be stored in a fixed amount of memory. As an example, consider
4096 bits of memory segmented into 64-bit real numbers. Only 32 cells are
possible if both the cell spacing and the density values are allowed to vary
reasonably. Although each of the stored values is accurate to one part in
1014, a general density profile is probably accurately represented to no
more than a percent or so at any particular point.
Consider these same 4096 bits divided up into 256 li ne segments of 16
bits each. Each line segment consists of two 8 bit numbers giving the
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

position and density displacements from one node to the next on the
continuous curve defining the profile. Both the density values and the
node displacements are quantized in this alternate "double-delta" rep­
resentation with an accuracy of one part in about 104, much coarser
than the cell values or locationsin the 64-bit conventional floating-point
representation. The overall specification of the profile, however, would be
at least 10 times more accurate because there are eight times as many
cells in the representation for the same number of bits. In addition true
discontinuities and even multivalued functions could be represented.
While there are also obvious drawbacks to this representation, such as
the need for careful attention to scaling and the inefficiency of current
supercomputers for executing this representation, these are no more pro­
hibitive than the implementation of cellular-automata models on con­
ventional computers. Indeed, it would be truly surprising if either I-bit
numbers (as the cellular-automata supporters claim) or 64-bit numbers (as
conventional CFD practitioners claim) turned out to be optimal.

2.3 Lattice-Gas Cellular-Automata Models


An alternate model to the Navier-Stokes equations for representing and
modeling fluid dynamics has been demonstrated recently in two spatial
dimensionsby Frisch et al ( 19 86) . This model is called a "lattice-gas
.

automaton" and is a particularly interesting example of a general class of


"cellular-automata" models for carrying out extensive complex calcu­
lations based on simple, local underlying rules. Figure 1 is a graphical
depiction of the hexagonal lattice for this standard cellular-automata rep­
resentation. The "atoms" of a lattice gas are computer idealizations of the
atoms of a real fluid (gas or liquid) bouncing back and forth in their random
thermal motions. The lattice-gas atoms conserve number, momentum, and
energy as in a real fluid. Hasslacher (1987) and Shimomura et al. (1987)
present a particularly readable presentation of cellular-automata fluid
models.
The allowed motions and interactions of the lattice-gas atoms are very
COMPUTATIONAL FLUID DYNAMICS 355
Two-Body Collisions Three-Body Collisions
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Figure 1 Schematic of the 2-D hexagonal lattice-gas cellular-automata grid with typical
collisions illustrated.

limited and stylized compared with their physical counterparts in order to


eliminate the cost of computing expensive collision integrals requiring high
numerical precision. Instead, collisions are implemented by simple local
rules requiring at most a few bits of precision. The hope is that the logic
required can be built into a special-purpose, large-scale integrated circuit
with many replications of the basic cellular automaton on a single chip.
Lattice-gas models are usually deterministic. The lattice-gas atoms of
the simulation reside on the nodes of a regular discrete lattice, typically at
the corners of closely packed squares or hexagons, as illustrated in Figure
1 for two dimensions. The atoms move only between adjacent nodes during
any one time step, so the atoms are limited to one of a few very specific
discrete velocities capable of carrying it exactly to one of the nearby lattice
sites.
The theory of this artificial system as a model for fluid dynamics is based
on the fact that enough of these highly stylized atoms interacting on a fine
enough lattice will statistically approach the macroscopic behavior of an
ideal fluid as the number of lattice nodes becomes large and the resolution
of the lattice becomes sufficiently fine. The practical application of these
cellular-automata models as an alternative to solving the continuum models
for fluid dynamics depends crucially on how quickly lattice-gas model be­
havior approaches that of the Navier-Stokes equations, known to approxi­
mate macroscopic behavior of real gases and fluids quite accurately.
To understand the physical basis for interest in lattice gases, recall that
fluids can be described at different levels of description or "chunking" (see
356 BORIS

Hofstadter 1979). At the atomic or molecular level, a swarm of individual


atoms collide according to essentially reversible laws of mechanics. At a
higher level of description, the enormous number of degrees of freedom
associated with tracking all th e atoms are replaced with distribution func­
tions of these particles. This reduces the number of degrees of freedom of
practical importance substantially but formally increases the total number
in the problem to infinity by averaging over instantaneous particle attri­
butes. This averaging process is responsible for dynamic irreversibility.
Further averages over these particle kinetic equations reduce the fluid
model to the approximate but much more tractable macroscopic fluid
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

equations, where the medium appears as a continuum.


Deriving the mathematical model of this third level of description is one
of the triumphs of the human intellect: the development and solution of
partial differential equations. Nevertheless, we have to solve these non­
linear fluid equations numerically when the problem is too hard to solve
analytically To do this, the continuous fluid equations have to be dis­
.

cretized numerically on some sort of grid or lattice because the number of


degrees of freedom is formally infinite. Instead of discretizing a set of
continuum equations, which are just smoothed versions of the discrete
atomic or molecular-dynamics equations anyway, the proponents of cellu­
lar automata argue that we should use simplified versions of the underlying
deterministic Newton's laws for the atoms and molecules. By averaging
over the locations and velocities of many small lattice-gas atoms in a small
region, approximations to the macroscopic fluid variables in that region
can be obtained.
The potential advantage of cellular-automata models is their suitability
for implementation in massively parallel computers, where, for example,
a billion closely linked microcomputers could all chew on one big problem
simultaneously. The practical hurdle that these new lattice gas models
must overcome is the need to dedicate very many degrees of freedom to
obtain solutions smooth enough to compare with experiments and with
conventional 32-bit and 64-bit calculations of moderate- and high - R ey­
nolds-number flows (see e.g. Yakhot & Orszag 1986, Dahlburg et al. 1987).
There are additional problems with scaling these simple lattice-gas
models to low density, to low Mach number, and to three-dimensional
flows that must be overcome before special cellular-automata computers
can compete with conventional supercomputers in simulating realistic
flows. On the other side of the ledger, cellular-automata algorithms have
been invented for the convection of passive scalars by Chen & Matthaeus
(l987a), for immiscible fluids with interfaces by D. R othman & J. Keller
(private communication),'and for magnetohydrodynamics by Montgomery &
Doolen (1987) and Chen & Matthaeus (1987b). Though these automata
COMPUTATIONAL FLUID DYNAMICS 357
models have not yet achieved the full range of physical flexibility of their
floating-point representation counterpoints, a very important question to
answer is, how far can this new CFD direction be taken?
History, perhaps, has a lesson to teach here. A number of years ago
Symon et al. (1 970) considered a cellular-automata model to solve the
Vlasov equation, a close relative of incompressible fluid equations. This
model was descriptively called "bit pushing" and was shown to be
equivalent to a simple continuous plasma-simulation algorithm called the
"nearest-grid-point" algorithm if the latter were performed with limited
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

resolution. The nearest-grid-point algorithm was in heavy use at the time,


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

but neither of these algorithms is in use today. The statistical fluctuations


in the particle density, in that case the electrons and ions in a flowing
plasma, were simply too large because the number of particles was so far
below that needed to model a real plasma. Today, plasma is simulated
using discrete particles, but these are represented as smoothly moving,
smoothly overlapping patches with at least bilinear floating-point inter­
polations to help smooth the unacceptable numerical fluctuations.
Pure cellular automata do offer real promise for contributing to a deeper
understanding of more universal and abstract processes in causal, com­
putational, and cognitive systems. Wolfram ( 1 985) observes that "there is
a close correspondence between physical processes and computations. On
one hand, theoretical models describe physical processes by computations
that transform initial data according to algorithms representing physical
laws. And on the other hand, computers themselves are physical systems,
obeying physical laws." Cellular automata provide a flexible tool for
probing not only the limits of physical computability but also, through
the study of locally connected "neural networks," the massivety parallel
self-organization of information that takes place in the brain (see Hof­
stadter 1979).

2.4 Molecular Dynamics


In the real world, fluid dynamics arises as the statistical average of the
motion of a very large number of molecules. The collisions between mol­
ecules generally conserve mass, momentum, and energy in the classical
sense. When quantum-mechanical internal degrees of freedom play a role,
they are generally absorbed into macroscopic material properties, such as
the equation of state relating the thermodynamic energy density to the
pressure. An attractive extension of the ideas of cellular automata is to
use floating-point numbers to represent the location, velocity, and forces
that accelerate the gas atoms. Since a continuum of collisions is possible
compared with the handful of discrete collision possibilities in cellular­
automata models, using molecular dynamics as a paradigm for CFD
358 BORIS

clearly removes the worst of the nonphysical aspects of cellular-automata


models, such as the anisotropy of the effective pressure tensor.
Indeed, molecular dynamics has been used for years to study the con­
nections between measured or calculated interparticle force laws and the
collective properties of an ensemble of the particles. Dimer formation
leading to the condensation of gases and dense-gas corrections to the
equations of state are among the macroscopic quantities that can be
simulated by molecular dynamics. To date, the main use of molecular
dynamics has been as a bridge between the microscopic and the macro­
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

scopic representations of matter. The new direction here is the notion


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

that computers have finally become fast enough and large enough for
molecular-dynamics models to extend all the way to the fluid regime in a
practical manner.
The molecular-dynamics approaches to fluid dynamics usually take a
greatly simplified force law designed to minimize the computations in
exactly the same spirit but to a lesser degree than cellular-automata models.
The trade-off is computational simplicity in the cellular-automata models
versus physical veracity in the molecular-dynamics models. Greenspan
(1985) has considered a num ber of the basic questions associated with this
use of molecular-dynamics techniques for fluids, such as convergence to
continuum behavior as the number of particles is increased. Rapaport &
Clementi (1 986) have carried the approach even further to demonstrate
eddy formation in obstructed fluid flow, as shown with cellular-automata
models by Shimomura et al. (1 987).
The basic computational problems in this molecular-dynamics approach
are also the same as with cellular-automata models. The number of par­
ticles required to reduce the level of statistical fluctuations to an acceptable
value in even small volumes of the fluid is usually unacceptably large. In
addition, there is another major computational problem in molecular­
dynamics modeling, namely determining the neighbors that are near
enough to be significant. In the cellular-automata models this problem is
absent because only thc neighboring lattice sites are potential sources of
interaction.

