Computational Aerodynamics for Aircraft Design

Antony Jameson

Abstract
This article outlines some of the principal issues in the development of numerical methods for the pre-
diction of ows over aircraft and their use in the design process. These include the choice of an appropriate
mathematical model, the design of shock-capturing algorithms, the treatment of complex geometric cong-
urations, and shape modications to optimize the aerodynamic performance.
While computational methods for simulating uid ow have by now penetrated a broad variety of elds, including
ship design, car design, studies of oil recovery, oceanography, meteorology, and astrophysics, they have assumed
a dominant role in aeronautical science. In the aircraft industry there is often a very narrow margin between
success and failure. In the past two decades the development of new commercial aircraft successful enough to
make a prot for the manufacturer has proved an elusive goal. The economics of aircraft operation are such that
even a small improvement in eciency can translate into substantial savings in operational costs. Therefore, the
operating eciency of an airplane is a major consideration for potential buyers. This provides manufacturers
with a compelling incentive to design more ecient aircraft.
One route toward this goal is more precise aerodynamic design with the aid of computational simulation.
In particular it is possible to attempt predictions in the transonic ow regime that is dominated by nonlinear
eects, exemplied by the formation of shock waves. The importance of the transonic regime stems from the
fact that to a rst approximation, cruising eciency is proportional to ML/D, where M is the Mach number
(speed divided by the speed of sound), L is the lift, and D is the drag. As long as the speed is well below the
speed of sound, the lift-to-drag ratio does not vary much with speed, so it pays to increase the speed until the
eects of compressibility start to cause a radical change in the ow. This occurs when embedded pockets of
supersonic ow appear, generally terminating in shock waves. A typical transonic ow pattern over a wing is
illustrated in Fig. 1. As the Mach number is increased the shock waves become strong enough to cause a sharp
increase in drag, and nally the pressure rise through the shock waves becomes so large that the boundary layer
separates. The most ecient cruising speed is usually in the transonic regime just at the onset of drag rise, and
the prediction of aerodynamic properties in steady transonic ow has therefore been a key challenge.
Prior to 1965 computational methods were hardly used in aerodynamic analysis, although they were widely
used for structural analysis. There was already in place a rather comprehensive mathematical formulation of
uid mechanics. This had been developed by elegant mathematical analysis, frequently guided by brilliant
insights. Well-known examples include the airfoil theory of Kutta and Joukowski, Prandtls wing and boundary
layer theories, von Karmans analysis of the vortex street, and more recently Joness slender wing theory (1), and
Hayess theory of linearized supersonic ow (2). These methods, however, required simplifying assumptions of
various kinds, and could not be used to make quantitative predictions of complex ows dominated by nonlinear
eects. The primary tool for the development of aerodynamic congurations was the wind tunnel. Shapes were
tested and modications selected in the light of pressure and force measurements together with ow visualization
techniques. In much the same way that Brunelleschi could design the dome of the Florence cathedral through

Reprinted from Science, Volume 245, pp. 361-371, 1989

The author is McDonnell, Professor of Aerospace Engineering, Princeton, NJ 08544

1
a good physical understanding of load paths, so could experienced aerodynamicists arrive at ecient shapes
through testing guided by good physical insight. Notable examples of the power of this method include the
achievement of the Wright brothers in leaving the ground (after rst building a wind tunnel), and more recently
Whitcombs discovery of the area rule for transonic ow (3). In fact, in the 80 years since the Wright brothers
ight, every conceivable conguration has been tried, and by a process of natural selection airplanes have very
rapidly evolved to the rather ecient forms we see ying today.
Experimental design is an expensive process, however. More than 20,000 hours of wind tunnel testing were
expended in the development of some modern designs such as the General Dynamics F 111 or the Boeing 747.
The computer opens up new possibilities for attacking these problems by direct calculation of solutions to more
complete mathematical models. The requirements to be met by an eective method include: (i) capability to
simulate the main features of the ow, such as shock waves and vortex sheets; (ii) prediction of viscous eects;
(iii) ability to handle geometrically complex congurations; and (iv) eciency in both computational and human
eort.
Figure 1: Pattern of transonic ow past an airfoil. The sonic line is where M = 1, that is, where the speed
equals the speed of sound
This is a formidable list that poses a severe challenge to the numerical analyst. Aside from the need to
treat nonlinear eects, the problem is also compounded by the fact that the steady ow equations are of
mixed type, elliptic in the subsonic regions of the ow, and hyperbolic in the supersonic regions. Moreover,
they are to be solved in the unbounded domain exterior to the aircraft. The problem is somewhat alleviated,
however, by the fact that airplanes y by achieving controlled and steady ows. In some ways it is easier to
predict the ow past an airplane than the ow past a nonspinning baseball. The trajectory of a knuckle ball is
unpredictable because vortices may be shed at any angle to the axis of ight. Experimental measurements of
the drag coecient of a sphere show drastic variations with speed associated with changes in the wake pattern
(4). There is a sudden reduction in the drag coecient from around 0.4 to around 0.09 when the ow becomes
turbulent and the ow separates further aft. Accurate prediction of the drag coecient would thus require
prediction of the change in the separation point as the ow becomes turbulent. The ow past an aircraft is
generally attached in normal operating conditions, with the exception of certain high-speed aircraft which use
controlled separation o sharp side edges to generate vortices which pass above the wing, thus enhancing the
lift. Otherwise the onset of separation typically marks the limits of the ight envelope, such as the stalling
point. Consequently, depending on the intended application, useful aerodynamic simulations can be achieved
with substantially simplied mathematical models.
This article outlines some of the key issues in the development of algorithms for uid ow simulation, and
their use for aircraft design. The volume of work in computational uid dynamics precludes a comprehensive
review in an article of this length, and the choice of topics must necessarily be personal. Some representative
examples of ow calculations have been drawn from the work of the author and his associates.
2
= 0
p
n

