Computational Fluid Dynamics JOURNAL vol.xx no.x February 2004 (pp.

–)
COMPRESSIBLE FLOW SUPG STABILIZATION
PARAMETERS COMPUTED FROM ELEMENT-EDGE
MATRICES
L. Catabriga

A.L.G.A. Coutinho

T.E. Tezduyar

Abstract
For the streamline-upwind/Petrov-Galerkin (SUPG) formulation of inviscid compressible flows,
we present stabilization parameters calculated based on the element-edge matrices and degree-of-
freedom submatrices of those element-edge matrices. In performance tests involving supersonic
flows and shocks, we compare these stabilization parameters with the ones defined based on the
element matrices. In both approaches the formulation includes a shock-capturing parameter. We
investigate the difference between updating the stabilization and shock-capturing parameters at
the end of every time step and at the end of every nonlinear iteration within a time step. The
formulation also involves an algorithmic feature that is based on freezing the shock-capturing
parameter at its current value when a convergence stagnation is detected.
Key Words: Euler equations, compressible flow, finite elements, edge-based data
structures, stabilization parameter
1 INTRODUCTION
In finite element computation of compressible flows,
the streamline-upwind/Petrov-Galerkin (SUPG) for-
mulation [19, 20, 11] is now one of the most estab-
lished and preferred stabilized formulations. Stabi-
lized formulations such as the SUPG formulation of
compressible flows, SUPG formulation of incompress-
ible flows [8, 1], and the pressure-stabilizing/Petrov-
Galerkin (PSPG) [14] formulation have some well-
pronounced advantages. They prevent numerical in-
stabilities in solving problems with high Reynolds or
Mach numbers and shocks or thin boundary layers,
as well as when using equal-order interpolation func-
tions for velocity and pressure. The SUPG and PSPG
Received on.
∗ Department of Computer Science - Federal University of
Esp´ırito Santo
† Department of Civil Engineering - COPPE - Federal Uni-
versity of Rio de Janeiro
‡ Mechanical Engineering, Rice University
c Copyright: Computational Fluid Dynamics JOUR-
NAL 2002
formulations are among the stabilized methods that
achieve these objectives without introducing excessive
numerical dissipation.
The SUPG formulation for incompressible flows was
first introduced in an ASME paper [8], with further
studies and examples in [1]. The SUPG formulation
for compressible flows was first introduced, in the con-
text of conservation variables, in a NASA technical re-
port [19]. A concise version of the technical report was
published as an AIAA paper [20], and a more thor-
ough version with additional examples as a journal
paper [11]. After that, several SUPG-like methods for
compressible flows were developed. Taylor–Galerkin
method [7], for example, is very similar, and under cer-
tain conditions is identical, to one of the SUPG meth-
ods introduced in [19, 20, 11]. Another example of the
subsequent SUPG-like methods for compressible flows
in conservation variables is the streamline-diffusion
method described in [12]. Later, following [19, 20, 11],
the SUPG formulation for compressible flows was re-
cast in entropy variables and supplemented with a
shock-capturing term [10]. It was shown in [13] that
2 L. CATABRIGA, A.L.G.A. COUTINHO, T.E. TEZDUYAR
the SUPG formulation introduced in [19, 20, 11], when
supplemented with a similar shock-capturing term, is
very comparable in accuracy to the one that was re-
cast in entropy variables. The PSPG formulation for
the Navier–Stokes equations of incompressible flows,
introduced in [14], assures numerical stability while
allowing us to use equal-order interpolation functions
for velocity and pressure. An earlier version of this
stabilized formulation for Stokes flow was introduced
in [9].
A stabilization parameter that is mostly known as
“τ” is embedded in the SUPG and PSPG formula-
tions. It involves a measure of the local length scale
(also known as “element length”) and other parame-
ters such as the element Reynolds and Courant num-
bers. Various element lengths and τs were proposed
starting with those in [8, 1] and [19, 20, 11], followed
by the one introduced in [22], and those proposed in
the subsequently reported SUPG and PSPG methods.
Here we will call the SUPG formulation introduced
in [19, 20, 11] for compressible flows “(SUPG)
82
”,