2.5 Near-Neighbors Algorithms


A major part of a molecular-dynamics calculation is the continual need to
determine which nodes are near neighbors in a large set of nodes that often
seem to be moving randomly. The expense arises because each node can
potentially interact with any of N-l other nodes in the system. This is
called the N2 problem because only a few of the N2 interactions possible
between pairs of nodes are potentially important. This near-neighbors
COMPUTATIONAL FLUID DYNAMICS 359

problem has been the subject of much research and has been reviewed, for
example, by Hockney & Eastwood (1981).
The Monotonic Lagrangian Grid (MLG) is a new method designed to
beat the N2 problem (Boris 1986a, Lambrakos & Boris 1987, Lambrakos
et al. 1988). The MLG is a compact data structure for storing the positions
and other data needed to describe N moving nodes. A node with three
spatial coordinates has three indices in the MLG data arrays. The data
relating to each node are ordered in the memory locations indicated by
these indices such that nodes that are close to each other in real space are
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

automatically near neighbors in the MLG data arrays. This simplifies


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

and speeds near-neighbor accounting for molecular-dynamics, Lagrangian


cellular-automata, and continuum Lagrangian CFD models. Though
lattice-gas models have been implemented on a fixed lattice to date, a
sparse lattice gas, capable of great variations in density, also is a natural
application of the Monotonic Lagrangian Grid representation. Only
the relatively small fraction of active lattice-gas atoms would have to be
represented. The unfilled lattice sites could be squeezed out of the descrip­
tion. Such an approach has been applied to cellular-automata and mo­
lecular-dynamics models by Lambrakos & Boris (1987). These alternate
fluid representations do not suffer the discontinuous grid fluctuations of
the fluid Lagrangian models, but instead suffer severe fluctuations of a
statistical nature, as discussed earlier.
A computer program based on the MLG data structure does not need
to check N-I possible distances to find which nodes are close to a particular
node. The indices of the neighboring nodes are automatically known
because the MLG node indices vary monotonically in all directions with
the Lagrangian node coordinates. The cost of the algorithm in practical
calculations is dominated by the calculation of the interactions of nodes
with their near neighbors, and the computation time scales as N.
Figure 2 shows an example of a small, two-dimensional MLG. Because
spatial coordinates define a natural ordering of positions, it is always
possible to associate two grid indices with a set of random locations in a
two-dimensional space. The MLG shown is one of the many possible 4 x 4
MLGs linking 16 nodes distributed nonuniformly in a two-dimensional
region.
The two-dimensional MLG shown satisfies the following indexing
constraints:

x(i,}) � x(i+ I,}) i = 1, ... , N - I, } = 1, ... , N,

y(i,j) $ y(i,j+ 1) i = 1, . . . , N, j = 1, . . . , N-1.

The 16 nodes are shown at their irregular spatial locations in Figure 2, but
they are indexed regularly in the MLG by a monotonic mapping between
360 BORIS

The MLG in Two Dimensions

0 p
"'- -- AI2. T- t<I_1--I- 414)
1{1, ) V 3,�) Vi.
I(V V l- I ,3 4 M 0 N P
II'. I- 2. �}.- -I"- r.tI- \
,11M 33}l"-t-... H 3 I K J L
y r--...G
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

V 7- , \
11: !t"r-� V F 2 G E H F
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

II U,2 lk [4, )
I Via. (. 1 1 B A C 0
BII 1/
11!f1-,"- A 1 2 3 4
2, )

space memory
Figure 2 An example of a two-dimensional Monotonic Lagrangian Grid.

the grid indices and the locations. Each grid line in each spatial direction
is forced to be a monotone index mapping. Nodes stored in the data
memory according to this prescription are at most two or three nodes
away from the neighboring nodes that can affect them. Thus, for gradient,
derivative, or force calculations, only a fraction of the possible node-node
interactions need to be considered. No search is necessary to locate near
neighbors, and the necessary logic and computation are ideal for parallel
or multiprocessing methods using this data structure. Computations of
interactions are only made between nodes located in a small contiguous
portion of computer memory. Although this approach results in com­
puting interactions for a few distant nodes, it provides a substantial
reduction in computational cost. The computations can be vectorized
efficiently because nodes that are close to each other are indexed through
contiguous memory.
A construction algorithm that scales as N log N can be used to build an
MLG from randomly located nodes. An MLG can always be found for
arbitrary data, although it is not necessarily unique. Further, when node
motions in real space destroy some of the monotonicity conditions,
another, faster N log N algorithm exists to restore the MLG order. Using
these MLG algorithms has reduced the time it takes to do a molecular­
dynamics calculation. On the eRAY, restructuring the MLG when the
nodes have moved appeciably typically takes 3-5% of each time step. In
COMPUTATIONAL FLUID DYNAMICS 361
addition, the MLG is fully vectorizable and can make optimum use of
parallel computers. Thus, it is also being investigated as the basis for fast
free Lagrange models of continuum fluid dynamics, as described in Section
3.5.

3. NOVEL ALGORITHMS
3.1 Numerical Solution Limitations
Accurate numerical simulation of a complex fluid system can only succeed
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

with a computational approach that stresses understanding and accurately


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

models the fundamental physics and fluid dynamics of the system. The
computational model must begin with the minimum set of processes and
mechanisms thought to reproduce the essential features of the known
analytic solutions and observations. Each of the major processes must be
treated by algorithms optimized to reproduce the qualitative and quan­
titative aspects of that process accurately, flexibly, robustly, and efficiently.
The overall model is constructed by coupling these individually optimized
algorithms together. Detailed numerical simulations constructed this way
eliminate the need for phenomenological "fudge factors" when bench­
marked against analytic solutions and careful experiments. Because they
focus on the underlying processes and mechanisms, these simulations make
much better fluid-dynamic "experiments" than do "kitchen-sink" models,
where a large number of approximate phenomenologies are thrown
together.
Fluid-dynamic convection in the absence of strong physical diffusion
effects is the most difficult flow process to simulate and thus is the pacing
limitation in CFD. The attempts to circumvent this primary difficulty have
spawned the extensive field ofCFD. Turbulence, mixing, and gas-dynamic
shocks are obvious practical instances illustrating this limitation. Extensive
research over the last four decades has continually sought improved algo­
rithms to solve continuity equations using finite resolution grids. Different
representations of the convective process have been used, and many differ­
ent algorithms have been proposed within each of these representations.
General background references include Courant et al. ( 1928), von Neu­
mann & Richtymyer ( 1950), Godunov ( 1 959), Lax & Wendroff ( 1 960,
1 964), Harlow ( 1 964), MacCormack ( 1 9 7 1 ), Potter ( 1 973), Roache ( 1 982),
Gottlieb & Orszag ( 1 977), Anderson et al. ( 1 984), Fritts et al. ( 1 985), and
Oran & Boris ( 1 987).
The inevitable limitations imposed by computer size and speed are
reflected in the need for improved accuracy with a fixed number of degrees
of freedom (e.g. grid points) in the representation, and the need to treat
more general and realistic physical systems. To cope with these limitations,
362 BORIS

the new directions in algorithms involve sophisticated adaptive gridding


to maintain high-resolution at gradients in c omple x flow fields, appli­
cations of finite-element methods with unstructured grids for complex flow
geometries, very fast near-neighbors algorithms for Lagrangian CFD and
molecular dynamics, more sophisticated methods for coupling models
of physical processes with disparate time scales, and software specially
optimized for parallel computer architectures. While there is long-term
promise in new representations and new architectures, it is a low-risk, high­
payoff approach to work on improving existing Eulerian and Lagrangian
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

finite-difference algorithms. Three promising new algorithms are described


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

below: Finite-EleMent Flux-Corrected Transport (FEM-FCT) for CFD


in complicated moving doma ins spectral-element methods for accurate
,

solutions of smooth flows in moderately complex geometries, and free­


Lagrangian-dynamics algorithms for fluid systems with moving and multi­
phase interfaces.