___
t
= 0

___
Viscousity = 0
Linearize

No
streamwisw
viscous terms
Vorticity = 0 Density = Const. Density = Const.
Compressible flow
Boundary
layer eqs.
NS eqs.
Thinlayer
eqs.
NavierStokes
Unsteady viscous
Euler eqs. Laplace eq.
Parabolized
NS eqs.
Perturb. eq.
PradtlGlauert
Potential eq.
Small Transonic small
perturb. eq.
Figure 2: Equations of uid dynamics for mathematical models of varying complexity. (Supplied by Luis
Miranda, Lockheed Corporation)
Mathematical Models of Fluid Flow
The Navier-Stokes equations state the laws of conservation of mass, momentum, and energy for the ow of a gas
in thermodynamic equilibrium. In the Cartesian tensor notation, let x
i
, be the coordinates, p, , T, and E the
pressure, density, temperature, and total energy, and u, the velocity components. Each conservation equation
has the form
w
t
+
F
j
x
j
(1)
For the mass equation
w = , F
j
= u
j
(2)
For the i momentum equation
w
i
= u
i
, F
ij
= u
i
u
j
+p
ij

ij
(3)
where
ij
is the visous stress tensor, which is proportional to the rate of strain tensor and the bulk dilatation.
If and are the coecients of viscosity and bulk viscosity then

ij
=
_
u
i
x
j
+
u
j
x
i
_
+
ij
_
u
k
x
k
_
(4)
Usually = 2/3. For the energy equation
w = E, F
j
= (E +p)u
j

jk
u
k

T
x
j
(5)
where is the coecient of heat conduction. The pressure is related to the density and energy by equation of
state
p = ( 1)
_
E
1
2
u
i
u
i
_
(6)
3
Linear
1 Mflops 100 Gflops
CRAY XMP
100 Mflops
CONVEX
10 Mflops
CDC 6600
2D Airfoil
Inviscid
flow
potential
Nonlinear
Stokes
Navier
Euler
3D Wing
Aircraft
aeraged
Reynolds
Figure 3: Complexity of the problems that can be treated with dierent calsses of computer (1 op=1 oating-
point operation per second; 1Mop = 10
6
ops, 1 Gop=10
9
ops)
in which is the ratio of specic heats.
An indication of the relative magnitude of the internal and viscous terms is given by the Reynolds number
R
e
=
UL

(7)
where U is a characteristic velocity and L a representative length. The viscosity of air is very small, and typical
Reynolds numbers for the ow past a component of an aircraft such as a wing are of the order of 10
7
or more,
depending on the size and speed of the aircraft. In this situation the viscous eects are essentially conned
to thin boundary layers covering the surface. Boundary layers may nevertheless have a global impact on the
ow by causing separation. Unfortunately, unless they are controlled by active means such as suction through
a porous surface, boundary layers are unstable and generally become turbulent.
The computational requirements for the full simulation of all scales of turbulence have been estimated as
growing proportionally to R
9/4
e
(5), and are clearly beyond the reach of current computers. Turbulent ows
may be simulated by the Reynolds equations, in which statistical averages are taken of rapidly uctuating
components. Denoting uctuating parts by primes and averaging by an overbar, this leads to the appearance of
Reynolds stress terms of the form u

i
u

j
which cannot be determined from the mean values of the velocity and
density. Estimates of these additional terms must be provided by a turbulence model. The simplest turbulence
models augment the molecular viscosity by an eddy viscosity that crudely represents the eects of turbulent
mixing, and is estimated with some characteristic length scale such as the boundary layer thickness. A rather
more elaborate class of models introduces two additional equations for the turbulent kinetic energy and the
rate of dissipation. Existing turbulence models are adequate for particular classes of ow for which empirical
correlations are available, but they are generally not capable of reliably predicting more complex phenomena,
such as shock wave-boundary layer interaction. Outside the boundary layer excellent predictions can be made
by treating the ow as inviscid. Setting
ij
= 0 and eliminating heat conduction from Eqs. 2, 3, and 5 yields the
inviscid Euler equations, which are a very useful model for predicting ows over aircraft. According to Kelvins
theorem, a smooth inviscid ow that is initially irrotational remains irrotational. This allows one to introduce
a velocity potential 4 such that u
i
= /x
i
. The Euler equations for a steady ow now reduce to

x
i
_

x
i
_
= 0 (8)
4
This may be expanded in quasilinear form as
c
2

x
2
i
u
i
u
j

x
i
x
j
= 0 (9)
where c is the speed of sound, which is determined by the relation c
2
= p/
If the ow is locally aligned, say, with the x
1
axis, Eq. 9 reads as
(1 M
2
)

x
2
i
+

2

x
2
2
+

2

x
2
3
(10)
where M is the Mach number u
1
/c. The change from an elliptic to a hyperbolic partial dierential equation
as the ow becomes supersonic is evident. According to Croccos theorem (6), the vorticity g in a steady ow
is related to entropy production through the relation u x + TS = 0, where S is the entropy. Thus the
introduction of a potential is consistent with the assumption of isentropic ow. If M

is the free stream Mach

number, then p =

/M
2

and = (M

c
2
)
1/(1)
The potential ow equation cannot exactly model shock waves, through which entropy is produced. Weak
solution admitting isentropic jumps that conserve mass but not momentum are a good approximation to shock
waves, however, as long as the shock waves are quite weak (with a Mach number 1.3 for the normal velocity
component upstream of the shock wave). Stronger shock waves tend to separate the ow, with the result that
the inviscid approximation is no longer adequate. Thus this model well balanced, and it has proved extremely
useful for the prediction of the cruising performance of transport aircraft. An estimate of the pressure drag
arising from shock waves is obtained because of the momentum decit through an isentropic jump.
If one assumes small disturbances about a free stream in the x
1
direction, and a Mach number close to unity,
Eq. 10 can be reduced to the transonic small disturbance equation in which M
2
is estimated as
M
2

[1 ( + 1)/x
1
]
The nal level of approximation is to linearize Eq. 10 by replacing M
2
by its free stream value M

. In
the subsonic case the resulting Prandtl-Glauert equation can be reduced to Laplaces equation by scaling the x,
coordinate by (1 M
2