and the set of τs introduced in conjunction with that
formulation “τ
82
”. The stabilized formulation intro-
duced in [22] for advection–diffusion–reaction equa-
tions included a shock-capturing term and a τ def-
inition that takes into account the interaction be-
tween the shock-capturing and SUPG terms. That
τ definition precludes “compounding” (i.e. augmen-
tation of the SUPG effect by the shock-capturing ef-
fect when the advection and shock directions coin-
cide). The τ used in [13] with (SUPG)
82
is a slightly
modified version of τ
82
. A shock-capturing parameter,
which we will call here “δ
91
”, was embedded in the
shock-capturing term used in [13]. Subsequent mi-
nor modifications of τ
82
took into account the interac-
tion between the shock-capturing and the (SUPG)
82
terms in a fashion similar to how it was done in [22]
for advection–diffusion–reaction equations. All these
slightly modified versions of τ
82
have always been used
with the same δ
91
, and we will categorize them here
all under the label “τ
82-MOD
”.
New ways of computing the τs based on the ele-
ment matrices and vectors were introduced in [21] in
the context of the advection–diffusion equation and
the Navier–Stokes equations of incompressible flows.
These new definitions are expressed in terms of the
ratios of the norms of the matrices or vectors. They
automatically take into account the local length scales,
advection field and the element Reynolds number.
Based on these definitions, a τ can be calculated for
each element or for each degree-of-freedom of each el-
ement, or, as it was proposed in [16, 17], for each
integration point of each element. It was proposed
in [21, 15, 18] that the stabilization parameters to
be used in advancing the solution from time level
n to n + 1 (including the parameter embedded in
a stabilization terms that resembles a discontinuity-
capturing term) should be evaluated at time level n
(i.e. based on the flow field already computed for time
level n), so that we are spared from another level of
nonlinearity.
In [4], the τ definitions based on the element matri-
ces were used in conjunction with the (SUPG)
82
for-
mulation supplemented with the shock-capturing term
involving δ
91
. In conjunction with the same formula-
tion, in [5], we investigated the performances of the
τ definitions based on the element matrices, where a
τ is calculated for each degree-of-freedom of each el-
ement. These concepts are extended in this paper to
the edge-based implementation that was introduced
in [2]. We reported this effort first in a conference pa-
per [6]. Our studies include comparisons between the
τs calculated for each element and for each degree-
of-freedom of each element. We also investigate the
performance differences between calculating the stabi-
lization and shock-capturing parameters at time level
n and at (every nonlinear iteration of) time level n+1.
The performance comparisons are based on test com-
putations involving supersonic flows and shocks. In
all test computations the shock-capturing parameter
is frozen when a convergence stagnation is detected.
2 EULER EQUATIONS
The system of conservation laws governing inviscid,
compressible flows are the Euler equations. In two
SUPG PARAMETERS COMPUTED FROM LOCAL MATRICES 3
dimensions these equations can be written in terms of
the conservation variables, U = (ρ, ρu, ρv, ρe), as
∂U
∂t
+
∂F
x
∂x
+
∂F
y
∂y
= 0 on Ω × [0, T]. (1)
Here ρ is the fluid density, u = (u, v) is the velocity
vector, e is the total energy per unit mass, F
x
and
F
y
are the Euler fluxes, Ω is a domain in IR
2
, and T
is a positive real number. We denote the spatial and
temporal coordinates respectively by x = (x, y) ∈ Ω
and t ∈ [0, T], where the superimposed bar indicates
set closure, and Γ is the boundary of domain Ω. We
consider ideal gases. Alternatively, Equation (1) can
be written as
∂U
∂t
+A
x
∂U
∂x
+A
y
∂U
∂y
= 0 on Ω × [0, T], (2)
where A
x
=
∂Fx
∂U
and A
y
=
∂Fy
∂U
. We assume that
we have an appropriate set of boundary and initial
conditions associated with Equation (2).
3 STABILIZED FORMULATION AND
STABILIZATION PARAMETERS
Considering a standard discretization of Ω into fi-
nite elements, the (SUPG)
82
formulation for the Eu-
ler equations in conservation variables introduced in
[19, 20, 11], supplemented with a shock-capturing
term [13], is written as
_