3.2 Optimal Monotone Methods


The key to simulating complex, compressible, time-dependent flow prob­
lems with steep fluid�dynamic gradients such as interfaces between tur­
bulent and nonturbulent flow, flames, or shocks is to use a nonlinear,
monotone, positivity-preserving method. The original methods of this type
were called Flux-Corrected Transport (FCT) (see Boris 1 97 1 , Boris & Book
1 976, Zalesak 1 979). Subsequent high-accuracy algorithms incorporating
monotonicity ideas include MUSCL by van Leer ( 1 979) and the Piecewise
Parabolic Method (PPM) of Colella & Woodward (1984). Recently, a
number of closely related Total Variation Diminishing (TVD) schemes (e.g.
Harten 1983, Sweby 1984) have been assembled out of slope-limiting and
flux-evaluation concepts introduced in FCT, MUSCL, and PPM . This
general class of monotone algorithms does not allow unphysical maxima
or minima resulting from a lack of numerical resolution to grow in a
computed profile. These algorithms are generally positivity preserving, in
the sense that originally positive quantities like mass densit y cannot
become negative unphysically, and they are intrinsically nonlinear. They
have been used extensively on quadrilateral finite-difference grids, and they
can easily be used with nonorthogonal grids.
With years of development already invested in monotone convection
algorithms, it is appropriate to ask whether this technology can be taken
appreciably farther. How much better can convective-flow algorithms get?
Obviously an imperfect but optimal solution exists, since an infinitely
accurate representation of a general continuous profile is impossible using
a finite number of degrees of freedom. This optimal solution depends on
the definition used for optimal and on the conditions that must be met by
COMPUTATIONAL FLUID DYNAMICS 363
the solution. It has finite error arising from insufficient resolution at steep
gradients.
A fluid profile consisting of a positive delta function, with value one at
one cell location and value zero at all others, is composed of many sine
waves. The Gibbs phenomenon forces the smoothest continuous function
constructed through this profile to have oscillations with nonphysical
extrema and even negative values between the specified cell values. Faithful
convection of this interpolated continuous profile allows the negative
values and spurious extrema associated with finite resolution to "appear"
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

on the grid in a few time steps. To protect against these unacceptable


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

errors, a convection algorithm must spread the delta-function profile so


that the interpolating profile in the nearest discrete representation has no
negative values between the discrete cell values. This spreading, to smooth
out the Gibbs phenomenon oscillations, gives the "optimal" discrete profile
a finite error relative to the infinite-resolution limit (see Oran & Boris
1 987).
The optimal solution for convection is quite good, typically an order of
magnitude more accurate than classical finite-difference techniques for
solving the continuity equation. The better nonlinear monotone algorithms
approach this optimal solution quite closely, typically well within a factor
of two of the minimal error. Thus, not much additional improvement is
possible compared with the gains that have already been made for fluid­
flow simulation using an essentially Eulerian grid.
The obvious next questions are, can algorithm design possibly circum­
vent this apparently unavoidable limit, and what are the other costs
and problems? Six approaches are possible to get around this accuracy
limitation based on numerical resolution, in effect a kind of uncertainty
principle:

1 . ignore the positivity/mono tonicity conditions of real convection


entirely or in part,
2. use weakly unstable algorithms to keep underresolved gradients
steep,
3. use special, problem-specific information available outside the
numerical algorithm to adjust the numerical solutions,
4. use Eulerian algorithms with additional degrees of freedom per cell,
such as specifying both a value and a slope in each cell,
5. use Lagrangian algorithms to transform the convection errors away,
or
6. use Eulerian algorithms with added Lagrangian characteristics, such
as additional degrees of freedom, to track the exact location of fluid­
dynamic interfaces, steep gradients, or discontinuities.
364 BORIS

The first two of these approaches are, in fact, often used but cannot be
relied upon to give accurate solutions of fluid-dynamic equations. The
third approach is quite useful when additional information about the
solution is available. Each of the remaining ways of circumventing the
accuracy limits imposed by resolution requires additional degrees of free­
dom in the numerical representation. Which of these last three approaches
to adopt is problem dependent and involves trading off many factors.
Sections 3.3-3.5 describe algorithms representing these three promising
avenues.
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

3.3 Adaptive and Unstructured Gridding


Finite-element techniques have been used for years to solve difficult, prac­
tical problems in structural engineering. More recently, finite-element
methods have been modified for a variety of problems in fluid dynamics
and heat transfer, and, most recently, for time-dependent fluid dynamics.
The book by Zienkiewicz & Morgan ( 1983) is a good introduction to the
use of finite elements for solving partial differential equations. Triangles
and tetrahedra are commonly used for finite-element grids in two and
three dimensions because they can be used to describe irregular objects
or strangely shaped internal boundaries and interfaces. Combining the
triangle-based, unstructured grid approach of a finite-element method with
the accuracy of a nonlinear, monotone finite-volume method brings near­
optimal accuracy and flexibility together in a single algorithm.
A major drawback of finite-element methods is that they can be very
expensive. It is usually necessary to solve a linear matrix problem at each
time step, and the cost in computer time for large matrix inversions at
each step is prohibitive. In multidimensions, finite-element methods also
require a great deal of computer storage because the cost of recomput­
ing geometric and shape functions for each variable is also prohibitive.
Thus, these variables are usually stored rather than recomputed, requiring
dozens of quantities per node in two dimensions and many more in three
dimensions.
There are also compensating advantages to hybrid finite-element, finite­
volume methods. They are relatively straightforward to generalize to multi­
dimensional problems in several dependent variables. The method is
well suited to problems with irregular boundaries. The domain can also
change in time to allow improvements in resolution where needed or to
allow for deformable structures in the flow.
Two new approaches to solving time-dependent fluid-dynamics prob­
lems deserve mention. These are the moving finite-element method (e.g.
Miller & Miller 1 98 1 , Gelinas et al. 1 98 1 , Djomehri et al. 1 988) and the
hybrid monotone method developed by Lohner et al. ( 1 985, 1 988). Both
COMPUTATIONAL FLUID DYNAMICS 365
of these are finite-element methods using an expansion in linear "tent"
functions that are, or can be made, multidimensional. However, there are
fundamental differences in the two approaches.
In the moving finite-element method, node positions are treated as
dynamic degrees of freedom. This couples ordinary differential equations
for node positions to the integration equations for the usual physical
variables. This method produces "best" values for the positions of the
nodes in addition to updating the variables at the nodes. However, there
are many more variables involved and the resulting equations are stiff, so
the computational cost can be substantial. Nodes tend to migrate toward
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

regions where there are abrupt changes in the variable values. Thus,
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

problems characterized by advancing surfaces are efficiently solved by such


methods because the number of nodes can be minimized, even though a
relatively large amount of computation is required for each node.
Finite-element methods have not been used as long for fluid calculations
as have finite-difference and spectral methods. In general, the number of
operations per finite-element node substantially exceeds that per mesh
point in finite-difference or finite-volume methods, so finite-element
methods must be more accurate if they are to be competitive. The algo­
rithms of Lohner et al. ( 1 985, 1 9 88), which incorporate the monotone FCT
algorithm into a finite-element framework, are competitive with finite­
difference methods because matrix inversion is avoided. An example of a
typical solution using this method, a shock passing over irregular objects,
is shown in Figure 3. In this method the mesh also adapts in time, but the
only dynamic degrees o f freedom are the values of the dependent variables.
The node locations are static degrees of freedom. The cells are subdivided
in front of advancing gradients and flow structures and recoarsened behind
them according to rules based on gradients in the known properties of the
evolving system. These algorithms and the adaptive-gridding procedure
combine the flexibility of finite elements with some of the robustness and
accuracy features of monotone finite-volume methods.
One drawback of this approach is the amount of computer memory per
mesh point that it requires, a serious issue when memory is at a premium.
However, these methods are accurate, relatively fast, and allow very flex­
ible gridding of complex geometries. Extensions to three dimensions are
becoming practical with the advent of computers with large high-speed
memories.

3.4 Spectral-Element Methods


Another class of algorithms designed to handle complex geometries with
structured grids has been called spectral elements by Patera ( 1 984). Spectral
elements combine the geometric capabilities of finite-element methods
366 BORIS
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Figure 3 Finite-element FCT calculation of a shock over two irregularly shaped obstacles.
(Figure is courtesy of R. Lohner).

with the potential for accuracy of spectral methods. In a spectral-element


method the computational domain is divided into small subdomains of
nodes supporting polynomial basis functions of high degree. Typically
5 x 5 nodes to 1 5 x 1 5 nodes in two dimensions are used for the flow
variables in each spectral element. These elements are "patched" together
at their common interfaces using techniques introduced by Orszag (1980)
to cover the entire computational domain. Because the expansion functions
are global within each subdomain, a high order of accuracy is obtained in
each spectral element. This delocalization is confined to the individual
subdomains, however. Spectral-element methods are reviewed by Canuto
et al. ( 1 988) from the perspective of spectral methods.
Spectral-element approaches offer a reasonable compromise between
COMPUTATIONAL FLUID DYNAMICS 367
local representations using a spatial grid and fully delocalized represen­
tations using a spectral expansion. They have potential advantages over
fully structured grids and completely unstructured grids, though the com­
putations tend to be more extensive than for low-order finite-volume
methods. Since the spectral elements can be given severe distortions, vir­
tually any complex geometry can be built up from properly connected
spectral elements. Good examples of this for incompressible flow are given
in Karniadakis et al. ( 1 986). The distortions of each spectral element are
introduced by isoparametric mappings, as shown by Korczak & Patera
( 1 986), so high-order approximations are possible in principle even when
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

the element's distortion is severe.


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

The drawbacks of spectral elements are their intrinsically nonlocal


nature, which incorrectly moves information ahead of propagating shocks,
and the fixed structure underlying the connectivity of the elements. These
hardly seem essential limitations and may be overcome by extensions to
existing techniques (Zalesak 1981). McDonald et al. ( 1 985) and McDonald
( 1 988) have extended positivity-preserving convection techniques to spec­
tral algorithms by generalizing the Zalesak (1 979) flux-correction pro­
cedure. This could presumably work as well on spectral elements as on
the global basis used first by Zalesak. In the same sense, an unstructured
grid technOlogy such as that introduced by Lohner ( 1 987) or Jameson et
al. ( 1 986) could be used to manipulate the connectivity of the macroscopic
spectral elements, rather than the connectivity of individual triangles and
tetrahedra as in current models.

