·'

(J5'/!. P Ii 5';)
ll/-a G/- 7/rl

AIAA-94-0061 C9()I/t;ff/'n ]) t;··

Recent Progress Toward a
Three-Dimensional Unstructured
Navier-Stokes Flow Solver
N. T. Frink
NASA Langley Research Center
Hampton, VA

32nd Aerospace Sciences Meeting

January 10-13, 1994/Reno, Nevada
. -
For permission to copy or republish, contact the American Institute of Aeronautics and Astronautics
370 L'Enfant Promenade. S.W.• WashinQton, D.C. 20024
 AIAA 94-0061
Recent Progress Toward a Three-Dimensional
Unstructured Navier-Stokes Flow Solver

Neal T. Frinkt
NASA-Langley Research Center

Abstract may soon produce the much needed capabilitylO-12.

Progress has been made toward solving three- Pirzadeh12 recently demonstrated the generation of a
dimensional viscous flow problems with unstructured 3-D tetrahedral viscous grid on a multielement wing by
grids. Enhancements have been added to an estab- the Advancing-Layer Method. Once this technology is
lished unstructured grid flow solver that upgrade its matured, it will only be a matter of time before viscous
formal accuracy to second order, improve the code ef- unstructured grid solutions become routine.
ficiency and speed, and accurately solve the Navier- This paper reports on recent enhancements made
Stokes equations on tetrahedral cells. Accuracy is in- to the unstructured grid flow solver described in Refs.
creased by a pseudo-Laplacian weighted averaging al- 13 to 15 toward solving viscous-flow problems, as well
gorithm which produces more robust convergence, and as to provide a general upgrade of the base methodol-
permits higher-order boundary conditions to be used, ogy. The accuracy and utility of the base Euler code
which is important on highly-stretched cells. An ex- has been well documented in Refs. 16 to 21. The
isting time-implicit Gauss-Seidel algorithm is installed subject enhancements were made to address four is-
which results in a six-fold reduction in total computer sues, 1) to improve some peculiar convergence prop-
time over that of an explicit algorithm for inviscid-flow erties encountered periodically through extensive ap-
calculations. Finally, an unstructured laminar solution plication of the inviscid method to complex configu-
on the ONERA M6 wing has been validated against a rations, which may potentially inhibit convergence of
structured-grid solution with encouraging results. viscous-flow solutions, 2) to increase the formal accu-
racy of the method, and 3) to improve code efficiency
and speedup the overall convergence rate, and 4) to
Introduction develop an accurate algorithm for solving the viscous
Unstructured grid methods provide a viable alter- shear fluxes for the Navier-Stokes equations on tetra-
native to the structured grid approach for solving the hedral cells. Toward these goals, a new weighted av-
Euler equations for flows over complex aerodynamics eraging scheme has been implemented which upgrades
shapes. With current grid generation technology, 1-3 the algorithm for reconstructing cell-averaged data to
isotropic tetrahedral grids can now be constructed the cell vertices to full second-order accuracy from one
around complex aircraft configurations in a matter that is slightly less accurate. The increased accuracy of
of days. The construction of a comparable block- the new scheme permits higher-order boundary condi-
structured grid generally requires much more time. As tions to be employed. A face-colored data structure has
a result, unstructured grid methodology has become also been implemented resulting in significant improve-
widely used throughout industry and government for ments in code efficiency over the prior cell-based data
rapidly solving complex aerodynamic problems which structure. The implicit time integration scheme from
are amenable to inviscid assumptions. Ref. 22 has been installed in the code to increase the
While many practical problems of interest can be support within the computational domain during time
addressed with inviscid methodology, the inclusion of stepping. The Navier-Stokes terms have been added
viscous effects is critical for the correct analysis of and validated against structured grid results for lami-
many complex configurations. The numerical solu- nar attached flow over an ONERA M6 wing.
tion of the Navier-Stokes equations on tetrahedral ele-
ments presents a formidable challenge. However, con-
siderable progress is being made toward achieving such Governing Equations
solutions,4-9 particularly on two-dimensional triangu- The fluid motion is governed by the time depen-
lar elements. The primary obstacle to progress in 3D dent Navier-Stokes. equations for an ideal gas which
viscous unstructured grid methodology been the ab- express mass, momentum, and energy for a compress-
sence of a highly stretched unstructured grid generation ible Newtonian fluid in the absence of external forces.
capability. Advances are being made in that area which

tSenior Research Engineer, Senior member AIAA

Copyrillh' @1994 by the AmeriC&lllnautuce of A.,.,uutica aDd Aatrouauuca.
IDe:. ~o copyright ia.-ned in the Uni~ StaWi unGer Title 17. U.S. Code. 1
The U.S. Govenuuent hu a royal'Y-free Iicenae to exerciH all righta under
the copyright claimed herein for Governmental purp_. All Other righta are
rMerVed by the copyright OWIIer.
The equations are given below in integral form for a . J£ aT
bounded domain with a boundaryn an q" = - Pr("'(-I) az
The equations are nondimensionalized with a free-
stream reference values for density Poo and a speed of
sound a oo • Here n:r:, ny, and n" are the Cartesian com-
ponents of the exterior surface unit normal D on the
where
boundary an. The Cartesian velocity components are

Q={~ }
u, v, and W in the :z:, '1/, and z directions, respectively.
The term eo is the total energy per unit volume. The
Prandtl number, Pr, is assigned a value of 0.72. Viscos-
ity, J£, is computed by Sutherland's law. With the ideal
gas assumption, the normalized values for pressure and
and temperature can be expressed as

F(Q) ·n= (V .n) { 2}+p{~}

eo +p 0
and
T=",(p/p
and where "'( is the ratio of specific heats and is prescribed
as 1.4 for air.
G(Q) ·n= MROO(n:r:Gl+nyG,+n,Gs)
eL
Spatial Discretization
A finite-volume discretization is applied to Eq. 1
T:r::r: which results in a consistent approximation to the con-
Gl = 0
T:r:y
} servation laws where the time rate of change of the
{ T:r:.z state vector Q within the domain n is balanced by the
UT:r::r: + VT:r:y + WT:r:.z - q:r: net fluxes of F and G across the boundary surface an.
The spatial domain is divided into a finite number of
tetrahedral cells, with each element serving as a com-
T:r:y
0 } putational cell. Thus, the discretized solution to Eq. 1
G, = Tyy
results in a set of volume-averaged state variables Q
{ Ty.z
which are in balance with the area-averaged fluxes (in-
UT:r:y + VTyy + WTy.z - qy viscid and viscous) across the cell faces.