W
h
.
_
∂U
h
∂t
+A
h
x
∂U
h
∂x
+A
h
y
∂U
h
∂y
_
dΩ +
n
el

e=1
_

e
τ
_
∂W
h
∂x
A
h
x
+
∂W
h
∂y
A
h
y
_
.
_
∂U
h
∂t
+A
h
x
∂U
h
∂x
+A
h
y
∂U
h
∂y
_
dΩ +
n
el

e=1
_

e
δ
91
_
∂W
h
∂x
.
∂U
h
∂x
+
∂W
h
∂y
.
∂U
h
∂y
_
dΩ = 0. (3)
Here W
h
and U
h
are the finite-dimensional test and
trial functions that are defined on standard finite el-
ement spaces. In Equation (3), the first integral cor-
responds to the Galerkin formulation, the first series
of element-level integrals are the SUPG stabilization
terms, and the second series of element-level integrals
are the shock-capturing terms added to the variational
formulation to prevent spurious oscillations around
shocks. The shock-capturing parameter, δ
91
, is eval-
uated here using the approach proposed in [13]. We
define the following element-level matrices:
m :
_

e
W
h
.
∂U
h
∂t
dΩ (4)
¯
k :
_

e
_
∂W
h
∂x
.A
h
x
A
h
x
∂U
h
∂x
+
∂W
h
∂x
.A
h
x
A
h
y
∂U
h
∂y
+
∂W
h
∂y
.A
h
y
A
h
x
∂U
h
∂x
+
∂W
h
∂y
.A
h
y
A
h
y
∂U
h
∂y
_
dΩ
(5)
¯c :
_

e
_
∂W
h
∂x
.A
h
x
∂U
h
∂t
+
∂W
h
∂y
.A
h
y
∂U
h
∂t
_
dΩ (6)
c :
_

e
_
W
h
.A
h
x
∂U
h
∂x
+W
h
.A
h
y
∂U
h
∂y
_
dΩ (7)
The conventional finite element data structure as-
sociates to each triangle e its connectivity, that is, the
mesh nodes I, J and K (see Fig. 1). In the edge-
s
I
J
K
L
f
e
Fig.1: Elements adjacent to edge s formed by nodes I
and J.
based data structure, each edge s is associated to the
adjacent elements e and f, thus to the nodes I, J, K
and L. Moreover, each element matrix can be disas-
sembled into its contributions to three edges, s, s + 1
and s + 2, with connectivities IJ, JK and KI, that
4 L. CATABRIGA, A.L.G.A. COUTINHO, T.E. TEZDUYAR
is,
_
¸
_
• • •
• • •
• • •
_
¸
_
. ¸¸ .
element e
=
_
¸
_
××0
××0
0 0 0
_
¸
_
. ¸¸ .
edge s
+
_
¸
_
0 0 0
0 × ×
0 × ×
_
¸
_
. ¸¸ .
edge s + 1
+
_
¸
_
×0 ×
0 0 0
×0 ×
_
¸
_
. ¸¸ .
edge s + 2
, (8)
where • and × are 4 × 4 submatrices, and 0 repre-
sents a 4 × 4 null matrix. Thus, all the contributions
pertaining to edge s will be present in the adjacent el-
ements e and f. The resulting edge matrix is the sum
of the corresponding sub-element matrices containing
all the contributions to nodes I and J, that is,
_
◦ ◦
◦ ◦
_
. ¸¸ .
edge s
=
_
××
××
_
. ¸¸ .
element e
+
_
××
××
_
. ¸¸ .
element f
, (9)
where ◦ and × also represent 4 × 4 blocks. Consider-
ing a conventional elementwise description of a given
finite element mesh, the topological informations are
manipulated, generating a new edge-based mesh de-
scription. Instead of the element-level matrices de-
fined in Equations (4)-(7), we will use the edge-based
matrices defined as
m =
(A
e
+ A
f
)
12
_
I I
I I
_
(10)
¯
k =
_
−scpg
1
scpg
1
scpg
2
−scpg
2
_
(11)
¯c =
_
−sm
1
sm
2
sm
1
−sm
2
_
(12)
c =
_
−scg
1
scg
1
scg
2
−scg
2
_
, (13)
where A
e
and A
f
are the areas of elements e and f,
and I is the identity matrix of order 4. The submatri-
ces are defined as,
scpg
1
= A
h
x
(s1A
h
x
+ s2A
h
y
) +A
h
y
(s3A
h
x
+ s4A
h
y
) (14)
scpg
2
= A
h
x
(s1A
h
x
+ s3A
h
y
) +A
h
y
(s2A
h
x
+ s4A
h
y
) (15)
sm1 =
1
6