3.5 Free Lagrangian Algorithms


Lagrangian algorithms appear to offer substantial gains over Eulerian
algorithms through the elimination of numerical diffusion in the solution
of continuity equations. A Lagrangian frame of reference moves with the
local fluid velocity so that the advective derivative is transformed away,
leaving a system of ordinary differential equations without convective
partial derivatives that is attractive for numerical calculation. Because
the convective derivatives are absent, finite resolution does not require
smoothing the computed profiles (numerical diffusion) to ensure posi­
tivity of mass and energy densities. Nodes of a Lagrangian grid that start
out on an interface move with that interface. Thus, Lagrangian methods
'
also appear to be ideal for heterogeneous and multiphase flow problems.
These potential gains are not free, however. Lagrangian methods simply
have other problems associated with them. Traditional applications of
Lagrangian algorithms used quadrilateral grids, which rapidly distort
when simulating a rotational flow. A twisting vortex, for example, distorts
a quadrilateral Lagrangian grid to the point where line segments cross
368 BORIS

each other, making a unique definition of cell volumes impossible. Though


regridding restores grid regularity, it requires an essential interpolation
that reintroduces numerical diffusion.
To get around this regridding problem for structured Lagrangian grids,
the obvious extension to unstructured Lagrangian grids of reconnecting
triangles and tetrahedra was explored by Crowley ( 1 97 1), Fritts & Boris
( 1 979), and Trease ( 1 985). The increased programming complexity and
slower execution speed required by these methods did not, however, bring
the expected improvements in accuracy. Though good performance was
obtained on some problems, reconnecting the grid introduced significant
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

fluctuations into the simulation that do not appear in Eulerian calculations.


The numerical diffusion of a sharp flow feature acroSS a few cells in
Eulerian representations did not occur, but other errors also related to the
finite resolution of the grid did.
There are other Lagrangian paradigms for simulating fluid motion in
addition to solving the partial differential Navier-Stokes equations. The
lattice-gas and the molecular-dynamics approaches to fluid dynamics
described above are Lagrangian in underlying philosophy. Solving the
Lagrangian fluid equations, however, does avoid the problem of statistical
fluctuations that plagues both the cellular-automata and the molecular­
dynamics approaches to CFD . Thus, research aimed at faster and
smoother Lagrangian algorithms is currently of great interest. Fritts et al.
( 1 98 5) have edited a volume entitled "Free Lagrangian Methods" that
contains pointers to much of the current work. The relative scarcity of
Lagrangian solutions to difficultCFD problems in the literature, however,
clearly indicates that the field has a long way to go before it is generally
competitive with Eulerian methods.
The degrees of freedom in a Lagrangian representation are used to
define the locations of the Lagrangian nodes as well as the values of the
fluid variables at those locations. This represents an additional expense
relative to Eulerian methods when the grid locations in the latter can all
be specified in one-dimensional tables. When Eulerian grids become either
arbitrarily distorted or fully unstructured, Lagrangian algorithms have no
appreciable memory penalty. Because the grid can move, however, new
types of instability are possible in a Lagrangian representation that cannot
occur in the Eulerian case.
Methods without an underlying node connectivity, denoted "free"
Lagrangian algorithms, are being studied carefully now in the hope of
avoiding the fluctuations associated with the discontinuous process of
reconnecting the Lagrangian nodes as their near neighbors change. Multi­
dimensional algorithms based on a Monotonic Lagrangian Grid to keep
track of near neighbors are being investigated by Fyfe et al. ( 1 987). The
COMPUTATIONAL FLUID DYNAMICS 369
general idea is to use information from all nodes in the neighborhood to
calculate gradients and other fluid-dynamic driving terms without defining
a connectivity that changes discontinuously. The influence of one node on
another is described by a weight that goes continuously to zero as the two
nodes separate. Thus, an infinitesimal change in a node location cannot
lead to a finite change in any term affecting the fluid dynamics at that
node.

4. NOVEL IMPLEMENTATIONS
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

4. 1 Today 's Supercomputer Architectures


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Today's supercomputers such as the Cray, ETA 10, NEC, Fujitsu, and
Hitachi rely on vector registers and pipelines to obtain their speed. In an
arithmetic pipeline the fetching of data from memory, the several stages
of each arithmetic operation, the storage of the computed results back into
memory (or vector registers), the loop-index increment, and the test for
completion of the loop are all carried out simultaneously for the sequential
components of a vector of similar operands. Because vector notation is well
suited to describing these operations logically, the term "vectorization" has
been adopted for the process of optimiZing algorithms to take full advan­
tage of these computers. In most supercomputers, however, the term is a
misnomer, since each element of the vector is really calculated sequentially
using these hardware "pipelines" of memory controllers, routers, and
arithmetic and logical processors. Even the supercomputers with vector
registers do not have separate processors for each element of the vector of
operations to be performed. The vector registers are merely fast, inter­
mediate locations to store the operands and results, ensuring optimal
sequential processing through the pipelines.
Conventional scalar computers are generally rated in millions of instruc­
tions per second, or "MIPS," reflecting their scalar nature. For scientific
computing a more suitable measure has been millions of floating-point
operations per second or "megaflops." Today's fastest supercomputers
perform at levels that can exceed a "gigaflop" (109 floating-point oper­
ations per second). Each flop requires 5 or 6 instructions, so there is a
corresponding factor between MIPS and megaflops. This factor of 5 or 6
in instruction count becomes translated into a factor of 1 0 or 20 in
execution speed because of additional economies possible using the
assumed regularity of the operand data in the supercomputer memory.
Nevertheless, current supercomputers are still operating essentially as
sequential computers.
It is not surprising that these sequentially pipelined computer archi­
tectures have reached a stage of development where speed improvements
370 BORIS

are both difficult and expensive to obtain. The hardware in these computers
is pushing the physical limits on the speed of uniprocessors, yet they are
still too small and too slow to bring solutions of realistic three-dimensional
problems to most fluid dynamicists.
In the last several years, several different supercomputer architectures
have emerged. These machines are categorized technically in several ways:
(a) by the maximum floating-point operation count in megaflops or
gigaflops, (b) by the number of processors available (i.e. the "grain" or
degree of parallelism possible), (c) by the power of the individual pro­
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

cessors, (d) by the communication and memory strategy interconnect­


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

ing the processors, and (e) by the method of program coordination and con­
trol. For this survey we consider five types of architectures for CFD
problems.

I. Supercomputers. These are the top-of-the-line, very fast, general­


purpose machines available today with speeds of a gigaflop or more. They
include the CRAY X-MP, CRAY 2, Fujitsu, Hitachi, NEC, and ETA 1 0
systems. A s described above, they have vector capabilities, implemented
by arithmetic pipelines, and some parallelism. In general, the degree of
parallelism is not very high, generally two to four processors, and is often
called "coarse-grained."
2. Superminicomputers. These have speeds of 10 to 1 00 megaflops and
some modest level of parallel processing, often more closely coupled than
in the true supercomputers. Nevertheless, they are designed in a more or
less standard way. They achieve their market share by being cheaper
and somewhat more cost effective than the true supercomputers. These
superminicomputers include the CDC CYBER, Convex C I and C2, SCS
40, Multiflow, Alliant, and Elexi machines. The performance is comparable
to current technology array processors with much better software and with
interactive operating systems. The parallelism in these systems is also quite
coarse.
3. Highly Parallel Processors. These have operation counts of 50 to
2500 megaflops using more radical architecture designs specifically to
exploit parallelism. They are usually expandable and promise to be com­
petitive with conventional supercomputers in the near future, although
perhaps not as versatile for some time to come. These highly parallel
processors include the TMC connection machine, the BBN butterfly
machine, the Navier-Stokes computer (Princeton), the various hypercube
machines such as the NCUBE and AMATEK systems, and systolic arrays.
The grain of the parallelism for this class of supercomputer is usually
described as "moderate" (indicating of order 1 6 to 1 28 processors) or
COMPUTATIONAL FLUID DYNAMICS 371
"fine" (indicating a high degree of parallelism, with 256 processors or
more).
4. Hybrid Systems. These are combinations of the above. One example
is the GAPS (Graphical and Array-Processing System), assembled in our
laboratory at NRL and described briefly below. These are heterogeneous
assemblies of computing components with overall performance in the
supercomputer range (currently starting around 200 megaflops). It is very
unlikely that any one highly parallel architecture will be efficient on all
classes of CFD problems. Structured and unstructured grids, for example,
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

have very different computational constraints, as do implicit and explicit


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

time-advancement strategies for CFD models. A connection machine, with


approximately 64,000 processors, could be used effectively as an array
processor with a Cray. Fluid-dynamic calculations might be executed
efficiently on the Cray as each node of the connection machine simul­
taneously integrates a set of ordinary differential equations describing
the chemistry in each of the fluid elements in the Cray. A heterogeneous­
element supercomputer system with optically networked components will
become the most flexible system overall in the near future, provided that
the network control software for such distributed computing becomes easy
to use.
5. Special-Purpose Computers. The opportunity now exists, as dis­
cussed further in Section 4.3, to develop special-purpose highly parallel
computational engines, where lack of generality in the end application
allows much of what is implemented in software today to be implemented
in hardware at the chip level. The data communication and control require­
ments are known ahead of time, so generality is not needed. Military
signal-processing systems are the first practical entrants in this class of
parallel-processing supercomputer, but specialized Navier-Stokes com­
puters and dedicated computational flow facilities have been the subject
of discussion and design effort for over a decade.