T:r:.z Inviscid Fluxes

0 }
Gs = Ty.z The inviscid flux quantities are computed across
{ T" each cell face using Roe's23 flux-difference-splitting
UT:r:" + VTy" + WT", - q" (FDS). The implementation of this methodology in the
where present approach is described in Refs. 13 to 15. Roe's
2 FDS technique reconstructs the fluxes by determining
T:r::r: = 2J£u:r: - 3'J£(u:r: + Vy + w,) an approximate solution to an underlying set of Rie-
2 mann problems which still describe the important non-
TYII = 2J£vy - 3'J£(u:r: + Vy + w,) linear behavior of the interacting waves. For that, dis-
continuous states are assumed to exist on either side of
2 a cell interface. The accuracy with which these states
Tn = 2J£w" - 3'J£(u:r: + Vy + w,)
are determined directly impacts the accuracy of the flux
T:r:y =J£(u y + v:r:), T:r:" = J£(u, + w:r:), computation.
The following two subsections describe the present
Ty" =J£(v, + wy) reconstruction approach that is based on a Taylor series
. J£ aT. J£ aT expansion of the cell-averaged solution. A key compo-
q:r: = - Pr( "'( _ 1) az' qy = - Pr( "'( - 1) ay , nent of the scheme is the reconstruction of surrounding
cell-averaged data to a common vertex by a weighted

2
averaging procedure. The prio; av~~aging procedure of Holmes6 •· .That approach is extended for three-
Ref. 13 has proven to be robust and accurate, but is dimensional tetrahedral cells in the following. The
known to be less than second-order accurat.e ils and has method is based on deriving weight factors in Eq. (3)
periodically exhibited some convergence difficulties. A which satisfy the Laplacians
new averaging procedure is presented below which re-
solves these issues. N
L(%n) = ~We,i(%e,i - %n) =0
Higher-Order Cell Reconstruction i=l
A simple, universal formula for reconstructing data N

on tetrahedral cells was introduced in Ref. 14 L(Yn) =L:We,i(Ye,i- Yn) =0 (5)

i=l
N
L(zn) =~We,i(Ze,i - zn) =0
where i=l
q == [p, u, v, w,pf This is a desirable property since the Laplacian of a
and as illustrated in Fig. 1, the subscripts nt, n2, ns linear function is exactly zero. The weights are deter-
denote the nodes comprising face 11,2,3 of cell c and mined by defining
n4 corresponds to the opposite node. The derivation
of this formula has only appeared in Ref. 15 which re- We,i = 1 + ~We,j (6)
ceived limited distribution, and is thus included in the
Appendix of this paper. This formula represents the where
analytical solution to a Taylor series expansion of the (7)
primitive variables from the centroid of a tetrahedral i=l
cell to the centroids of its triangular faces. The state
at the nodes is assumed to be known and is determined is a cost function. The cost function is minimized by
by the reconstruction process described in the next sub- solving an optimization problem subject to the con-
section. straints of Eqs. (5). This optimization problem is solved
by the method of Lagrange multipliers were ~ we,i is
Weighted Averaging Procedure given by
Estimates of the solution are determined at each
node by a weighted average of the surrounding cell-
centered solution quantities:
The solution to the constrained optimization problem
N N yields the Lagrange multipliers
qn = (L:We,iqe,i)/(L:We,i)' (3)
i=l i=l >':11 = [-R:II(1yy1u - I;,,)
+Ry(1.,,,lu - 1:11"1,,,,) (9a)
The subscripts n and c, i refer to the node and sur-
rounding cell-centered values, respectively. The -R"(1.,y1y,, - 1",,1:11")]/ D
method of Ref. 13 employs an inverse-distance weight-
ing factor, i.e. >'11 = [R:r:(1:1111 1u - 1:11,,111 ,,)
1 (9b)
wei= -rj (4) -Ry(I:11::1u - I:,,)
,
+R"(1::::111 ,, - 1:1111 1:11,,)]/D
where
>." = [-R:II(1:11111y" - l y,I:11")
+Ry(I.,::ly" - 1:11111::,,) (9c)
-R"(I":II1y,, - l~y)]/D
This weighting has been used quite successfully for com-
puting the flow over a number of complex configura- where
tions. However, its accuracy is known to be less than D =I:11::(1y"lu - I;,,)
second-order. -I:11y(1:11y1U - 1:11,,111 ,,)
A fully second-order accurate averaging procedure +I:11"(I:11,ly,, - 111 ,1::,,)
was presented by Rausch, et. al. 24 for two-dimensional
triangular cells which was an extension of work by