)
1/2
. Ideal incompressible ow satises the Laplaces equation, as can be seen by setting
= constant in Eq. 8.
The hierarchy of mathematical models is illustrated in Fig. 2. With limits on the available computing power,
and the cost of the calculations, one has to make a trade-o between the complexity of the mathematical model
and the complexity of the geometric conguration to be treated. The rst major success in computational
aerodynamics was the development of boundary integral methods (7, 8), which enabled the linearized potential
ow equation to be solved routinely for arbitrary congurations. In the 1970s eective methods were devised
for the treatment of the transonic potential ow equation (9-13). These were instrumental, for example, in the
development of rened wing designs for aircraft such as the Canadair Challenger and the Boeing 757, and they
are still widely used today. More recently there have been intensive eorts to develop good algorithms for the
Euler and Navier-Stokes equations.
The computational requirements for aerodynamic simulation are a function of the number of operations
required per mesh point, the number of cycles or time steps needed to reach a solution, and the number of mesh
points needed to resolve the important features of the ow. Algorithms for the three-dimensional transonic
potential ow equation require about 500 oating point operations per mesh point per cycle. The number of
operations required for an Euler simulation is in the range of 1000 to 5000 per time step, depending on the
complexity of the algorithm. The number of mesh intervals required to provide an accurate representation of
a two-dimensional inviscid transonic ow is of the order of 160 wrapping around the prole, and 32 normal to
the airfoil. Correspondingly, about 100,000 mesh cells are sucient to provide adequate resolution of three-
dimensional inviscid transonic ow past a swept wing, and this number needs to be increased to provide a good
simulation of a more complex conguration such as a complete aircraft. The requirements for viscous simulations
by means of turbulence models are much more severe. Good resolution of a turbulent boundary layer needs
5
about 32 intervals inside the boundary layer, with the result that a typical mesh for a two-dimensional Navier-
Stokes calculation contains 512 intervals wrapping around the prole, and 64 intervals in the normal direction. A
corresponding mesh for a swept wing would have, say, 512 X 64 X 256 = 8,388,608 cells, leading to a calculation
at the outer limits of current computing capabilities. Figure 3 gives an indication of the boundaries of the
complexity of problems with can be treated with dierent levels of computing power. The vertical axis indicates
the geometric complexity, and the horizontal axis the equation complexity.
Algorithms for Flow Simulation
The case of the inviscid Euler equations is used here to illustrate a general approach to the design of algorithms
for aerodynamic simulation (14). The underlying idea is simply to integrate the timedependent equations of
uid ow until the solution evolves to a steady state. This may be accomplished by dividing the domain of the
ow into a large number of small subdomains and applying the conservations laws in integral form

t
_

wd +
_

F.dS = 0 (11)
Here F is the vector appearing in Eq. 1, and dS is the directed surface element of the boundary d of the domain
. The use of the integral form has the advantage that no assumption of the dierentiability of the solutions is
implied, with the result that it remains a valid statement for a domain containing a shock wave. In general the
subdomains could be arbitrary, but it is convenient to use either hexahedral or tetrahedral cells. Hexahedral
cells are naturally generated by the use of body conforming meshes corresponding to curvilinear coordinate
systems. Where the geometric complexity of the conguration is such that it becomes dicult to generate a
structured mesh of this kind, one may prefer to resort to an unstructured decomposition into tetrahedral cells.
Alternative discretization schemes may be derived by storing sample values of the ow variables at either the
cell centers or the cell vertices.
With a tetrahedral mesh, each face is a common external boundary to exactly two control volumes. Therefore
each internal face can be associated with a set of ve mesh points consisting of its three corners 1, 2, and 3, and
the vertices 4 and 5 of the two tetrahedra based on the face. Vertices 4 and 5 are the centers of the two control
volumes inuenced by the face. It is now possible to generate the discrete approximation by presetting the ux
balance at each mesh point to zero, and then performing a single loop over the faces. For each face one rst
calculates the uxes of mass, momentum and energy across the face, and then one assigns these contributions
to the vertices 4 and 5 with positive and negative signs, respectively. Because every contribution is transferred
from one control volume into another, all quantities are perfectly conserved. Mesh points on the inner and outer
boundaries lie on the surface of their own control volumes, and the accumulation of the ux balance in these
volumes has to be correspondingly modied.
An alternative route to the discrete equations if provided by the Galerkin Method. Multiplying Eq. 1 by a
test function , and integrating by parts over space, one obtains the weak form

t
_ _ _

wd =
_ _ _

F.d
_ _

F.dS (12)
that is also valid in the presence of discontinuities in the ow. By choosing test functions with local support,
separate equations are obtained for each node. For example, if is piecewise linear, with a nonzero value only
at a single node, and one also assumes that ux vector to be piecewise linear, one obtains a ux balance for the
node that is equivalent to the use of trapezoidal integration in the integral conservation law.
These procedures lead to nondissipative approximations to the Euler equations. Dissipative terms may be
needed for two reasons. First there is the possibility of frozen oscillatory modes. The second reason is to
allow the clean capture of shock waves and contact discontinuities without undesirable oscillations. An extreme
overshoot could resultogn a negative value of an inherently positive quantity such as the pressure or density.
6
Consider a general semidiscrete scheme of the form
d
dt
v
j
=

k
c
jk
(v
k
v
j
) (13)
A maximum cannot increase and a minimum cannot decrease if the coecients c
jk
are nonnegative. Positivity
conditions of this type lead to diagonally dominant schemes, and they are a key to the elimination of unwanted
oscillations. They may be realized by the introduction of dissipative terms or by the use of upwind biasing in the
discrete scheme. Unfortunately they may also lead to severe restrictions on the accuracy unless the coecients
have a complex nonlinear dependence on the solution.
Following the pioneering work of Godunov (15), a variety of dissipative and upwind schemes designed to
have good shock capturing properties have been developed during the past decade (16-25). The one-dimensional
scalar conservation law
du
dt
+

x
f(u) = 0 (14)
provides a useful method for the analysis of these schemes. The total variation
_

u
x

dx
of a solution of this equation does not increase, provided that any discontinuity appearing in the solution satises
an entropy condition (24). Harten proposed that dierence schemes ought to be designed so that the total
variation cannot increase (19). General conditions on coecients which result in total variation diminishing
(TVD) for schemes of this kind were stated and proved by Jameson and Lax (27), and also by Osher and
Chakravarthy (28). In the case of a three-point scheme they are equivalent to the positivity conditions stated
above.
A conservative semidiscrete approximation to the one dimensional conservation law can be derived by sub-
dividing the line into cells. Then the evolution of the value vj in the jth cell is given by
x
d
dt
v
j
+h
j+1/2
h
j1/2
= 0 (15)
where h
j+1/2
is an estimate of the ux between cells j and j+1. The simplest estimate is the average (f
j+1
+f
j
)/2
but this leads to a scheme that does not satisfy the positivity conditions. To correct this one may add a dissipative
term and set
h
j+1/2
=
1
2
(f
j+1
+f
j
) +
j+1/2
(v
j
v
j1
) (16)
In order to estimate the required value of the coecient
j+1/2
, let a
j+1/2
be a numerical estimate of the wave
speed f/u,
a
j+1/2
=
_
_
_
fj+1fj
vj+1vj
if v
j+1
= v
j
f
v

v = v
j
if v
j+1
= v
j
(17)
Now
h
j+1/2
= f
j
+ (f
j+1
f
j
)/2
j+1/2
(v
j+1
v
j
)
= f
j
(
j+1/2