4.2 Parallel Processing for Computational Fluid Dynamics


Parallel computing, a research topic for about ten years, is today a com­
mercial reality. Hybrid systems interconnecting multiple-array processors,
user-friendly scalar processors, and other peripherals on a high-bandwidth
data-transfer bus can be assembled off the shelf. Hypercube networks have
been expanded from a handful of relatively weak nodes to relatively closely
coupled systems with up to 1 024 nodes, each of which can support a 1 0-
20 megaflop, pipelined, floating-point processor. Even more finely grained
parallel systems show exceptional performance on specialized problems
like image processing and cellular automata (e.g. Hillis 1987). Of the five
372 BORIS

classes of "supercomputers" identified in the previous section, two of these


(supercomputers and superminicomputers) require no further discussion.
In this section and the next, one example each of the last three classes is
discussed where exploitation of parallelism is an intrinsic contributor to
the expected performance.
A major technological change, such as that introduced by parallel pro­
cessing to the CFD community's way of doing business, tends to nullify the
large investment already made in programs, methodologies, and software
utilities that cannot be adapted readily to the new parallel-computing
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

architectures. Highly parallel architectures allow not only vectorization,


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

but also closely coupled and loosely coupled parallel computations. The
availability of such computers makes it necessary to develop represen­
tations of the fluid state that are optimal on these machines and to re­
think how we write algorithms and software for them. Luckily, CFD is
a very "parallelizable" discipline. Thus, significant progress is also
expected on the algorithmic and software side. Indeed, parallel processing
is currently finding its way into a number of applications in CFD.
Nature "solves" fluid problems in a fully parallel manner, so the most
effective CFD representations and algorithms seem to be those that rep­
licate what nature does to a significant degree. Thus new numerical algo­
rithms optimized for parallel processing as well as old ones recast in parallel
form are becoming available for both computational fluid dynamics and
molecular dynamics. Though some of the currently popular methods may
be eclipsed because of the difficulty of constructing efficient parallel
implementations, other models, better suited to the directions of parallel
processing, will take over their role.
Waterman ( 1 988) suggests differentiating parallel systems based on the
way that the computational elements communicate control information
and data with each other. Communication between the processors over a
common data bus represents a natural extension of the usual minicomputer
architecture. The advantages of this communications technology are sim­
plicity and familiarity. The disadvantage is contention for the limited
communication bandwidth. With more than a few processing elements on
the data bus, some have to sit idle waiting for data while other elements
are allowed to communicate with each other, with the host (control)
processor, or with shared memory. This is like having a single phone line
connecting a number of offices.
The Graphical and Array Processing System (GAPS) is an example of
a hybrid supercomputer systcm assembled for computational fluid dynam­
ics, molecular dynamics, and reactive flows. It was constructed from com­
mercially available hardware and features the hierarchical parallelism of
a number of separate array processors with powerful vectorization in the
COMPUTATIONAL FLUID DYNAMICS 373
individual processors (see Clementi et al. 1 985, Boris 1 986b, Boris & Oran
1 988). GAPS is not quite as powerful (fast and capable of multiusers) as
a true supercomputer, but it does support direct graphic interaction with
the evolving simulations. Its configuration is becoming standard for bussed
array-processor systems.
GAPS is an asynchronous, multitasking, distributed-control system con­
sisting of an APTEC 2400 I/O computer with 1 2 megabytes of additional
fast memory and about 3 gigabytes of online disk storage connected to a
VAX 1 1 /780. GAPS contains several major computational components,
including six Numerix MARS 432 array processors (maximum per­
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

formance of 30 megaflops each). The six array processors are programmed


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

in Fortran and have a library of vectorized routines and synchronization


software. Three high-resolution color graphics monitors-a Tectronix
4 1 28 raster system, an Iris 4D vector system, and a Metheus 03600-
are connected to the GAPS. Results from GAPS simulations can be dis­
played as they are calculated using a high-bandwidth graphics package that
selects data from the VAX, from the GAPS array processors, or from the
CRAY.
An important part of multiprocessor research has been to develop
efficient algorithms that can take advantage of an asynchronous multi­
tasking parallel architecture such as the GAPS. The fluid dynamics is
modeled using Flux-Corrected Transport algorithms, designed to take
advantage of more parallelism than the pipeline and vector-register archi­
tectures used in current supercomputers. The Reactive Flow Model
(RFM), currently implemented on the GAPS, can be run on relatively
high-resolution fluid simulations for days or weeks of computer time
because it runs in the background without affecting the normal interactive
uses of the host. When all six array processors are working on the same
problem, performance of about 80% of a Cray XMP processor can be
obtained while loading the Aptec communications buses less than 50%.
The performance of systems like these is limited by the contention
among many increasingly fast processors for the limited communications
bandwidth. More communication pathways is a solution to this problem.
For a modest number of processors, a crossbar switch is possible with a link
connecting each processing element with every other processing element.
There are two problems with this approach. First, the number of links
that connect the processors must scale with the square of the number of
processors. Second, significant communication burdens are now trans­
ferred into each processor, which must now schedule and juggle N com­
munication links simultaneously. The crossbar approach seems better
suited to optical communications than to electronic data transfer, since
photons can pass through each other while electrons cannot. Thus it is not
374 BORIS

suprising that crossbar communications do not seem to underpin the most


powerful of the currently useful parallel processors.
Other interprocessor communications approaches have been designed
to reduce the number oflinks coming out of each processor to a manageable
number. The hypercube architectures (see, for example, Fox & Otto 1 984,
Seitz 1 985, and Gustafson et al. 1988) have 10gzN links to each processor,
one for each bit of the address used to number the processors. One way
of viewing this is to consider the processors laid out in a long linear array.
In a hypercube, each processor is connected to its neighbor, to the next
nearest neighbor, to the processor four away, eight away, and so on in
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

powers of two up to the processor that is NI2 away. Each link connects a
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

processor to another processor whose address, expressed in binary, differs


by only one bit in each of the log2N positions. To communicate data
between any two processors, a data packet moves from one node to
another in a series of "jumps" along the communication links in the
hypercube. Each jump that the data packet takes changes one of the binary
bits of the source-node address in the hypercube to the corresponding bit
of the destination address. Thus five jumps are required if the source and
destination addresses differ in five of the log2N bits needed to number each
of the N processor nodes.
Since each node handles all of the data packets passing through it on
the way to their separate destination addresses, collisions occur between
data packets moving from one arbitrary processor to another. These data­
packet collisions, and the delays required to resolve them, are the price
paid for reducing the number of links from NZ to Nlog2N. This collision
problem is not as significant in CFD as in artificial-intelligence applications
and other forms of general information processing because the physics of
the fluid being simulated lends a natural structure to the information
transfers that can be used to reduce or eliminate data-packet collisions.
Gustafson et al. ( 1 988) report extensive research on the use of a l024-node
hypercube in structural analysis, wave mechanics, and fluid dynamics.
They chose the FCT fluid-dynamic algorithm to solve compressible shear­
flow problems as a test case and were able to reduce the cost of com­
municating data between adjacent processors to below a couple of percent
of the overall processing time. Such efficient communication for parallel
processing will be possible with many CFD algorithms.

4.3 A Hybrid Cellular-Automata Computer


A hybrid cellular automaton is a generalization of the usual cellular­
automata models in which the governing rules are entirely local and can
be applied simultaneously to all the nodes defining a given state of the
COMPUTATIONAL FLUID DYNAMICS 375
system to obtain the next state, but where they act on local data specified
as floating-point numbers. In this sense, an explicit finite-difference or
finite-volume model with a spatially local template and a sufficiently gen­
eral set of update formulas is a hybrid cellular automaton. By accepting
systems with continuous, accurately specified macroscopic fluid quantities
as hybrid cellular automata, the generally unacceptable scaling of statis­
tical fluctuations as N- 1/2 in cellular-automata models, where N is the
number of lattice-gas atoms in a small volume of macroscopic extent, can
be eliminated. Further, relating the "rules" to the desired physics is easier
for a broad range of systems, allowing compressible fluids to be simulated
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

accurately.
Actually, there is no hard . distinction between the classical cellular­
automata models and these hybrid versions, because floating-point num­
bers of finite precision are discrete. Further, more complex cellular-autom­
ata models (for example, in three dimensions) already require many bits
of data at each lattice site to specify the state of the system. The distinctions
can be viewed as a choice of whether to average over the real physical
fluctuations, however represented, before or after one discretizes the prob­
lem for digital computation. As noted in Section 2.2, optimal represen­
tations with a few bits of precision may well be found.
There are numerous examples of CFD algorithms that already satisfy
the criteria to be hybrid cellular-automata models, including some of the
most general and flexible of the monotone convection algorithms. These
algorithms are entirely local, and exactly the same instructions are used
for all the cells. For these models a computer could be built in which a
generic cubical block of 3-D computational cells can be stacked like "Lin­
coln Logs" or "Lego" in any size or configuration. Inital data loadings of
the cells could specify distorted geometries in the sense of spectral elements,
and boundary conditions for each block could be chosen from a menu of
allowed situations. General geometries could be represented by block
structuring, and internal obstacles might be implemented by replacing
certain blocks with null-operation blocks or special-purpose blocks repre­
senting different materials or fluids. Communication of data from one
block to the next could be coupled through the abutting faces of adjacent
blocks. Power could be delivered by conducting rods threaded through the
assembly along one of the coordinate directions. The physical phenomenon
limiting the size of such a special-purpose computer is heat conduction,
since the algorithms and communications for a hybrid cellular-automaton
are all local. As the linear size and hence spatial resolution are doubled,
the heat-generation rate would increase eightfold but the surface area to
radiate it would increase by only a factor of four.
376 BORIS