3
and Viscous Fluxes
N
The viscous fluxes G(Q) are approximated at
R~ = E(Ze,i - zn) the cell-face centroids by first computing the viscous
i=l stresses at the cell vertices, then averaging each compo-
N nent from the three vertices comprising the face. This
R" = E(Ye,i - Yn) approach alleviates the need to store the set of nine
i=l stress terms for each tetrahedral face.
N As is evident in Eq. 1, the viscous stress and heat
Rz = E(Ze,i - Zn) flux terms are constructed from the Cartesian deriva-
i=l tives of te, tI, w, and T. These derivatives are computed
N
directly at the nodes by applying the midpoint trape-
zoidal rule to the gradient theorem:
1~~ = ~)Ze,i - Zn)2
i=1
N VtPn = ,~ 1 tPndS. (10)
vOho
11/1/ = ECYe,i - Yn)2
i=1 where (} is the volume of the domain. A compact sten-
N cil is formed by prescribing for (} an interior subset
1u = L:(Ze,i - Zn)2 of the tetrahedra surrounding a node as illustrated in
i=1 Fig. 2 for a 2D simplification. The domain is defined
N
by scaling down the tetrahedra from I to I', where the
new face centroid coincides with the centroid of cell c.
1~1/ = L:(Ze,i - Zn)(lIe,i - lIn)
Because of the invariant distances noted in Fig. 1, the
i=l
face areas scale in 3D as SI' = (3/4)2 SI' Similarly, the
N
reduced volume is defined by Vel = (3/4)3Ve. Apply-
l~z = E(Ze,i - zn)(Ze,i - zn) ing the midpoint rule to Eq. 10 for the sealed domain
i=l results in
N
Il/z = L:(lIe,i - Yn)(Ze,i - zn).
i=l (11)
These weights are constructed entirely from geomet-
ric information. For some severely distorted grids even
these 'optimum' weights can differ significantly from
unity. Holmes6 proposed clipping all weights to the where Vn denotes the accumulated volume of the un-
range (0,2). The weights were only modified in the sealed tetrahedra Ve,i surrounding node n, nSI,i is the
present work for the viscous solutions. directed area of face i on the unsealed tetrahedra, and
Boundary points are treated in the same manner as lI:(n) are the outer faces I of cells surrounding node n ..
interior points through the use of ghost cells, described The parameter tPe,i == tP I' ,i is the cell-averaged value for
in a subsequent section. In contrast, the scheme of tP in tetrahedral cell c, i.
Ref. 13, which employed the inverse-distance weighting Figure 3 illustrates for a 2D example how bound-
ofEq. (4), constructs the state at boundary nodes from ary faces provide the final closure of the domain sur-
the surrounding face-centered boundary conditions and rounding a boundary node nb. For boundary faces en-
respective distances. countered in Eq. 11, define SI = SI,b and
The weighting factor of Eqs. 6 and 8 produces
an exact interpolation of the nodal quantity Qn for
a linear variation of qe,i on an arbitrary tetrahedral
grid. The entire procedure of 1) averaging to nodes where tPb is the face-centered boundary-condition state
via. Eqs. 3, 6, and 8, and 2) interpolating to cell face value of face band tPn is the value of tP at node n as
via Eq. (2) has been tested for a linear function on determined from the weighted averaging of Eq. 3.
an arbitrary grid. The linear solution is reconstructed
to within machine accuracy, both at the vertices and Improved Data Structure
at the cell faces. Thus, for a nonlinear variation of The FORTRAN computer code reported in Refs.
the cell-averaged state, this pseudo-Laplacian weight- 13 to 15 utilized a cell-based data structure. The flux
ing procedure is second-order accurate in space. calculation and residual accumulation were performed

4
 in two steps. The fluxes were first computed and stored Thus, the solution is computed by solving for all the
by cycling over the cell faces. Then theresidlllUs were unknowns in a partiCUlar color by a point-Jacobi subit-
accumulated by cycling through the tetrahedra and eration step before proceeding to the next color. Since
summing the stored fluxes from the four faces. This the solution of the unknowns for each group can depend
process required an additional connectivity array to on those from previously computed groups, a Gauss-
scatter the information. One advantage of this ap- Seidel-like effect is realized. The method has the ad-
proach was the ease with which it vectorized on current vantage of being completely vedorizable.
supercomputers without further special treatment. Because of the number of operations required to
A face-based data structure21i was explored in an invert a matrix depends on the matrix bandwidth, the
attempt to eliminate the need to store the fluxes for left-hand side of the system of linear equations is eval-
each face, and thereby reduce storage requirements. uated with first-order differencing to reduce both re-
The code was modified to sum the residuals by scatter- quired storage and computer time. Convergence of the
ing the fluxes to the two adjacent cells during the flux subiterations is further accelerated by using Van Leer's
calculation. This code modification not only eliminated Flux Vector Splitting (FVS) on the left-hand side. Thus
the need to store the face fluxes, but also eliminated an in the present study, first-order differencing and FVS
unnecessary connectivity array. are applied to the left-hand side, and higher order dif-
A similar strategy was used to reconstruct the ferencing and FDS to the right-hand side. The viscous
cell-averaged state quantities at the nodes. The pre- terms are not included in'the left-hand side of the equa-
vious approach cycled through the nodes and per- tion at this time.
formed a weighted summation of the surrounding cell- It is necessary to store [A]", which is a 5 x 5 ma-
. centered quantities. The routine was modified to cycle trix for each cell. Thus, storage requirements are in-
through the faces and scatter the adjacent weighted creased from 45 words/cell for the explicit code to 180
cell-averaged quantities to the appropriate nodes. words/ cell for the implicit code. The code requires 64
In order for this scheme to vectorize properly, the JJs/cell/cycle on a CRAY Y-MP with 20 subiterations
faces must be reordered into groups or "colors" such and higher-order differencing. The viscous terms in the
that no two faces share the same cell and opposing right-hand side are absorbed within workspace in the
node. The result of this modification is a 3D-percent re- code and do not require additional memory.
duction in memory from 64 words/cell to 45 words/cell
and a 16-percent speedup from 21 JJS/cell/cycle to 17.5 Boundary Conditions
JJS/cell/cycle on a CRAY Y-MP supercomputer for the The application of the pseudo-Laplacian averag-
explicit Euler code. ing to boundary nodes requires the implementation of
ghost cells. Ghost cells are produced by constructing an
Time Integration image cell across the exterior boundary of an adjacent
interior cell. The geometric information is supplied by
Explicit vector relations
The explicit scheme advances the solution in time
using a 3-stage Runge-Kutta time stepping scheme with Z,h = (Zc,i - Z",) - 2Xn#:
implicit residual smoothing and local time stepping. 1I,h = (1Ic,i - 11",) - 2X n~ (13)
Pertinant details are reported in Refs. 13 through 15.
Z,h = (Zc,i - Z",) - 2Xn"
Implicit
where
Details of the implicit algorithm are described in
Ref. 22. The scheme uses the linearized, backward Eu-
ler time differencing approach to update the solution at
each time step for the set of equations
is a contravariant vector component of distance, and
(12) the subscript nil denotes the a boundary node. These
coordinates are used only in the initial generation and
where storing of the weighting factors and are not used there-
V 8R"
[A]" = Atl+ 8Q after.
The flow information within the ghost cells must
The linear system of equations are solved at each time be updated throughout the solution process. For solid
step with a subiterative procedure where the tetrahe- boundaries, the flow tangency condition is imposed at
dral cells are grouped into "colors" (different from face-
coloring) such that no two cells share a common face.