1
2
a
j+1/2
)(v
j+1
v
j
)
(18)
and similarly
h
j1/2
= f
j
(
j1/2
+
1
2
a
j1/2
)(v
j
v
j1
) (19)
Thus the positivity conditions are satised if

j+1/2

1
2

a
j+1/2

7
for all j. The minumum sucient value of just one half the wave speed produces the upwind scheme
h
j+1/2
=
_
f
j
if a
j+1/2
0
f
j+1
if a
j+1/2
< 0
(20)
It may be noted that the successful treatment of transonic potential ow also involved the use of upwind
biasing. This was rst introduced by Murman and Cole to treat the transonic small disturbance equation
(9). The authors rotated dierence scheme (10), which extended their technique to treat the general transonic
potential ow equation, proved to be very robust. TVD schemes can yield sharp discrete shock waves without
oscillations, but in this simple form they are at best rst-order accurate. Schemes that are second-order accurate
almost everywhere can be devised with the aid of ux limiters. These are used to limit the magnitude of higher
order antidiusive or corrective terms depending on the ratio of the dierences
j+1/2
= v
j+1
v
j
, and
j1/2
in
adjacent cells. This technique may be traced to the work of Boris and Book (16), and it was also independently
advanced by van Leer (17).
In order to apply these ideas to a system of equations one may split the ux into components corresponding
to the dierent wave speeds. A convenient way to do this was proposed by Roe (19). Another promising
approach is to reduce the multidimensional Euler equations directly to a diagonal form (29, 30). The use of ux
splitting allows the precise matching of the dissipative terms to introduce the minimum amount of dissipation
needed to prevent oscillations, but it is computationally expensive. In practice the use of a blend of low- and
high-order dissipative terms with adaptive coecients conditioned to the maximum local wave speed can yield
eective shock capturing schemes. Recently Engquist, Lotstedt, and Sjogreen have shown that oscillations can
be eectively controlled by post-processing the result after each time step with a lter which is applied only
if a new extremum is detected (31). These procedures lead to a set of coupled ordinary dierential equations,
which can be written in the form
dw
dt
+R(w) = 0 (21)
where w is the vector of the ow variables at the mesh points, and R(w) is the vector of the residuals,
consisting of the ux balances dened by the space discretization scheme, together with the added dissipative
terms. These are to be integrated to a steady state. If the objective is simply to reach the steady state and
details of the transient solution are immaterial, the time-stepping scheme may be designed solely to maximize
the rate of convergence. The rst decision is whether to use an explicit scheme, in which the space derivatives
are calculated from known values of the ow variables at the beginning of the time step, or an implicit scheme,
in which the formulas for the space derivatives include as yet unknown values of the ow variables at the end
of the time step, leading to the need to solve coupled equations for the new values. The permissable time step
of an explicit scheme is limited by the Courant-Friedrichs-Lewy (CFL) condition, which states that a dierence
scheme cannot be a convergent and stable approximation unless its domain of dependence contains the domain
of dependence of the corresponding dierential equation (32).
One can anticipate that implicit schemes will yield convergence in a smaller number of time steps, because
the time step is no longer constrained by the CFL condition. This will be ecient, however, only if the decrease
in the number of time steps outweighs the increase in the computational eort per time step consequent upon
the need to solve coupled equations. The prototype implicit scheme can be formulated by estimating w/t at
t +t as a linear combination of R(w
n
) and R(w
n+1
). The resulting equation
w
n+1
= w
n
t
_
(1 )R(w
n
) +R(w
n+1
)
_
(22)
can be linearized as
_
I +t
R
w
_
w+ tR(w
n
) = 0 (23)
This reduces to the Newton iteration if one sets = 1 and lets t . In a three dimensional case with an
N x N x N mesh its bandwidth is of order N
2
. Direct inversion requires a number of operations proportional
to the number of unknowns multiplied by the square of the bandwidth, that is, of the order of N
7
. This is
prohibitive, and forces recourse to either an approximate factorization method or an iterative solution method.
8
Figure 4: Multigrid W cycle for managing the grid calculation. Symbols: E, calculate the change in ow for
one time step, , transfer the data without updating the solution
9
Alternating direction methods, which introduce factors corresponding to each coordinate, are widely used
for structured rectangular meshes (33, 34), but they cannot be implemented on unstructured tetrahedral meshes
that do not contain identiable mesh directions. If one chooses to adopt the iterative solution technique, the
principal alternatives are variants of the Gauss-Seidel and Jacobi methods. A symmetric Gauss-Seidel method
with one iteration per time step is essentially equivalent to an approximate lower-upper (LU) factorization of the
implicit scheme (35-37). On the other hand the Jacobi method with a xed number of iterations per time step
reduces to a multistage explicit scheme, belonging to the general class of Runge-Kutta schemes (38). Schemes
of this type have proved very eective for wide variety of problems, and they have the advantage that there can
be applied equally easily on both structured and unstructured meshes (39-42).
Let w
n
be the result after n steps. The general form of an m stage scheme is
w
(0)
= w
n
w
(1)
= w
(0)

1
tR
(0)
. . .
w
(m1)
= w
(0)

m1
tR
(m2)
w
(m)
= w
(0)
tR
(m1)
w
(n+1)
= w
(m)
(24)
In cases where only the steady state solution is needed, it is helpful to separate the residual R(w) into its
convective and dissipative parts Q(w) and D(w). Then the residual in the (q+1)st stage is evaluated as
R
(q)
=
4

r=0
_

qr
Q(w
(r)
)
qr
D(w
(r)
)
_
(25)
where
4

r=0

qr
= 1,
4

r=0

qr
= 1
Blended multistage schemes of this type, which have been analyzed elsewhere (42), can be tailored to give
large stability intervals along both the imaginary and negative real axes.
Radical further improvements in the rate of convergence to a steady state can be realized by the multigrid
time-stepping technique. The concept of acceleration by the introduction of multiple grids was rst proposed by
Federenko (43). There is by now a fairly well-developed theory of multigrid methods for elliptic equations (44,
45), based on the concept of the updating scheme acting as a smoothing operator on each grid. This theory does
not hold for hyperbolic systems. Nevertheless, it seems that it ought to be possible to accelerate the evolution
of a hyperbolic system to a steady state by using large time steps on coarse grids so that disturbances will be
more rapidly expelled through the outer boundary. Several multigrid time-stepping schemes designed to take
advantage of this eect have been proposed (46-50).
One can devise a multigrid scheme using a sequence of independently generated coarser meshes that are not
associated with each other in any way. In the case of a structured mesh it is convenient, however, to generate
the coarser meshes by eliminating alternate points in each coordinate direction. In order to give a precise
description of the multigrid scheme subscripts may be used to indicate the grid. Several transfer operations
need to be dened. First the solution vector on grid k must be initialized as
w
(0)
k
= T
k,k1
w
k1
(26)
where w
k1
is the current value on grid k 1, and T
k,k1
is a transfer operator. Next it is necessary to transfer
a residual forcing function such that the solution on grid k is driven by the residuals calculated on grid k 1.
This can be accomplished by setting
P
k
= Q
k,k1
R
k1
(w
k1
) R
k
(w
(0)
k
) (27)
10
whre Q
k,k1
is another transfer operator. Then R
k
(w
k
) is replaced by R
k
(w
k
)+P
k
in the time stepping scheme.
Thus, the multistage scheme is reformulated as
w
(0)
k
= w
(0)
k