5. NOVEL APPLICATIONS OF CFD


5.1 Toward Simulations of More Complex Problems
The application of computational fluid dynamics technology is, and always
will be, limited by the speed and size of the available computers. Complex
geometry, discussed in Section 3, causes as many problems as complex
physics. Today CFD methods can simulate flow either in complex
geometry with simple physics or with complex physics in relatively simple
geometry, but they cannot do both. Thus research continues on a broad
front to develop algorithms that are faster, more robust, more flexible,
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

easier to use, and more accurate. As the number of phenomena in a


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

problem increases, the possible number of interactions between them


increases at least quadratically. Each of these possible interactions will
have to be represented and resolved numerically to be simulated accurately.
In practice, every interaction will have its own parameter regime within
which the computational predictions will be valid and outside of which
they will be suspect. Thus the rapidly increasing number of interactions
possible in a complex fluid system is a potential source of weakness.
Major weaknesses recognized in CFD are the detailed representation
and simulation of turbulence and of chemical reactions. Both of these
are particularly important for CFD, since they are intrinsic to the fluid
itself rather than being imposed externally (as boundary conditions or
forcing terms) or occurring only in limited regions (such as at interfaces).
To resolve small-scale turbulence or full chemical kinetic systems in a
multidimensional CFD model imposes unacceptable costs if it can be done
at all. In each case the detailed treatm�nt of the turbulence (or chemistry)
requires many more degrees of freedom than are practical to resolve in the
problem representation. Nevertheless, CFD is perhaps the most used and
useful of the computational disciplines because it fits so naturally the large
gap between what analysis and experiment can accomplish. Being an
intrinsically macroscopic discipline, CFD matches well the requirements
of designing and improving real-world fluid systems.

5.2 Use of Phenomenologies in CFD Simulations


One way to deal with resolution limits and overall simulation cost is
to isolate the expensive processes and to replace them with simplified
phenomenological models. These phenomenologies are physically reason­
able, approximate models with parameters that are guesses, fits to experi­
ments, or calibrated from more detailed but more restricted numerical
models. For example, a realistic chemical-reaction mechanism contains
many chemical species and perhaps hundreds of reaction rates linking
them. Integrating the stiff ordinary differential equations for the evolution
COMPUTATIONAL FLUID DYNAMICS 377
of the individual chemical species and the fluid temperature in a multi­
dimensional CFD code is possible in theory, but it is extremly expensive
in practice. An alternate approach is to use the detailed reaction mechanism
to calculate bulk properties, such as final temperatures and pressures, as
a function of the initial temperatures and pressures. This table of macro­
scopic results forms the basis of a computationally inexpensive phenom­
enology for use in CFD models. In general, the more accurate the under­
lying theoretical framework for these phenomenologies, the smaller the
tables of results needed and the broader the validity of the model.
There is always a tendency to overestimate how much a phenomenology
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

can actually predict. Physical interactions cannot really be predicted when


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

the controlling processes are not resolved accurately in time, in space, or


are not included in the computational model. If one of the controlling
physical or fluid-dynamic processes in a simulation is being treated by a
phenomenological model, the entire simulation is no more accurate than
the phenomenology. This may be true even if the other effects are treated
by more detailed models, because there really is no substitute for resolving
all the space and time scales affecting the problem.
Figure 4 shows schematically what levels of detailed CFD simulations
are practical for complex, multidimensional, turbulent, and reactive
'
systems. Two "dimensions" of difficulty are considered in the figure: the
complex physics of chemical kinetics and local multiphase material effects
along the horizontal axis, and the geometry-related resolution-bound pro­
cesses of fluid dynamics and turbulence on the vertical axis. Each axis
spans a range of difficulty from EASY to HARD.
On the horizontal axis, an easy problem might involve several coupled
chemical reactions among a few species. A hard problem involves hundreds
of chemical reactions among dozens of chemical species. Even more diffi­
cult problems, such as the multiphase properties of dusts, sprays, droplets,
and other heterogeneous fluid-dynamic phenomena, are indicated to the
right of the vertical dashed line. These latter processes and interactions
almost always need a phenomenological representation to be included in
a practical CFD simulation.
Along the vertical axis, an easy fluid-dynamic problem might involve a
one-dimensional shock calculation. A hard problem involves transient
multidimensional fluid dynamics, where both divergence and curl com­
ponents of the flow field interact. Detailed modeling of turbulence is a
very hard fluid-dynamic problem. Generally, phenomenologies are used to
represent the macroscopic fluid effects of turbulence. This situation is
indicated high on the vertical axis, above the dashed horizontal line.
The region indicated CAN DO in the figure includes problems with
either an easy fluid-dynamic component or an easy chemistry component.
378 BORIS

UJ
U
Z
UJ
...J

1
;j
CD
a:
;j
I-
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

o
a:
«
::t:


(J)
«
w

EASY HARD DUSTS. SPRAYS

CHEMISTRY ..

Figure 4 The interaction of phenomenological models and detailed models in computational


fluid dynamic simulations of complex systems.

The computational effort can be concentrated on the harder aspect of the


problem. Today, detailed modeling of complex fluid-dynamic processes is
possible for certain problems with one difficult aspect, provided that the
other aspects are easy enough. The COULD DO region is an extension of
the CAN DO region. If ample computational resources are available and
the co.mputations are directed at answering a few, specific, well-focused
questions, the COULD DO region can be simulated. No new technology
is needed, but 1 0 to 40 hrs of supercomputer time are required, compared
with the I to 4 hrs for a CAN DO calculation.
New supercomputer systems are aimed primarily at the COULD DO
problems. Hundreds of computational hours on today's supercomputers
will be available on newer systems in a few hours of computation, and
these systems in turn could also be run around the clock. This reduction
in running time for what is today a COULD DO problem will eventually
make it a CAN DO problem. There is a tendency for CFD calculations
to migrate toward runs a few hours long on a fairly loaded computer
COMPUTATIONAL FLUID DYNAMICS 379
system. Overnight turnaround of a large complex calculation matches
well the CFD practitioner's ability to absorb and understand what has
happened and to modify his program accordingly. In addition, the scien­
tific value often well-coordinated 4-hr runs usually exceeds the information
gained from one 40-hr run.
The next region is marked ONE PHENOMENOLOGY to indicate that
the overall problem difficulty is so great that full resolution of all the
important physics and fluid dynamics cannot be done simultaneously. At
least one of the major aspects has to be treated phenomenologically_ The
outer region of Figure 4 is labeled ALL PHENOMENOLOGY. Both the
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

fluid dynamics and other physical aspects of the problem are so difficult
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

that detailed solution of either is considered impractical. Thus, interacting


phenomenologies must be constructed. Note that if one phenomenology
is a suspect representation, the interactions between two are at least doubly
suspect. Conservation laws help maintain some aspects of the global va­
lidity of the simulations, but detailed prediction of new problems cannot
be expected. Turbulent reacting flow models fall into this category.
Using local phenomenologies such as equations of state or chemical­
kinetics models has one advantage over fluid-dynamic phenomenologies.
We know, in principle, how to incorporate the former processes into large­
eddy simulations. Though not all of the input chemical rates or specific
heats necessary for particular reactions or species are known, we have
some confidence in the equations and how they fit into and drive the
overall CFD model. In turbulence, the modeler is in a much weaker
position with respect to what is known and how to use that knowledge.
The effects of turbulence are already in the conservation equations, but
various limitations and costs keep one from being able to resolve all the
active space scales in detailed simulations. This problem has been dealt
with by proposing phenomenologies such as the k-s and eddy-diffusivity
models and fitting the constants in them to experiments and other cal­
culations. The difficulty relative to the equation of state and chemistry
phenomenologies is that turbulence really is not a local phenomenon on
the macroscopic scales of interest. All space and time scales are excited in
turbulence, so there is no natural scale at which to distinguish turbulence
from complex but computable large-scale flows.
With the improvements that are becoming available in computer hard­
ware and software, we expect to see some changes in this picture. First, it
will be easier to do large simulations routinely in three dimensions. Second,
it will also be possible to include a larger range of space scales in each
dimension. When several scales of interacting coherent structures can be
resolved conveniently, the details of how to represent the smaller unre­
solved scales will be far less important.
380 BORIS

5.3 Using Artificial-Intelligence Techniques in CFD


The traditionally separate disciplines of scientific computing, signal pro­
cessing, and computer science are growing together and strengthening each
other as each approaches a level of relative maturity and user friendliness.
The current trend toward parallel processing in CFO and the increasing
availability of low-cost pipelined and vector processors are direct conse­
quences of strides made in signal processing. The GAPS system described
earlier and the powerful processors of the Cornell Supercomputer Center
and the IBM parallel-processing system (e.g. Rapaport & Clementi 1 986)
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

are a direct metamorphosis of signal-processing systems into general­


Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

purpose systems for scientific use.