5
the nodes by constructing an image of the interior ve- domain. The invariants are used to determine the lo-
locity vector within the adjacent ghost cell. As be- cally normal velocity component and speed of sound.
fore, this is accomplished by subtracting twice the con- At an outflow boundary, the two tangential velocity
travariant velocity from the adjacent interior cell veloc- components and the entropy are extrapolated from the
ity vector: interior, while at an inflow boundary they are specified
Ugh = Ue,i - 2U n~ as having far-field values. These five quantities provide
Vgh = Ve,i - 2Uny (14) a complete definition of the flow in the far field.
Implicit boundary conditions are employed with
Wgh = Wc,i - 2Un,
the implicit time integration scheme as described in
where Ref. 22.
U =Uc,in~ + Ve,iny + We,in,.
Results
For viscous, no-slip boundaries, negative values of the
cell-centered velocity components are assigned to the Results are presented in this section to 1) demon-
corresponding ghost cell. The density and pressure strate improved convergence properties of the second-
within the ghost cell are set equal to the values in the order averaging scheme for a representative "problem"
adjacent interior cell. case, 2) assess convergence characteristics of the im-
The face-centered boundary conditions can be de- plicit time integration scheme, and 3) validate the algo-
fined by either a low-order or a higher-order approach. rithm for solving the Navier-Stokes equations for lam-
The low-order approach involves setting the five prim- inar, attached flow. The inviscid results will focus on
itive variables on the boundary faces to their cell- convergence characteristics since the code accuracy is
centered values. The higher-order approach utilizes the well documented in literature16 - 21 .
Taylor series expansion formula of Eq. (2) to construct
a more accurate estimate of the state on the boundary. Weighted Averaging Scheme
Flow tangency is enforced on solid inviscid bound- A surface grid is shown in Fig. 4 for the ONERA
ary faces by subtracting the contravariant velocity com- M6 wing which has a leading edge sweep of 30 de-
ponent normal to the surface. A condition of zero grees, an aspect ratio of 3.8, a taper ratio of 0.56, and
mass and energy flux through the surface is ensured symmetrical airfoil sections. The complete grid con-
by setting the left and right states of solid boundary tains 288,869 cells and 53,791 nodes. An inviscid so-
faces equal to the boundary conditions prior to comput- lution was obtained for Moo = 0.84 and Q = 3.060
ing the fluxes with Roe's approximate Riemann solver. with inverse-distance averaging, Eq. (4), using Runge-
This technique only permits a flux of the pressure terms Kutta time stepping. Convergence was accelerated by
of the momentum equations through a solid boundary. local time stepping and implicit residual smoothing
The viscous, no-slip boundary conditions take ad- with a CFL number of 4. The computational domain
vantage of inherent "structure" in the grid imposed by is bounded by a rectangular box with boundaries at
the Advancing-Layers Method (ALM) of Ref. 12. The -6.5 ~ :z: ~ 11.0, 0.0 ~ y ~ 2.5, and - 6.5 < z < 6.5
ALM produces thin-layer tetrahedral grids with nodes relative to the wing with a semispan of 1. As c~ be
in the "viscous region" distributed along predetermined noted in Fig. 5, the L2-norm enters a limit cycle. This
surface vectors. Thus, as for structured grids, the pres- type of convergence behavior does not occur often with
sure can be extrapolated to the boundary nodes by a Eq. (4), but has surfaced occasionally during wide use
two-point formula. Density is determined by either an of the code. The initial spike at cycle 500 is caused
adiabatic condition or a prescribed temperature, and by the code automatically shifting from first-order to
velocity at the nodes is explicitly set to zero. Once higher-order differencing.
the nodal quantities are set on the boundary, Eq. 2 is A second run was made with the pseudo-Laplacian
applied to determine the face boundary condition for averaging scheme of Eq. (6) with the same low-order
pressure and density. The velocity is explicitly set to boundary conditions. The result in Fig. 5 is a signif-
zero on the boundary faces. icant improvement in L2 convergence and a slight in-
Characteristic boundary conditions are applied to crease in lift coefficient (plotted on an expanded scale).
the far-field subsonic boundary using the fixed and ex- The disturbances in the L2-norm at cycles 2000 and
trapolated Riemann invariants corresponding to the in- 4000 are due to initialization problems during restart,
coming and outgoing waves. The incoming Riemann and are not characteristic of the algorithm.
invariant is determined from the freestream flow and A third run is shown in Fig. 5 with Eq. (6) and
the outgoing invariant is extrapolated from the interior higher-order boundary conditions. (Attempts to ap-
ply higher-order boundary conditions with Eq. (4) were
unsuccessful.) There is little difference in the L2