1
t
k
[R
(0)
k
+P
k
]
. . .
w
(q+1)
k
= w
(0)
k

q+1
t
k
[R
(q)
k
+P
k
]
(28)
The result w
(m)
m
then provides the initial data for the grid k+1. Finally, the accumulated correction on grid k has
to be transferred back to detailed analysis of multigrid time-stepping schemes is provided elsewhere (50). With
properly optimized coecients multistage time-stepping schemes can be very ecient drivers of the multigrid
process. A W cycle of the type illustrated in Fig. 4 proves to be a particularly eective strategy managing the
work split between the meshes. In a three-dimensional case the number of cells is reduced by a factor of eight
on each coarser grid. On examination of the gure, it can therefore be seen that the work measured in units
corresponding to a step on the ne grid is of the order of
1 + 2/8 + 4/64 +. . . < 4/3 (29)
Examples of Flow Simulations
Some examples are presented here to illustrate the kind of results that are attainable with these methods. The
calculations can be performed on readily accessible equipment, including the mini-supercomputers that are now
becoming widely available.
The rst example, displayed in Fig. 5, shows the result of a transonic ow calculation for a swept wing.
In this case my program FL067 was used to solve the Euler equations on a structured mesh with 192 intervals
wrapped around the wing, 32 intervals normal to the wing, and 48 intervals in the spanwise direction, for a total
of 294,912 cells. The gure shows the pressure distribution on the upper and lower surfaces measured by the
pressure coecient C
p
= (p p

)/(q
2

/2), with negative C

p
, toward the top, following the usual convention
in aeronautical science. The triangular shock pattern is clearly visible. This calculation, requiring the solution
of coupled nonlinear equations for 1,474,560 unknowns, reached an essentially steady state in 15 multigrid
cycles, with the lift converged within 1 part in 1,000. Because the algorithm is explicit it permits concurrent
calculations at separate mesh points, so it is fully amenable both to vector processing and to massively parallel
computation. The cover shows the vortical ow generated by a delta wing at a high angle of attack. The
ow separates o the sharp side edges and rolls up to form vortices which pass over the top of the wing. The
associated low pressure is an important mechanism for enhancing the lift of a high-speed aircraft. In the gure
the vortices are illustrated by tracking particle paths. (This ow calculation was also performed with FL067,
while the mesh generation and the graphic visualization was the work of G. Volpe and M. Siclari, using facilities
at the Grumman Corporate Research Center.)
Figure 6 gives an indication of the present state of the art for viscous ow simulation. It shows a comparison
between the predicted pressure distribution over the RAE2822 and experimental data obtained in a wind tunnel.
These results were obtained by Martinelli, using a new turbulence model based on renormalization group theory
(51). Figure 7 shows a comparison between inviscid and viscous ow simulations for the planned Hermes space
shuttle in hypersonic ow at Mach 8, and an angle of attack of 30 degrees. While the calculation demonstrates
the capability of the LU implicit algorithm (37), the temperatures reached in the ow are high enough that
a more complex model ought to be introduced to allow for real gas eects, such as dissociation and chemical
reactions.
If computational methods are to be really useful to airplane designers, they must be able to treat extremely
complex congurations. A major pacing item of the eort to attain this goal has been the problem of mesh
generation. For simple wing body combinations it is possible to generate rectilinear meshes without too much
diculty (52): for more complicated congurations containing, for example, pylon-mounted engines, it becomes
increasingly dicult to produce a structured mesh, which is aligned with all solid surfaces. A popular mesh
11
Figure 5: Flow over a design fot the Hermes Space Shuttle at M=8 and 30 degrees angle of attack. Comparison
of Mach number distributions for (A) inviscid and (B) viscous ow. (Bottom) The color contours represent the
local Mach number, with balck showing the free stream (M=8), yellow-red the range M=3 to 0. (Supplied by
H. Rieger, Dornier, GmbH.)
12
generation procedure, pioneered by Thompson (53), is to generate grid surfaces as solutions of elliptic equations.
Hyperbolic marching methods have also proved successful in some applications (54). These methods are typically
combined with domain decomposition to produce multiblock meshes in which the mesh inside each block is
separately generated. The simulation by Sawada and Takanasha of the ow over a four-engined shorttake o
aircraft provides a striking example of what can be achieved by these methods (55).
An alternative procedure is to use tetrahedral cells in an unstructured mesh which can be adapted to conform
to the complex surface of an aircraft. Such an approach was successfully implemented for the potential ow
equation by Bristeau et al. (13) whose simulation of transonic ow past a complete aircraft was a major
advance. Because an arbitrary set of points admits a triangulation, the problem can be simplied by separating
the procedure for generating mesh points from the procedure for triangulating them. A cluster of mesh points
surrounding the aircraft can be created in any convenient manner. One may take, for example, the union of the
points belonging to separately generated meshes around each component. The swarm of mesh points is then
connected together to form tetrahedral cells, which provide the basis for a single niteelement approximation
for the entire domain. The triangulation of a set of points to form disjoint tetrahedra is, in general, non-unique:
one procedure is to generate the Delaunay triangulation (56, 57). This is dual to the Voronoi diagram that
results from a division of the domain into polyhedral neighborhoods, each consisting of the subdomain of points
nearer to a given mesh point than any other mesh point.
Combined with a nite-element method based on the concepts outlined in this article, this approach to mesh
generation yields an eective method to perform ow calculations for complete aircraft (58, 59). Figure 8 shows
a result for the Boeing 747. The color contours denote the surface pressure distribution. The calculation includes
internal ow through the engine ducts, but as yet no eort has been made to simulate propulsive eects. Using
our Convex C2 mini-supercomputer at Princeton University, we nd that calculations with, say, 125,000 to
150,000 nodes and 750,000 to 900,000 tetrahedrons require 15 to 18 hours. We have also been fortunate enough
to have been given access to a Cray 2 by Cray Research, which allows us to perform the same calculations in 3
to 4 hours. They also require a lot of memory, typically about 60 million words. Other eorts to develop nite
element methods to treat supersonic and hypersonic ows over complex congurations are under way (60, 61).
The Design Problem
In considering the objectives of computational aerodynamics, three levels of desirable performance can be
identied: (i) capability to predict the ow past an airplane or its important components in dierent ight
regimes such as take-o or cruise, and o design conditions such as utter; (ii) interactive calculations to allow
rapid improvement of the design; and (iii) integration of the predictive capability into an automatic design
method that incorporates computer optimization.
lthough the results presented in this article demonstrate that substantial progress has been made toward
the rst objective, various problems of viscous separated ows still remain beyond our reach. Also in relatively
simple cases, such as an airfoil or wing in inviscid ow, calculations can be performed fast enough that the
second objective is attainable. The third objective must eventually prevail. What the designer really needs is a
method of determining shapes that will have the desired aerodynamic properties. The ability to predict the ow
over a given shape is not good enough for this purpose, as it does not provide any guidance on how to change
the shape if it is unsatisfactory. It has been recognized that an experienced designer generally has an idea of
the kind of pressure distribution that will lead to favorable characteristics. For example, he can avoid adverse
pressure gradients that will induce premature separation of the boundary layer. Thus, in addition to the direct
problem of calculating the pressure distribution over a given shape, the inverse problem of nding the shape
that will yield a specied pressure distribution can also play an important role. The problem of designing a two-
dimensional prole to attain a desired pressure distribution was solved by Lighthill for the case of incompressible
ow by means of conformal mapping (62). Subsequently there has been continuing interest in the problem, and
a variety of methods have been proposed for the solution of the inverse problem in compressible ow (63-66).
One source of diculty is that the desired pressure distribution is not necessarily attainable, unless it satises
certain constraints, with the result that the problem needs to be very carefully formulated.
13
Figure 6: Flow bast a Boeing 747-200 at Mach number 0.84 and 2.73 degrees angle of attack. (Top) The color
contours represent the surface pressure, with red indicating low pressure in the supersonic zone. (Bottom)
Another view showing streamlines around aircraft. (Calculated by A. Jameson and T. J. Baker.)
14
The hodograph transformation oers an alternative approach to the design of airfoils to produce shock-free
transonic ows. Garabe dian and Korn achieved a striking success by using the method of complex characteristics
to solve the equations in the hodograph plane (67). There have been several studies that have explored the
possibility of meeting desired design objectives by using constrained optimization (68, 69). The conguration
is specied by a set of parameters, and any suitable computer program for ow analysis is used to evaluate
the aerodynamic characteristics. In principle this method allows the designer to specify any reasonable design
objectives. The method becomes extremely expensive, however, as the number of parameters is increased. In a
recent paper I suggested that there are benets in regarding the design problem as a control problem in which
the control is the shape of the boundary (70). A variety of alternative formulations of the design problem can
then be treated systematically by using the mathematical theory for control of systems governed by partial
dierential equations (71).
I have used this approach to accomplish the automatic redesign of airfoils to improve their performance for
the case of transonic potential ow. The method proceeds as follows: A two-dimensional prole in the z plane
is generated by conformal mapping from a unit circle in the plane. Since the transformation is dened by an
analytic function, it is fully determined by the value of its modulus h = |dz/d| on the boundary. Thus we can
take f = log h as the control. Suppose now that we dene a cost function, such as
I =
1
2