Computer science is making correspondingly large contributions to
computational fluid dynamics. The obvious items are compilers, interactive
operating systems, networks, and file-manipulation facilities. Other con­
tributions are beginning to have major impact. As interactive graphics
workstations become the standard human-machine interface for CFD, the
"human engineering" technologies of computer science are destined to
play an increasing role in speeding the comprehension rate ofthe enormous
quantities of raw data that are generated by a running simulation. As CFD
fi nd s its w ay in creasin gl y into real-time systems (for example, i n predict­
ing the perfomance of a racing sailboat in situ or the developing flow
forces on an airplane trying to land in turbulent conditions), the need for
rapid assimilation of the results of a CFD computation become even more
acute.
Artificial-intelligence (AI) research, long an esoteric componeni of com­
puter science , is also reaching a level of maturity and commercial avail­
ability sufficient to make major contributions to CFD. As problems and
computer systems become more complex, a real need for expert systems
and automatic-programming systems is arising in order to organize and
systematize the development of big simulation models, to guide the analysis
and interpretation of the individual computer runs, and to organize the
enormous data bases that result from storing the simulation results.
Andrews ( 1 988) summarizes as follows: "These first AI/CFD systems
demonstrate that present AI technology can be successfully applied to
well-formulated problems that are solved by means of classification or
selection of preenumcratcd solutions (as opposed to construction, where
solutions must be synthesized). Attempts to incorporate topics that are
still in the realm of AI research or to apply AI technology to poorly
understood or poorly formalized CFD tasks have some benefits but gen­
erally result in long system development times and a large investment of
effort with no guaranteed payoff ."
COMPUTATIONAL FLUID DYNAMICS 381

5.4 Computation in Experimental Fluid Dynamics

Laboratory experiments and field observations have long been the main
approach to the study of fluid-dynamic behavior. Despite the decades of
research, understanding is still increasing rapidly because of recent strides
made in diagnostic, recording, and data-processing technology. The advent
of very localized, highly accurate, and virtually noninvasive diagnostic
techniques, generally made possible by the rapid development of laser and
computer technology in the last two decades, has underpinned these recent
advances.
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

Two distinct approaches to fluid-dynamic experimentation deserve men­


tion: global flow visualization, and localized quantitative measurement.
Flow visualization is still the best way to extract understanding of the
overall dynamics taking place in a fluid system. Lasers are now used to
illuminate selected planes and regions of a flow and to excite specific dyes
and diagnostic chemical reactions. Film and videotape are the generally
used recording media, and direct recording with arrays of charge-coupled
devices (CCDs) or special video-recording equipment is being used increas­
ingly to acquire flow-visualization information directly in computer-read­
able form.
Quantitative analyses of the flow can be accomplished at some difficulty
by digitizing several successive images and using modern image-processing
techniques to identify corresponding locations or structures at several
times in the flow and then to deduce their dynamics from frame-io-frame
changes.
The second class of experimental approaches involves the use of probes
to make in situ quantitative measurements in the fluid. As in the flow
visualizations, lasers and computer technology go hand in hand to measure
local particle velocities by laser-Doppler ve10cimetry (LDV) techniques.
Lasers are also used to measure the local density of trace components of
mixing streams to measure the time history of the mixedness.
Recent advances in computer technology are tending to blur the dis­
tinction between flow visualization and point-probe diagnostics. Flow
visualizations can be digitally encoded and analyzed off-line from the
experiment by a number of techniques designed to extract quantitative
information about specific locations in the flow from successive frames of
the record. Using computers, one can calculate accurately and auto­
matically the location and motion of features such as a passive interface
between two distinct fluid components or a dynamic surface such as a
flame front or a shock. The motion of speCific illuminated markers can be
deduced from successive frames, giving information on both local fluid
velocity and density, assuming that the marker particles or bubbles are
382 BORIS

small enough that buoyancy and inertial effects can be neglected. These
digital records also facilitate the direct comparison of CFD simulations
and the experiments.
Computer technology, through high-speed digitization and the use of
sophisticated graphics, is also allowing local quantitative measurements
of a flow field to be converted to flow visualizations. When data at enough
points in the fluid can be taken and recorded with a high enough repetition
rate, computers can interpolate the flow properties between the measure­
ment points, allowing subsequent reconstruction of "visualizations" of
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

otherwise "invisible" fields like the fluid velocity, vorticity, or flow Mach
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

number. These relatively new laboratory applications of CFD-super­


computer technology to fluid-dynamic research comprise a rapidly grow­
ing component of computational fluid dynamics.

ACKNOWLEDGMENTS

This work was sponsored by the Office of Naval Research through the
Naval Research Laboratory, by the Defense National Agency, and by the
Defense Advanced Research Projects Agency. I would particularly like to
thank Ms. Lorraine Mundo for her help in preparing the manuscript, Dr.
David M osher for his valued suggestions toward improving the manu­
script, Dr. Elaine Oran for her help in and dedication to the task of con­
verting a body of work to a coherent reactive-flow modeling discipline,
and Drs. Timothy Coffey and Alan Berman for their support of CFD at
NRL throughout the years. Finally, I would like to acknowledge my
colleagues, many of whom are cited above, who have helped to leave a
lasting mark on how fluid-dynamic research is done through their numer­
ous creative contributions to computational fluid dynamics.

Literature Cited

Anderson, D. A., Tannehill, J. C., Pletcher, bors" algorithm of order N using a mono­
R. H. 1 984. Computational Fluid Me­ tonic logical grid. J. Comput. Phys. 66:
chanics and Heat Transfer. New York: 1-20
Hemisph ere/McGraw-H ill Boris, J. P. 198 6b . Supercomputing at the
Andrews, A. E. 1988. Progress and chal­ U.S. Naval Research Laboratory. In Opti­
lenges in the application of artificial intel­ cal and Hybrid Computing, SPIE Vol. No.
ligence to computational fluid dynamics. 634, pp. 7-23. SPIE Inst. Adv. Techno!.
A IA A J. 26(1): 40-46 Boris, 1. P., Boo k , D. L. 1 976. Solutio n of
Book, D. L., ed, 1 98 1 . Finite-Difference the continuity equati on by the method of
Techniques for VeclOrized Fluid Dynamics flux-corrected transport. In Methods in
Calculations. New York: Springer-Verlag. Computational Physics, ed. B. Alder, S .
226 pp Fembach, M. R o tenberg, 16: 85-129.
Boris, J. P. 1 97 1 . A fluid transport algorithm New York: Academic
that works. In Computing as a Language Boris, 1. P., Oran, E. S. 1 988. The numerical
of Physics, pp. 171-189. Vienna: Int. At. simulation of compressible and reactive
Energy Agency turbulent structures. Proc. Joint US/Fr.
Boris, J. P. 1986a. A vectorized "near neigh- Workship Turbul. React. Flows, ROURn,
COMPUTATIONAL FLUID DYNAMICS 383
Fr., ed. B. G. Murthy, R. Borghi. Berlin: methods for numerical computation of
Springer-Verlag. In press discontinuous solutions of the equations
Canuto, C., Hussaini, M . Y . , Quarteroni, A., of fluid dynamics. Mat. Sb. 47: 271-
Zang, T. A. 1988. Spectral Methods in 306
Fluid Dynamics. New York: Springer-Ver­ Gottlieb, D., Orszag, S. A. 1977. Numerical
lag. 557 pp. Analysis of Spectral Methods: Theory and
Chen, H., Matthaeus, W. H. 1987a. Cellular Applications. Philadelphia: SIAM
automaton formulation of passive scalar Greenspan, D. 1985. Computer studies in
dynamics. Phys. Fluids 30: 1 235-37 particle modelling of fluid phenomena.
Chen, H . , Matthaeus, W. H. 1987b. New Math. Modelling 6: 273-94
cellular automaton model for magneto­ Gustafson, J. L., Montry, G. R., B enner, R.
hydrodynamics. Phys. Rev. Lett. 58: 1 845 E. 1 9 8 8 . Development of parallel methods
(Erratum in Phys. Rev. Lett. 59: 1 55) for a 1 024-processor hypercube. SIA M J.
Clementi, E., Corongiu, G., Detrich, J. H . Sci. Stat. Comput. 9(4): 1-32
1985. Parallelism i n computations i n Harlow, F. H. 1964. The particle-in-cell com­
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

quantum and statistical mechanics. Com­ puting method for fluid dynamics. In
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

put. Phys. Commun. 37: 287-94 Methods in Computational Physics, ed. B .