6
convergence properties, but a '~iigKt difference in lift The c~ii~~fgence history for the implicit scheme
coefficient. The conclusion is that the" second..order ~th.CFL=100 and 20 sub iterations is compared in Fig.
pseudo-Laplacian averaging scheme substanti~ly im- 13 with the explicit RK history from Ref. 26. The im-
proves the robustness and convergence properties of the plicit solution converges in over 6-times less computer
flow solver. time than the explicit scheme. Figure 14 again confirms
that the two solutions yield identical surface pressure
Implicit Time-Integration distributions. The poor agreement in Fig. 14 between
The implicit time integration scheme is investi- the computed solutions and the experimental data from
gated for inviscid flow using the Low-Wing Transport Ref. 27 is unfortunately typical of that produced by
(LWT) configuration shown in Fig. 6 which contained inviscid methods on supercritical airfoils. At the test
135,207 cells and 25,400 nodes. A study is performed Reynolds number of 2.5 x 10 6 , the boundary layer has
to determine combinations of CFL number and num- a pronounced effect of decreasing the aft-camber which
bers of subiterations for the Jacobian matrix which pro- dramatically alters the circulation and resulting shock
duce rapid convergence. Computations are performed 10 cation. 28
at both subsonic and transonic speeds using the the im-
plicit Gauss-Seidel (GS) algorithm and compared with Laminar Viacous Solution
computations reported in Ref. 26 using the explicit The viscous unstructured algorithm is validated for
three-stage Runge-Kutta (RK) time stepping scheme. laminar attached flow over the ONERA M6 wing by
Convergence of the RK scheme is accelerated to steady comparing with structured-grid solutions from
state with implicit residual smoothing, while both RK CFL3D29 on a common base grid. A tetrahedral vis-
and GS employ local time stepping. cous grid was constructed by subdividing a structured
Subsonic. Figure 7 depicts the effect of CFL 97 x 41 x 37 0-0 grid (chord-radial-span) into tetra-
number on solution convergence for Moo = 0.498 and hedral cells (Fig. 15). Thus, each hexahedral cell is
c¥ = 1.615 0 • Convergence to four-orders of magnitude divided into 6 tetrahedra, producing an unstructured
is most rapid for CFL=75. Note that convergence dete- grid with 829,440 cells and 143,705 nodes. A normal
riorates for CFL=100 as a result of decreased diagonal grid spacing of 1.5 x 10- 5 root chord was prescribed at
dominance of [A] in Eq. (12). Figure 8 shows the ef- the surface and stretched geometrically to yield 11 to
fect of the number of subiterations, NSUB, with CFL 12 points in the mid-chord laminar boundary layer for
fixed at 75. The number of global time steps to achieve a Reynolds number of 1 x 106
4-orders of residual reduction, NCYC,are shown for Dual computations were performed with the un-
reference. Note that 20 subiterations produce the most structured and structured codes for Moo = 0.5, c¥ = 30 ,
rapid convergence which is consistent with results from and Re = 1 X 106 • The CFL numbers were ramped lin-
Ref. 22. early from 0.1 to 5 over the first 1000 cycles, and main-
The convergence history from the implicit scheme tained at 5 for the remaining cycles. The unstructured
with CFL=15 and 20 subiterations is compared with algorithm used 20 sub iterations of the Jacobian ma-
the explicit RK history from Ref. 26 in Figure 9. The trix. Figure 16 compares the structured and unstruc-
solution converges to four orders of residual reduction tured surface grid, upper surface pressure contours, and
with 5 to 6 times less computer time for the implicit normalized velocity vectors. Reasonable agreement is
scheme. The lift coefficient for the implicit result con- noted in the flow parameters. The normalized velocity
verges more rapidly and has better damping of low fre- vectors were taken from the first node off the surface
quency oscillations than the explicit scheme. Figure 10 and reflect the directionality of the flow. An aberration
verifies that the surface pressures computed from the in flow direction is present near the apex for the un-
implicit algorithm are identical to those from the ex- structured case which will be addressed in future work.
plicit scheme, which are in reasonable agreement with A comparison of skin-friction coefficients is shown
experimentai data from Ref. 21. in Fig. 17 at 0.44 and 0.90 semispan which shows good
Transonic. Similar results are presented for the agreement between unstructured and structured solu-
LWT at the transonic flow conditions Moo = 0.768 and tions. The differences in resolution at the leading edge
c¥ = 1.1160 in Figs. 11 through 14. Figure 11 shows may be due to the higher spatial resolution of the tetra-
that CFL=100 yields the most rapid convergence of hedral grid. However, a grid sensitivity study is needed
the L2-norm. The effect of the number of Jacobian before drawing firm conclusions. A comparison of pres-
subiterations is shown in Figure 12 where the best value sure coefficients at the same span stations is shown in
is difficult to determine. However, the plots suggest Fig. 18 with excellent agreement. Some anomilies are
that 20 subiterations would be a good compromise. present in the leading-edge recompression region of C,

1
at 17 = 0.90 in Fig. 18, and in the unstructured skin- In two-dimensions, Eq. (A.2) represents a line integral
friction coefficients near the trailing edge in Fig. 17. around some closed path surrounding an area. In three-
Work is underway to investigate the source of these dimensions, it represents a surface integral over a sur-
features. face enclosing a volume. The unit normal Ii is assumed
A comparison of convergence histories is shown in to point outward from the domain.
Fig. 18. The lift coefficient converges to 0.210 for the The components of the gradient in Eq. (A.1) can
unstructured algorithm and 0.202 for CFL3D. Peculiar be computed as in Ref. 13 by coalescing the surround-
convergence behavior is noted for the unstructured s0- ing cell-averaged information to the vertices of the can-
lution after 3500 iterations which needs further investi- didate cell, and then approximating the solution to
gation. The rate of convergence may improve with the Eq. (A.2) with the midpoint trapezoidal rule. Consider
inclusion of the viscous terms in [A]n of Eq. (12). the arbitrary triangular cell in sketch a.

Concluding Remarks
Several improvements to an established three- 3
dimensional unstructured grid flow solver have been
reported which result in a more efficient and robust
flow analysis tool with viscous-flow capability. A new
weighted averaging algorithm is developed which in-
creases the formal accuracy of the base method to
second-order and leads to improved inviscid solution
convergence for complex configurations. A new data Sketch a
structured based on tetrahedral face coloring has been The goal is to expand the solution from the cell cen-
implemented within the code, resulting in a 30-percent troid to the centroids of the edges, e.g. to edge 2-3.
reduction in memory usage and a 16-percent reduc-
Note that the vector ~r is defined along the line passing
tion in required computer time per cycle for the time-
from node 1 through the cell centroid to the centroid of
explicit option of the code. An existing time-implicit
the opposite edge 2-3. When the midpoint trapezoidal
Gauss-Seidel algorithm has been installed which results
rule is applied to the integral in Eq. (A.2) around the
in a factor 6 decrease in total computer time needed to
perimeter of the triangle, then the z-component is writ-
obtain a steady state solution, with a factor of four in-
ten as
crease in memory. Finally, a 3D unstructured laminar
solution on the ONERA M6 wing has been validated
against a structured-grid solution. +(qn~ + qna)n"'23 123
Appendix - Derivation of Extrapolation Formula +( qna + qnl )n"'al1S1]
(A.3)
In the following, a universal expression for the Tay- I
= A [{q12)~Y12
lor series expansion within a triangular cell will be de-
rived. A similar derivation could be carried out for +{q23)~Y23
a tetrahedral cell. The resulting expression requires +(q31)~Y31]
knowledge of the state at the vertices or nodes, which
where A is the area of the triangle, 112 , 123 , 131 are
is obtained by a weighted average from the state at the
edge lengths and n"'12 , n"'2a' n",Sl the z-component of
centers of the cells surrounding the node.
the direction cosines for edges 1-2, 2-3, and 3-1,
The Taylor series expansion within a cell can be
written as (q12) = l/2(qnl + ~2)
q(z, y, z) = q(zc, Yc, zc) + V'qc·.o.r + O(~r2) (q23) = 1/2(qn2 + qraa)
~ q(zc, Ye, zc) + <ulc' (z - Ze) (q31) = l/2(qna + qraJ
+ 'hIe' (y - Ye). and
(A.1)
The area-averaged solution gradient is estimated by
~Y12 =Y2 - Yl, ~Y23 =Y3 - Y2, ~Y31 =Yl - Ys·

evaluating the closed integral over the boundary an Similarly, the y-component can be written as
enclosing the domain n
-1
'h = A [{q12)~Z12 + (q2S)~Z23 + {q31)~Z31]
V'qc = Alo J;JO
1 qildl. (A.2) (A A)