1
_
C
(p p
d
)
2
d +
2
CD +
1
2

3
(CL CL
d
)
2
(30)
where the integral is taken over the boundary C, is the angle in the circle plane, p is the surface pressure,
p
d
is the desired surface pressure, CD is the drag coecient, CL is the lift coecient, CL
d
is the desired lift
coecient, and
1
,
2
and
3
are parameters to be chosen by the user. Note that in inviscid ow, minimization
of the drag without additional constraints would yield a at plate at zero angle of attack. Suppose that a
perturbation analysis yields an estimate of the variation BI induced by a charge in the mapping function f of
the form
I =
_
C
gfd (31)
where g is independent of f. Then g can be identied as the gradient or derivative of I with respect to f. If we
adjust f by a change
f = g (32)
where is any suciently small positive variable, then
I =
_
C
g
2
d (33)
and the cost must decrease unless the gradient g is already zero, indicating that a stationary point has been
attained. At rst sight it might appear that one must calculate the change in the solution for each possible
change f. Actually this is not the case. The change in the potential that results from a change f in the
mapping function satises an equation of the form
L = J (34)
where L is the perturbation operator, which is self-adjoint, and J depends on f. This is a constraint, and
we can augment the variation I by the integral over the domain of this equation multiplied by a Lagrange
multiplier ,
_ _
D
(L J)dS (35)
Let now be chosen to satisfy the adjoint equation
L = 0 (36)
with appropriately chosen boundary conditions. Then, an integration by parts, all terms explicitly depending
on can be eliminated. In this way it is possible to search for the mapping function f that minimizes I by
15
Figure 7: Redesign of the RAE 2822 airfoil by means of control theory to reduce its shock-induced pressure
drag. (A) Initial prole. Drag coecient is 0.0175. (B)Redesigned prole after ve cycles. Drag coecient is
0.0018
an iterative process, in which the ow equation and an adjoint equation of roughly equal complexity must be
solved at each iteration.
Figure 9 illustrates a result (previously unpublished) obtained by this method. It shows the automatic
redesign of a rather well-known airfoil, the RAE 2822, to reduce its shock-induced pressure drag at a Mach
number of 0.730, while the lift is held roughly constant. The gure shows the initial airfoil, and the modied
airfoil after ve design cycles. The corresponding distribution of the pressure coecient is also shown. The
dierence between the upper and lower surface pressures indicates the distribution of lift. It can be seen that
during the ve cycles the drag coecient is reduced from an initial value of 0.0175, which would be unacceptably
large, to a nal value of 0.0018, a reduction by a factor of 10. Each design cycle takes about 9 s on the Convex
C2, and the complete calculation requires less than 2 min. It appears entirely feasible to extend this method to
the optimization of wings in three-dimensional ow.
Future Possibilities and Challenges
We are now at a point where a variety of ecient algorithms for the solution of the Euler and Navier-Stokes
equations have been developed, and the principles underlying their construction are quite well understood.
16
Progress in viscous ow simulations (72-75) is paced by availability of sucient computing power, and the need
for more reliable turbulence models. Eorts are under way to treat more complex ows. In particular the
simulation of internal ows has lagged behind that of external ows. Devices such as compressors and turbines
require the treatment of unsteady ows with strong viscous eects. The unsteady ow induced by a helicopter
in forward ight is another example of a very complex ow which extends present methods of simulation to the
limit of their capabilities. The accurate prediction of hypersonic ow requires the introduction of much more
complex models to allow for dissociation and chemical reactions at high temperatures.
There remain many opportunities for improvement. Procedures for automatic renement of the mesh in
regions requiring better resolution are likely to be widely adopted. Such methods have been developed for
both rectangular and triangular meshes (60, 61, 7680). The data structures of nite element methods that use
unstructured meshes provide a particularly natural framework for adaptive mesh generation adoption as the
calculation proceeds. The ideal method would also be capable of providing both a solution and an accompanying
estimate of bounds on the error introduced by discretization. The advent of massively parallel computers will