Colella, P., Woodward, P. 1984. The piece­ Alder, S. Fernbach, M . Rotenberg, 3: 3 19-
wise parabolic method (PPM) for gas­ 43. New York: Academic
dynamic simulations. J. Comput. Phys. Harten, A. 1983. High resolution schemes
54(1): 1 74-20 1 for hyperb olic conservation laws. J. Com­
Courant, R., Friedrichs, K. 0., Lewy, H . put. Phys. 49: 357-93
1928. Ober die partiellen DitIerenzen­ Hasslacher, B. 1987. Discrete fluids. Los
gleichungen der mathematischen Physik. Alamos Sci: Spec. Iss. 1 5 : 175-2 1 7
Math. Ann. 100: 32-74 Hillis, W. D. 1987. The connection machine.
Crowley, W. P. 1 97 1 . FLAG: A free­ Sci. Am. 256(6): 108-15
Lagrange method for numerically simu­ Hockney, R. W., Eastwood, J. W. 198 1 .
lating hydrodynamic flows in two dimen­ Computer Simulation Using Particles, Chap.
sions. Proc. Int. Con[. Numer. Methods in 8, pp 267-309. New York: McGraw­
Fluid DYIl., 2nd, ed. M. Holt, pp. 327--43. Hill
New York: Springer-Verlag Hofstadter, D. 1979. Godel, Escher and Bach:
Dahlburg, J., Montgomery, D., Doolen, an Eternal Golden Braid. New York: Basic
G. 1987. Noise and compressibility in Books
lattice gas fluids. Phys. Rev. A 36: 2471- Jameson, A., Baker, T. J., Weatherhill, N.
74 P. 1986. Calculation of inviscid transonic
Djomehri, M. J., Doss, S. K., Gelinas, R. flow over a complete aircraft. AlAA Pap.
J., M iller , R. J. 1988. Applications of the No. 86-0103
moving finite clement method for systems Karniadakis, G. E., Bu\lister, E. T., Patera,
in 2-D. J. Comput. Phys. In press A. T. 1986. A spectral element method for
Fox, G. C., Otto, S. W. 1984. Algorithms for solution of the two- and three-dimensional
concurrent processors. Phys. Today 37(5): time-dependent incompressible Navier­
50-59 Stokes equations. Proc. Eur. US Conf.
Frisch, U., Hasslacher, B., Pomeau, Y. 1986. Finite Elem. Methodsfor Nonlinear Probl. ,
Lattice gas automata for the Navier­ eds. P. Bergan, K. J. Bathe, pp. 803- 1 7 .
Stokes equation. Phys. Rev. Lett. 56: Berlin: Springer-Verlag
1 505 8 Kim, J., Moin, P., Moser, R. 1987. Tur­
Fritts, M., Boris, J. P. 1979. The Lagrangian bulence statistics in fully developed flow
solution of transient problems in hydro­ at low Reynolds number. J. Fluid Mech.
dynamics using a triangular mesh. J. Com­ 177: 1 3 3--66
put. Phys. 3 1 (2): 173-2 1 5 Korczak, K. Z., Patera, A. T. 1986. Iso­
Fritts, M . J., Crowley, W. P., Trease, H . , eds. parametric spectral element method for
1985. The Free-Lagrange Method. Lect. solution of the Navier-Stokes equations in
Notes Phys., Vol . 238. Berlin: Springer­ complex geometry. J. Comput. Phys. 62:
Verlag 361-82
Fyfe, D., Oran, E. S., Boris, J. P. 1987. Fluid Lambrakos, S, G., Boris, J. P. 1987 Geo­
dynamics on a monotonic Lagrangian metric properties of the monotonic La­
grid. Bull. Am. Phys. Soc. 32(10): 2063 grangian grid algorithm for near neigh­
(Abstr.) bors calculations. J. Comput. Phys. 73(1):
Gelinas, R. J., Doss, S. K., Miller, K. 1 98 1 . 1 83-202
The moving finite element method: appli­ Lambrakos, S. G., Boris, J. P., Oran E. S.,
cations to general partial ditIerential equa­ Chandrasekhar, I., Nagumo , M. 1988. A
tions with multiple large gradients. J. constraint algorithm for maintaining rigid
Comput. Phys. 40: 202-49 bonds in molecular dynamics simulations
Godunov, S. K. 1959. Finite ditIerence of large molecules. J. Comput. Phys. In
384 BORIS

press. Also 1988. US Nav. Res. Lab. Memo. Orszag, S. A. 1 980. Spectral methods for
Rep. 6174 problems in complex geometries. J.
Lax, P. D., Wendroff, B. 1 960. Systems of Comput. Phys. 37: 70-92
conservation laws. Commun. Pure Appl. Ortega, J. M . , Voigt, R. G. 1 985. Solution
Math. 1 3 : 2 1 7-37 of partial differential equati ons on vector
Lax, P. D. , Wendroff, B. 1 964. Difference and parallel computers . lCASE Rep. 85-
schemes for hyperbolic equations with 1, l nst . Comput. App!. Sci. Eng. NASA
high order accuracy. Commun. Pure Appl. Langley Res. Cent., Hampton, Va.
Math. 1 7: 381-98 Patera, A. T. 1 984. A spectral element
Lohner, R. 1987. An adaptive finite element method for fluid dynamics: laminar flow
scheme for transient problems in CFD. in a channel expansion. J. Comput. Phys.
Compu!. Meth. Appl. Mech. Eng. 61: 323- 54: 468-88
38 Potter, D. 1 973. Computational Physics. New
Lohner, R., Morgan, K., Zienkiewicz, O. C. York: Wiley
Rapaport, D. c., Clementi, E. 1 986. Eddy
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org

1985. An adaptive finite element proced­


ure for compressible high speed flows. Com­ formation in obstructed fluid flow: a
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

put. Meth. Appl. Mech. Eng. 5 1 : 441--65 molecular dynamics study. Phys. Rev.
Lohner, R., Morgan, K., Vahdati, M . , Boris, Lett. 6(57): 695-98
J. P., Book, D. L. 1 988. FEM-FCT: com­ Roache, P. J. 1 982. Computational Fluid
bining unstructured grids with high res­ Dynamics. Albuquerque, NM: Hermosa
oluti on . Commun. Appl. Numer. Methods. Rodrigue, G., ed. 1 982. Parallel Computa­
In press tions. New York: Academic. 403 pp.
MacCormack, R. W. 1 97 1 . Numerical solu­ Seitz, C. L. 1 985. The cosmic cube. Commun.
tion of the interaction of a shock wave ACM 28: 22-33
with a laminar b oundary layer. Proc. Int. Shimomura, T., Doolen, G. D., Hasslacher,
Con! Numer. Methods in Fluid Dyn., 2nd, B., Fu, C. 1987. Calculations using lattice
ed. M . Holt, pp. 1 5 1--63. New York: gas techniques. Los Alamos Sci: Spec. Iss.
Springer-Verlag 1 5 : 201-10
McDonald, B. E. 1988. Flux-corrected Spalart, P. R. 1 986. Numerical study of sink­
pseudospectral method for scalar hyper­ flow boundary layers. J. Fluid Mech . 172:
bolic conservation laws. J. Comput. Phys. 307-28
In press Spalart, P. R. 1 988. Direct simulation of a
McDonald, B. E., Ambrosiano, J., Zalesak, turbulent boundary layer up to Re =
S. 1 985. The pseudo spectral flux cor­ 1 4 1 0. J. Fluid Mech. 1 87: 61-98
rection (PSF) method for scalar hyper­ Sweby, P. K . 1 984. High resolution schemes
bolic problems. Proc. Int. Assoc.for Math. using flux limiters for hyperbolic con­
and Comput. in Simulation World Congr., servation laws. SIAM J. Numer. Anal. 2 1 :
11th, ed. R. Vichnevetsky, pp. 67-70. New 995-1 0 1 1
Brunswick, NJ: Rutgers Univ. Press Symon, K . R., Marshall, D. , Li, K . W. 1 970.
Metcalfe, R. W., Orszag, S. A., Brachet, M . Bit-pushing and distribution pushing
E., Menon, S., Riley, J. J. 1987. Secondary techniques for the solution of the Vlasov
instability of a temp orally growing mixing equati on . Proc. Con! Numer. Simulation
laye r. J. Fluid Meeh. 1 84: 207--44 of Plasmas, 4th, pp. 68-125. Washington,
Miller, K., Miller, R. 1 98 1 . Moving finite DC: US Govt. Print. Off. [No. 085 1 00059]
elements, part I. SIAM J. Numer. Anal. Trease, H. 1985. Three-dimensional free
1 8: 1 0 1 9-32 Lagrangian hydrodynamics. See Fritts et
Montgomery, D., Doolen, G. D. 1 987. Two a!. 1 985, pp. 145--5 7
cellular automata for plasma computa­ van Leer, B. 1979. Towards the ultimate con­
tions. Complex Syst . 1 (4): 83 1-38 servative difference scheme. Proc. Int.
Morton, K. W., Baines, M. J., eds. 1 986. Con! Numer. Methods in Fluid Dyn. Lect.
Numerical Methods for Fluid Dynamics. Notes Phys., ed. E. Kraus, 170: 507-12.
Oxford: Clarendon New York: Springer-Verlag
Moser, R. D., Moin , P. 1987. The effects of von Neumann, 1., Richtmye r , R. D. 1950. A
curvature in wall-bounded turbulent method for the numerical calculation of
flows. J. Fluid Mech. 1 7 5 : 479-5 1 0 hydrodynamic shocks. J. Appl. Phys . 2 1 :
National Academy o f Sciences/National 232-57
Research CounciL 1986. Current Capa­ Waterman, P. J. 1988. Parallel processing
bilities and Future Directions in Com­ finds a place. Defense Comput. I: 24-28
putational Fluid Dynamics . Washington, Winkler K.-H. A., Chalmers, J. W . , Hodson,
DC: NatL Acad. Press S. W., Woodward, P., Zabusky, N. J.
Oran, E. S., Boris, J. P. 1987. Numerical 1987. A numerical laboratory. Phys. Today
Simulation of Reactive Flow. New York: 40(10): 28-37
Elsevier. 601 pp. Wolfram, S. 1985. Undecidability and in-
COMPUTATIONAL FLUID DYNAMICS 385
tractability in theoretical physics. Phys. 3 1 : 335-62
Rev. Lett. 54(8): 735-38 Zalesak, S. 1 98 1 . Very h igh order and
Woodward, P., Colella, P. 1984. The numeri­ pseudo spectral flux-corrected transport
cal simulation of two-dimensional fluid algorithms for conservation laws. In
flow with strong shocks. J. Comput. Phys. Advances in Computer Methods/or Partial
54(1): 1 1 5-73 Differential Equations, ed. R. Vichne­
Yakhot, V., Orszag, S. A. 1986. Reynolds vetsky, R. S. Stepleman, pp. 1 2�34. New
number scaling of cellular automaton Brunswick, NJ: Int. Assoc. Math. Com­
hydrodynamics. Phys. Rev. Lett. 56: 1691- put. Simulation
93 Zienkiewicz, O. c., Morgan, K. 1983. Finite
Zalesak, S. 1979. Fully multidimensional Elements and Approximation. New York:
flux-corrected transport. J. Comput. Phys. Wiley
Annu. Rev. Fluid Mech. 1989.21:345-385. Downloaded from www.annualreviews.org
Access provided by Sun Yat-Sen University on 12/12/21. For personal use only.

You might also like