8
where A similar ~eriy:~tion in three-dimensions on a tetrahe-
dral cell yields the expression

'lIl.2.a = Qc+ 1/4 (l/S{Qnl + Q"2 + Q"a) - Qn.]. (A.l0)

As shown in sketch a,
Thus, the state at the edge or face centroid can be
Ar = 1/S{[1/2{Z2 + ZS) - z1]i + [1/2(Y2 + YS) - Y1ti} readily determined by Eq. (A.9) in two dimensions and
Eq. (A.10) in three dimensions, respectively. Recall
= 1/6[(~Z12 - ~zsl)i + (~Y12 - ~1I31)j].
that the edge or face centroid locations are precisely
. (A.S) where the averaged Huxes are evaluated in the finite-
Equation (A.S) applies to any arbitrary triangle due to
volume formulation.
the invariant features that 1) a line extending from a
The Eqs. (A.9) and (A.10) are formally second-
vertex through the cell centroid will always intersect the
order accurate. This becomes evident by observing that
centroid of the opposing edge,. and 2) the distance from
the Taylor series in Eq. (A.1) is second-order accurate
the vertex to the centroid of the triangle is always two-
and that the midpoint trapezoidal rule used to evaluate
tI;Urds of that frpm th~ vertex to the opposing edge.
the integral in Eq. (A.2) is also second-order accurate.
SInce 'Vq = CUi + Cluj, then the first order term in
A second-order accurate nodal averaging procedure is
Eq. (A.l) becomes
presented in the main body of this paper.
1
'Vqc . Ar = 6A [({q12}~Y12 + (q2S}~Y2S Aclcnowledgements
+(q31}~Y31)(~Z12 - a Z31) The author would like to express appreciation to
Dr. Kyle Anderson, NASA Langley Research Center,
-«q12}a Z12 + (q23} aZ 23 for providing subroutines and expertise for installing
+(q31}a Z31)(aY12 - aY3t)] (A.6) his implicit time integration scheme into the How solver.
1 Thanks is also extended to Dr. Tom Roberts, NASA
= 6A [(q23}(aZ12aY23 - aZ23aY12) Langley Research Center for providing the 0-0 grid
+(q23}(~Z23aY31 - aZ 3I aY2S) used in the ONERA M6 wing viscous calculation, and
to Dr. Paresh Parikh, ViGYAN, Inc for the inviscid M6
+({q12) + (Q31})( aZ 12 a Y31 - aZ31~Y12)].
grid. Appreciation is expressed to Dr. David Halt of
By defining the edges as vectors McDonnell Douglas Research Laboratories for bringing
to the authors attention in the literature the second-
L12 = ~Z12i + ~yd order averaging scheme of Rausch et. al, and for sug-
gesting the change in code data structure. The compu-
tations were performed on the CRAY Y-MP at NASA
L 23 = aZ 2si + aY2s.i
Langley Research Center, and on the Numerical Aero-
L31 = aZ31 i + aY3J dynamic Simulation (NAS) facility at NASA Ames Re-
search Center.
then area of the triangle can be computed as
References
2A =1 L12 x L23 1= aZ12aY23 - a Z 23 a Y12
=1 L23 x L31 1= ~Z23aY31 - aZ31~Y23 (A.7) 1 Parikh, P., Pirzadeh, S., and Lohner, R., "A Package
for 3-D Unstructured Grid Generation, Finite-Element
=1 L31 x L12 1= ~Z31~Y12 - ~Z12~1I31' Flow Solutions, and Flow-Field Visualization", NASA
CR-182090, September 1990.
Substitute Eq. (A.7) into Eq. (A.6) yields
2 Pirzadeh, S., "Structured Background Grids for Gen-

'Vq. Ar = 1/3(2(q2S} - «q12) + (q31})] eration of Unstructured Grids by Advancing Front

(A.8) Method", AIAA Journal, Vol. 31, No.2, February
= 1/s(1/2(q"2 + q"a) - q"l]' 1993, pp. 257-265.
3 Pirzadeh, S., "Recent Progress in Unstructured Grid
Substituting Eq. (A.S) into Eq. (A.1) gives the simple
relation Generation", AIAA Paper 92-0445, January 1992.
4 Anderson, W. K., and Bonhaus, D. L.: "An implicit
Upwind Algorithm for Computing Turbulent Flows on
Unstructured Grids", Computers Fluids, Vol. 23, No.1,
pp. 1-21, 1994.