force a reappraisal of the trade-os in the design of algorithms in favor of schemes allowing concurrent calculation.
As the simulations increase in complexity the need for sophisticated pre- and post-processing procedures becomes
more pressing, and it is becoming increasingly necessary to integrate computer graphics software with numerical
simulation methods. Further development of optimization methods should nally allow ow simulations to be
exploited to their full potential for improving aerodynamic design.
References
[1] R. T. Jones, Properties of Low Aspect Ratio Pointed Wings at Speeds Below and Above the Speed of
Sound, Report TR835 (National Advisory Committee for Aeronautics, Washington, DC, 1946).
[2] W. D. Hayes, Linearized Supersonic Flow, Report AL-222 (North American Aviation, Inc., Los Angeles,
CA, 1947).
[3] R. T. Whitcomb, A Study of the Zero L( Drag-Rise Characteristics of Wing-Body Combinations Near the
Speed of Sound, Report TR 1273 (National Advisory Commit tee on Aeronautics, Washington, DC, 1956).
[4] H. Schlichting, Boundary Layer Theory (McGraw-Hill, New York, 1968), p. 17.
[5] C. Canuto, M. Y. Hussaini, A. Quarteroni, D. A. Zang, Spectral Methods in Fluid Dynamics (Springer-
Verlag, New York, 1987).
[6] H. W. Liepmann and A. Roshko, Elements of Gas Dynamics (Wiley, New York, 1957), p. 193.
[7] J. L. Hess and A. M. O. Smith, Calculation of Non-Ding Potential Flow About Arbitrary Three-Dimensional
Bodies, Report ES 40622 (Douglas Aircraft Corp., Long Beach, CA, 1962).
[8] P. E. Rubbert and G. R. Saaris, A General Three-Dimensional Potential Flow Method Applied to V/STOL
Aerodynamics, Technical Paper 680304 (Society of Automotive Engineers, Warrendale, PA, 1962).
[9] E. M. Murman and J. D. Cole, AIAA J. 9,114 (1971).
[10] A. Jameson, Commun. Pure. Appl. Math. 27, 283 (1974).
[11] and D. A. Caughey, Numerical Calculation of the Transonic Flow Past a Swept Wing, Report COO-3077-
140 (New York University, New York, 1977).
[12] A. Jameson, in Proceedings of the IMA Conference on Numerical Methods in Applied Fluid Dynamics
(Academic Press, New York, 1980).
[13] M. O. Bristeau et al., Computer Methods Appl. Mech. Engineering 51, 363 (1985).
17
[14] A. Jameson, Commun. Pure. Appl. Math. 41, 507 (1988).
[15] S. K. Godunov, Mat. Sbornik 47, 271 (1959) [translated as JPRS 7225 (U.S. Department of Commerce,
Washington, DC, 1960)].
[16] J. P. Boris and D. L. Book, J. Comp. Phys. 11, 38 (1973).
[17] B. van Leer, ibid. 14, 361 (1974).
[18] J. L. Steger and R. F. Warming, ibid. 40, 263 (1981).
[19] P. L. Roe, ibid. 43, 357 (1981).
[20] S. Osher and F. Solomon, Math. Comp. 38, 339 (1982).
[21] A. Harten, J. Comp. Phys. 49, 357 (1983).
[22] S. Osher and S. Chakravarthy, SIAM J. Num. Analysis 21, 955 (1984).
[23] P. K. Sweby, ibid., p. 995.
[24] B. K. Anderson, J. L. Thomas, B. van Leer, paper 85-0122 presented at the AIAA 23rd Aerospace Sciences
Meeting, Reno, January 1984.
[25] A. Jameson, in Large Scale Computations in Fluid Mechanics (Lectures in Applied Mathematics), B. E.
Engquist, S. Osher, R. C. J. Sommerville, Eds. (American Mathematical Society, Providence, 1985), vol.
22, p. 345.
[26] P. D. Lax, Regional Series on Applied Mathematics (Society for Industrial and Applied Mathematics,
Philadelphia, 1973), vol. 11.
[27] A. Jameson and P. D. Lax, Conditions for the Construction of Multi-Point Total Variation Diminishing
Dierences Schemes, MAE Report 1650 (Princeton University, Princeton, NJ, 1984).
[28] S. Osher and S. Chakravarthy, Very High Order Accurate TVD Schemes, ICASE Report 88-44 (Institute
for Computer Applications in Science and Engineering, City?, 1984).
[29] C. Hirsch, C. Lacol, H. Decominch, paper 87-1163 presented at the AIAA 8th Computational Fluid Dy-
namics Conference, Hawaii, June 1987.
[30] K. G. Powell and B. van Leer, paper 89-0095 presented at the AIAA 27th Aerospace Sciences Meeting,
Reno, January 1989.
[31] B. Engquist, P. Lotstedt, B. Sjogreen, Math. Comp. 52, 509 (1989).
[32] E. Isaacson and H. B. Keller, Analysis of Numerical Methods (Wiley, New York, 1966), p. 489.
[33] R. W. Beam and R. F. Warming, J. Comp. Phys. 23, 87 (1976).
[34] T. H. Pulliam and J. L. Steger, paper 35-0360 presented at the AIAA 23rd Aerospace Sciences Meeting,
Reno, January 1985.
[35] A. Jameson and E. Turkel, Math. Comp. 37, 385 (1981).
[36] S. Yoon and A. Jameson, paper 87-0600 presented at the AIAA 25th Aerospace Sciences Meeting, Reno,
January 1987.
[37] H. Rieger and A. Jameson, paper 88-0619 presented at the AIAA 26th Aerospace Sciences Meeting, Reno,
January 1988.
18
[38] R. Chipman and A. Jameson, paper 79-1555 presented at the AIAA 12th Fluid and Plasma Dynamics
Conference, Williamsburg, VA, July 1979.