(yKI + yIL)A
h
x
+ (xIK + xLI)A
h
y
)

(16)
sm2 =
1
6

(yJK + yLJ)A
h
x
+ (xKJ + xJL)A
h
y

(17)
scg
1
= sm1 (18)
scg
2
= sm2, (19)
where
s
1
=
1
4A
e
y
JK
y
KJ
+
1
4A
f
y
LJ
y
IL
(20)
s
2
=
1
4A
e
y
JK
x
IL
+
1
4A
f
y
LJ
x
LI
(21)
s
3
=
1
4A
e
x
KJ
y
KI
+
1
4A
f
x
JL
y
IL
(22)
s
4
=
1
4A
e
x
KJ
x
IK
+
1
4A
f
x
JL
x
LI
. (23)
In the equations above, x
ij
= x
i
− x
j
, y
ij
= y
i
− y
j
,
i, j = I, J, K, L.
We define the SUPG parameters from the edge ma-
trices, as given in [21]:
τ
g
=
_
1
τ
r
S1
+
1
τ
r
S2
_
−1/r
, (24)
where
τ
S1
=
c

¯
k
τ
S2
=
∆t
2
c
¯c
. (25)
Here ∆t is the time step, b = max
1≤j≤nee
{|b
1j
| +
|b
2j
| + . . . + |b
nee,j
|}, n
ee
is the number of edge equa-
tions (number of edge nodes times the number of
degrees-of-freedom per node), and r is an integer pa-
rameter.
We can calculate a separate τ for each edge matrix
degree of freedom. The resulting stabilization param-
eter matrix is written as
τ
dof
=
_
¸
¸
¸
¸
_
τ
ρ
τ
u
τ
v
τ
e
_
¸
¸
¸
¸
_
, (26)
where the subindexes (ρ, u, v, e) are the primitive vari-
ables. Each τ
i
can be calculated by using the expres-
sion
τ
i
=
_
1
τ
r
S1i
+
1
τ
r
S2i
_
−1/r
, (27)
SUPG PARAMETERS COMPUTED FROM LOCAL MATRICES 5
where
τ
S1i
=
c
i

¯
k
i

τ
S2i
=
∆t
2
c
i

¯c
i

. (28)
Here c
i
,
¯
k
i
and ¯c
i
are the submatrices of the edge ma-
trices for each degree-of-freedom i = ρ, u, v, e. Theses
submatrices are defined by edge matrix coefficients.
Let a general degree-of-freedom be denoted by i. The
corresponding edge submatrix b
i
can be expressed as
b
i
=
_
b
p1,1
b
p1,2
. . . b
p1,8
b
p2,1
b
p2,2
. . . b
p2,8
_
, (29)
where
if i = ρ then (p1, p2) ≡ (1, 5)
if i = u then (p1, p2) ≡ (2, 6)
if i = v then (p1, p2) ≡ (3, 7)
if i = e then (p1, p2) ≡ (4, 8)
. (30)
The norms of these submatrices, used in Equation
(28), are computed by b
i
= max
1≤j≤nee
{|b
p1,j
| +
|b
p2,j
|}.
The solution is advanced in time by the implicit
predictor-multicorrector algorithm given in [11]. The
resulting linear systems of equations are solved by
a nodal-block diagonal element-by-element precon-
ditioned GMRES method. Although local time-
stepping can be used as shown in [3], all solutions
in this work are obtained using a fixed time-step size
with CFL = 1.
4 NUMERICAL RESULTS
In this section we describe the test computations car-
ried out for two steady-state problems using the ele-
ment and edge data structures. The tolerance of the
preconditioned GMRES algorithm is 0.1, the dimen-
sion of the Krylov subspace is 5, and the number of
multicorrections is 3. All computations are initialized
with the inflow values. The shock-capturing param-
eter is frozen when a convergence stagnation is de-
tected. The symbol τ
g
represents τs calculated based
on element-level and edge-level matrices. The symbol
τ
dof
represents τs calculated based on the degree-of-
freedom sub-matrices of the element-level and edge-
level matrices.
4.1 Oblique Shock
The first problem is a Mach 2 uniform flow over a
wedge, at an angle of −10