9
IS Nakahashi, N., "FDM-FEM Zonal Approach for Vis- 20 Potsdam, M. A., Intemann, G. A., Frink, N. T.,
cous Flow computations Over Multiple Bodies" AIAAJ Campbell, R. L., Smith, L. A., and Pirzadeh, S.:
Paper 87-0604, January, 1987. "Wing/Pylon Fillet Design Using Unstructured Mesh
6 Holmes, D. G., and Connell, S. D., "Solution of the Euler Solvers", AIAA Paper No. 93-3500, August,
2D Navier-Stokes Equations on Unstructured Adap- 1993.
tive Grids" , presented at the AIAA 9th Computational 21 Ghaft'ari, F.: "On The Vortical-Flow Prediction Ca-
Fluid Dynamics Conference, June, 1989. pability Of An Unstructured Grid Euler Solver", AIAA
7 Mavriplis, D. J., "Multigrid Solution of Compressible Paper No. 94-0163, January, 1994.
Turbulent Flow on Unstructured Meshes Using a Two- 22 Anderson, W. K.: "Grid Generation and Flow So-
Equation Model", AlAA Paper 91-0237, January, 1991. lution Method for Euler Equations on Unstructured
8 Barth, T. J., "Numerical Aspects of Computing Vis- Grids", NASA TM-4295, April 1992.
cous High Reynolds Number Flows on Unstructured 23 Roe, P. L.: Characteristic Based Schemes for the
Meshes", AIAA 91-0721, January, 1991. Euler Equations. Annual Review of Fluid Mechanics,
9 Marcum, D. L.; and Agarwal, R. K.: "A Three- Vol. 18, 1986, pp. 337-365.
Dimensional Finite Element Navier-Stokes Solver with 24 Rausch, R. D., Batina, J. T., and Yang, H. T. Y.,
k-€ Turbulence Model for Unstructured Grids." AlAA "Spatial Adaption Procedures on Unstructured Meshes
Paper 90-1652, 1990. For Accurate Unsteady Aerodynamic Flow Computa-
10 Ward, S., and Kallinderis, Y.: "Hybrid Prismatic/ tion." AIAA Paper 91-1106, April, 1991.
Tetrahedral Grid Generation for Complex 3-D Geome- 25 Jameson, A. and Baker, T. J.: "Improvements to the
tries", AlAA Paper No. 93-0669, January 1993. Aircraft Euler Method", AlAA Paper 87-0452, January
11 Pirzadeh, S.: "Unstructured Viscous Grid Genera- 1987.
tion by Advancing-Layers Method", AIAA Paper No. 26 Frink, N. T., Parikh, P., and Pirzadeh, S., "Aerody-
93-3453, August, 1993. namic Analysis of Complex Configurations using Un-
12 Pirzadeh, S.: "Viscous Unstructured Three-Dimen- structured Grids", AIAA Paper 91-3292, September
sional Grids by the Advancing-Layers Method" , AIAA 1991.
Paper No. 94-0417, January 1994. 27 Pendergraft, O. C.; Re, R. J.; and Kariya, T. T.:
13 Frink, N. T.: "Upwind Scheme for Solving the Eu- "N acelle/Pylon Interference Study on a 1/ 17th-Scale,
ler Equations on Unstructured Tetrahedral Meshes." Twin-Engine, Low-Wing Transport Model", AIAA Pa-
AIAA Journal, Vol., No.1, January, 1992, pp. 70-77. per No. 89-2480, July 1989.
14 Frink, N. T., Parikh, P., and Pirzadeh, S., "A 28 Melson, N. D. and Streett, C. L.: "Updated User's
Fast Upwind Solver for the Euler Equations on Three- Guide for TAWFIVE With Multigrid", NASA TM-
Dimensional Unstructured Meshes", AIAA 91-0102, 4109, 1989.
Jan. 1991. 29 Thomas, J. L., Taylor, S. L., and Anderson, W. K.:
15 Frink, N. T., ''Three-Dimensional Upwind Scheme for "Navier-Stokes Computations of Vortical Flow8 Over
Solving the Euler Equations on Unstructured Tetrahe- Low Aspect Ratio Wings," AIAA Paper No. 87-0207,
dral Grids" ,Ph. D. Dissertation, Virginia Polytechnic 1987.
Institute and State University, September 1991.
16 Dodbele, S. S.: ''Three Dimensional Aerodynamic
Analysis of a High-Lift Transport Configuration", Pa-
per No. AlAA 93-3536, August, 1993.
17 Fouladi, K.: "Unstructured Grids for Sonic-Boom
Analysis" , AIAA Paper No. 93-2929, July, 1993.
18 Kwon, O. J., and Hah, C.: ''Three-Dimensional
Unstructured Grid Euler Method Applied to Turbine
Blades", AIAA Paper No. 93-0196, January, 1993.
19 Parikh, P., Pirzadeh, S., and Frink, N. T.: "Unstruc-
tured Grid Solutions to a Wing/Pylon/Store Config-
uration .Using VGRID3D/USM3D", AIAA Paper No.
92-4572, August, 1992.

10
 4

Cell centroid

Face centroid

Figure 1.- Geometrically invariant features of tetrahedra.

• node

X centroid of cell c

Figure 2.- Stencil for computing viscous stresses at interior nodes (2D simplification).

Figure 3.- Stencil for computing viscous stresses at boundary nodes (2D simplification).

11
 Figure 4.- Surface mesh for ONERA M6 wing.
288,869 cells, 53,791 nodes.

Averasins / B.C.
1 0.30
- - Eq. " / low-order
.. . .. . Eq. 6 / low-order
o . _. -. Eq. 6 / high-order 0.29

-
~ -1
~ ,
0.28
Averaaina / B.C.
:,-
,.'
~-2 .. ,.,
..
,, , - - Eq. " / low-order
~,

,,
.2 ,, . . ... . Eq. 6 / low-order
. --~-~:..
, l'
. _. -. Eq. 6 / high-order
-3 ,. 0.26
.. ~ ~

_, '---'-....&._L....""-.....L.--IL-""-.....L......I~..I.-~-I xl03 0.25 .....................--.."""-""'--'----'"--""'--'----'_"'-.......--1 xl03

o 1 2 3 , 5 6 0 1 2 3 , 5 6
I~aUoD I~aUoD

(b) Lift coefficient

Figure 5.- Effect of weighted-averaging scheme and boundary condition

on convergence history for ONERA M6 wing, Moo = 0.84, a = 3.06 0 •

12
 Figure 6.- Surface grid for Low-Wing Transp ort (LWT) configuration.
135207 cells, 25400 nodes

1 0.9

o 0.8
CFL / NCYC
- - 50 /162 0.7
...... 75 / 139 en. / NCYC
" ----- 100/2 73 01.
- - 50 /162
':';'"
,.,...
.. ....
0.6 ----.- 75 / 139
....:. .......
~.~.:.:.
----- 100/2 73
-3 '->:::-.- 0.5
".

~
-,~~~~--~~~~~--~~~ 0." 25 30
25 30 0 5 10 15 20
o 5 1015 20
CrILY Y-UP cpu time, (min.)
Cray Y -UP cpu time. {min.}
(a) L2-norm (b) Lift coefficient

Figure 7.- Effect of eFL number on solution convergence history for LWT.
20 subiterations, Moo = 0.5, ex = 1.62
0•

13
 1 0.9

NSUB / NCY'C
o 0.8

-o
~
-1
--
------
----.
15/183
20 / 139
25 / 130 0.7
NSUB / NCY'C
---. 30/ 125 Cr.
~
tIII- 2 0.6
-- 15/183
.9 ------ 20 / 139
----. 25 / 130
-3 0.5 ---- 30/125

-.~--------~------~~~----~~
o 5 10 15 20 25 30
0 .•
0 5 10 15 20 25 30
Clay Y-UP cpu time, (min.) Cray Y-UP cpu time, (min.)