[39] A. Jameson, in Proceedings of the IMA Conference on Numerical Methods in Aeronautical Fluid Dynamics
(Academic Press, New York, 1982), p. 29.
[40] A. Jameson, W. Schmidt, E. Turkel, paper 81-1259 presented at the AIAA 14th Fluid and Plasma Dynamics
Conference, Palo Alto, CA, June 1981.
[41] A. Rizzi and L. E. Eriksson, J. Fluid Mech. 148, 45 (1984).
[42] A. Jameson, in Numerical Methods in Fluid Dynamics, F. Brezzi, Ed. (Springer Verlag, New York, 1985),
p. 156.
[43] R. P. Federenko, USSR Comp. Math. Math. Phys. 4, 227 (1964).
[44] A. Brandt, Math. Comp. 31, 333 (1977).
[45] W. Hackbusch, Computing 20, 291 (1978).
[46] R. H. Ni, AIAAJ. 20,1565 (1982).
[47] A. Jameson, Appl. Math. Comput. 13, 327 (1983).
[48] M. G. Hall, paper presented at IMA Conference on Numerical Methods for Fluid Dynamics, Reading, U.K.,
April 1985.
[49] P. W. Hernker and S. P. Spekreijse, paper presented at the Oberwolfach Meeting on Multigrid Methods,
December 1984.
[50] A. Jameson, in Multigrid Methods II (Lecture Notes in Mathematics Series), W.Hackbusch and U. Trot-
tenberg, Eds. (Springer-Verlag, New York, 1986), p. 166.
[51] L. Martinelli and V. Yakhot, paper 89-1950 presented at the AIAA 9th Computational Fluid Dynamics
Conference, Bualo, June 1989.
[52] T. J. Baker, Appl. Numerical Math. 2, 515 (1986).
[53] J. F. Thompson, F. C. Thames, C. W. Mastin, J. Comp. Phys. 15, 299 (1974).
[54] J. L. Steger and D. S. Chaussee, SIAMJ. Sci. Statistical Computing 1, 431 (1980).
[55] K. Sawada and S. Takanashi, paper 86-0103 presented at the AIAA 24th Aerospace Sciences Meeting,
Reno, January 1986.
[56] B. Delaunay, Bull. Acad. Sci. USSR VII: Class Scil. Mat. Nat. 793 (1934).
[57] G. Voronoi, J. Reine Angew. Math. 134, 198 (1908).
[58] A. Jameson, T. J. Baker, N. Weatherill, paper 86-0103 presented at the AIAA 24th Aerosoace Sciences
Meetine, Reno, January 1986.
[59] A. Jameson and T. J. Baker, paper 87-0452 presented at the AIAA 25th Aerospace Sciences Meeting, Reno,
January 1987.
[60] B. Stoufet, J. Periaux, F. Fezoui, A. Dervieux, paper 87-0560 presented at the AIAA 25th Aerospace
Sciences Meeting, Reno, January 1987.
19
[61] O. Hassan, K. Margan, J. Peraire, paper 89-0363 presented at the AIAA 27th Aerospace Sciences Meeting,
Reno, January 1989.
[62] M. J. Lighthill, A New Method of Two Dimensional Aerodynamic Design, ARC Rand Report M 2112
(Aeronautical Research Council, London, 1945).
[63] G. B. McFadden, An Articial Viscosity Method for the Design of Supercritical Airfoils, Report COO-
3077-158 (New York University, New York, 1979).
[64] G. Volpe and R. E. Melnik, Int. J. Numer. Methods Eng. 22, 341 (1986).
[65] P. A. Henne, paper 80-0330 presented at the AIAA 18th Aerospace Sciences Meeting, Pasadena, January
1980.
[66] M. Giles, M. Drela, W. T. Thompkins, paper 85-1530 presented at the AIAA 7th Computational Fluid
Dynamics Conference, Cincinnati, 1985.
[67] P. R. Garabedian and D. G. Korn, in Proceedings of the SYNSPADE 1970, B. Hubbard, Ed. (Academic
Press, New York, 1971), pp. 253-271.
[68] R. M. Hicks and P. A. Henne, paper 79-0080 presented at the AIAA 17th Aerospace Sciences Meeting, New
Orleans, January 1979.
[69] G. B. Constentino and T. L. Holst, Numerical Optimization Design of Advanced Transonic Wing Cong-
urations, Report TM 85950 (National Aeronautics and Space Administration, Washington, DC, 1984).
[70] A. Jameson, ICASE Report 88-64 [J. Sci. Comput., vol. 3 (1988)].
[71] J. L. Lions, Optimal Control of Systems Governed by Partial Dierential Equations (Springer-Verlag, New
York, 1971).
[72] R. W. MacCormack, paper 85-0032 presented at the AIAA 23rd Aerospace Sciences Meeting, Reno, January
1985.
[73] T. J. Coakely, paper 87-0416 presented at the AIAA 25th Aerospace Sciences Meeting, Reno, January 1987.
[74] T. H. Pulliam, in Numerical Techniques for Viscous Flow Calculations in Turbomachinery Bladings (Von
Karman Institute for Fluid Dynamics, Rhode St. Genetic, Belgium, 1986).
[75] M. Berger and A. Jameson, AIAAJ. 23, 561 (1985).
[76] J. F. Dannenhofer and J. R. Baron, paper 86-0495 presented at the AIAA 24th Aerospace Sciences Meeting,
Reno, January 1986.
[77] R. Lohner, K. Morgan, J. Peraire, paper 86-0499 presented at the AIAA 24th Aerospace Sciences Meeting,
Reno, January 1986.
[78] D. G. Holmes and S. A. Lamson, Proceedings of the International Conference on Numerical Grid Generation
in Computational Fluid Dynamics, Landshut, July 1986.
[79] J. T. Oden et al., Proceedings of the ASME Winter Meeting, Computational Mechanics Advances and
Trends, Anaheim, CA, December 1986.
[80] D. Mavriplis and A. Jameson, in Multigrid Methods, Theory, Applications and Supercomputing, S. Mac-
Cormick, Ed. (Dekker, New York, 1988).
20