(b) Lift coefficient

Figure 8.- Effect of Jacobian subiterations on convergence history for LWT.

CFL=75, Moo = 0.5, c:r = 1.620 •

1 0.9

o 0.8

-o
o
-1
- - Explicit. RIC
. ----- Implicit. GS 0.7

,-
Z

~ 2
Cr.
0.6
.
- Explicit. RIC
. .. ... Implicit. as

-3 0.5

_.L-~~

o 25
__ L-~~~~~--~~~

50 75 . 100 125
0 .• U1-..J---'_-'---'-_.L-~~~~-..I.---"
o 25 50 75 100 125
Clay Y-WP cpu. (min.) Cray Y-WP cpu. (min.)

(b) Lift coefficient

Figure 9.- Effect of time-integration scheme on solution convergence history for LWT.
= =
Moo 0.5, c:r 1.620 •

14
 2.0

1.5
o Expt. (Ref. 27) o Expt. (Ref. 27)
- - Explicit. RIC - Explicit. RIC
. . . . . . Implicit. GS . . . . . . Implicit. GS
1.0

-C,O.5 o 0

o 0
0.0

-0.5

-1.0 _-"-_...........- - 0..............-.11000-_._ _ _ _- - '

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
x/c x/c
(a) 7] = 0.277 (b) 7] = 0.550
Figure 10. Effect of time-integration scheme on chordwise Cp distribution for LWT.
Moo =0.5, =1.62°.
Q

1 1.2
CFL / NCYC
a
-B
:2
-1
--
......
.-._.
--_.
50 /191
75 / 16"
100/ 156
125 / 158
1.0

CP'L / HC'lC
C, 0.8
~
QII- 2
-- 50 / 191
..9 ...... 75 / 164
.-._. 100/ 156
-3 --_. 125 / 158

5 10 15 20 25
Cray Y -UP cpu time. (min.)

(b) Lift coefficient

Figure 11.- Effect of eFL number on solution convergence history for LWT.
15 subiterations, Moo = 0.77, Q = 1.12°.

15
 1 1.2

NSUB / NCYC
o
1.0
-&
~ -1
- - 15/191
·····-20/115
'-'-·25/96 NSUB / NCYC

,-
::£ 2
--_. 30/88
35/87
C, 0.8

0.6
- - 15/191
...... 20/ 115
,-,-·25 / 96
-3 --_. 30/88
35/87

-.~------------~--~--
o 5 10 15
Cray Y-UP cpu Ume, (min.)
.... 20----~
25
0.' L..Ia............_-'---"_"'---£._L............L--..J
o 5 10 15 20
Cray Y-UP cpu time, (min.)
25

(b) Lift coefficient

Figure 12.- Effect of Jacobian subiterations on convergence history for LWT.

CFL=100, Moo = 0.77, Q = 1.120 •

1.2

1.0
- Explicit., RIC
. . . . . . Implicit. GS

C, 0.8
- - Explicit. RIC
. . . . .. Implicit, GS

0.6

25 50 75 100
0.' o'"'"-------'-------"'"-----.....
25 50 75
------
100
Cray Y -UP cpu. (min.) Cray Y-wPcpu, (min.)

(b) Lift coefficient

Figure 13.- Effect of time-integration scheme on solution convergence history for LWT.
Moo = 0.77, Q = 1.12 0 •

16
 0 Expt. (Ref. 27) 0 Expt. (Ref. 27)
- Explicit. RIC - - Explicit. RIC
. . . . . . Implicit. GS Implicit. GS
1.5

0 0 0 0
1.0

-C~0.5 0

0.0

-0.5

-1.0
0.0 0.2 0.' 0.6 0.8 1.00.0 0.2 0.' 0.6 0.8 1.0
x/c x/c
(a) 11 =0.277 (b) 11 =0.550
--
Figure 14. Effect of time-integration scheme on chordwise C" distribution for LWT.
Moo = 0.77, ex = 1.12°.

Figure 15.- Lower hemisphere of structured and unstructured 97 x 41 x 37 0-0 grid for ONERA M6 wing.

17
 Unstructured
Structured

(a) .Surface grid.

(b) Upper-surface Cp contours.

(c) Normalized upper-surface velocity vectors.

Figure 16.- Comparison of structured and unstructured grid and parameters for ON ERA M6 wing.
Moo = 0.5, ex = 3°, Re = 1 X 10 6 , laminar.

18
 - - -', '. , ~

" 0'

0.015

..
0.010
USM3D - - USM3D
C, 0.005 CFL3D CFL3D

0.000

-0.005 - - - - - - - - " -.......---'--....

1 2 3 4 2.5 3.0 3.5 4.0 4.5 5.0
x x
(a) '1 = 0.44 (b) '1 = 0.90
Figure 17.- Comparison of skin-friction distributions for structured and unstructured
laminar solutions for ONERA M6 wing. Moo = 0.5, a = 30 , Re = 1 X 10 6 •

1.5

1.0 - - USM3D USM3D

CFL3D CFL3D
0.5
-cl'
0.0

-0.5
-1.0 ___......_ ......_ ......_I0000._----'I0000._---1
1 2 3 4 2.5 3.0 3.5 4.0 4.5 5.0
x X

" (a) '1 = 0.44 (b) '1 = 0.90

Figure 18.- Comparison of CT' distributions for structured and unstructured
laminar solutions for ONERA M6 wing. Moo = 0.5, Q = 30 , Re = 1 X 10 6•

19
 1 0.5

-B
&
-1
- - USN3D
...... CFL3D
0 .•
- - USW3D
...... CFL3D

CI. 0.3
~
QII- 2
.9
-', 0.2
'.
-3
'.
-.~~--~~--~~--~~--~~~ 0.1~~~~~~~~~--~~--~~
o o
Iteration Iteration

(b) Lift coefficient

Figure 19.- Comparison of convergence characteristics for structured and unstructured

= =
laminar solutions for ONERA M6 wing. Moo 0.5, a 30 , Re 1 X 10 6 • =

20