with respect to a horizontal
wall. The solution involves an oblique shock at an
angle of 29.3

emanating from the leading edge of the
wedge, as shown in Fig. (2).
x = 0.9
29.3
ο
x
y
M = 2.0
M = 1.64052
Fig.2: Oblique shock – problem description.
The computational domain is a square with 0 ≤ x ≤
1 and 0 ≤ y ≤ 1. Prescribing the following inflow data
on the left and top boundaries results in a solution
with the following outflow data:
Inflow
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
M = 2.0
ρ = 1.0
u = cos10
0
v = −sin10
0
p = 0.17857
Outflow
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
M = 1.64052
ρ = 1.45843
u = 0.88731
v = 0.0
p = 0.30475
(31)
Here M is the Mach number and p is the pressure.
Four Dirichlet boundary conditions are imposed at
the left and top boundaries, the slip condition with
v = 0 is set at the bottom boundary, and no boundary
condition is imposed at the outflow (right) boundary.
A 20×20 mesh with 800 linear triangles and 441 nodes
is employed.
Figures (3)-(6) and Tables (1) and (2) show the per-
formances of the element and edge versions of τ
g
and
τ
dof
.
6 L. CATABRIGA, A.L.G.A. COUTINHO, T.E. TEZDUYAR
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
0 0.2 0.4 0.6 0.8 1
edge
element
exact
(a) Density profile at x = 0.9.
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
0 100 200 300 400 500 600 700 800 900
edge
element
(b) Evolution of density residual.
Fig.3: Oblique shock – solutions and residuals ob-
tained with the element and edge versions of τ
g
and
the iteration update.
Table 1: Oblique shock – computational costs (in
number of time steps, GMRES iterations and CPU
seconds) for the element and edge versions of τ
g
.
Iteration update
Data structure N
GMRES
N
steps
time(sec)
Edge 4,162 797 84
Element 4,287 846 149
Time-step update
Data structure N
GMRES
N
steps
time(sec)
Edge 4,845 833 85
Element 4,246 838 145
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
0 0.2 0.4 0.6 0.8 1
edge
element
exact
(a) Density profile at x = 0.9.
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0 100 200 300 400 500 600 700 800 900
edge
element
(b) Evolution of density residual.
Fig.4: Oblique shock – solutions and residuals ob-
tained with the element and edge versions of τ
g
and
the time-step update.
Table 2: Oblique shock – computational costs (in
number of time steps, GMRES iterations and CPU
seconds) for the element and edge versions of τ
dof
.
Iteration update
Data structure N
GMRES
N
steps
time(sec)
Edge 4,088 787 88
Element 4,902 833 97
Time-step update
Data structure N
GMRES
N
steps
time(sec)
Edge 4,965 822 87
Element 4,857 819 93
SUPG PARAMETERS COMPUTED FROM LOCAL MATRICES 7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
0 0.2 0.4 0.6 0.8 1
edge
element
exact
(a) Density profile at x = 0.9.
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0 100 200 300 400 500 600 700 800 900
edge
element
(b) Evolution of density residual.
Fig.5: Oblique shock – solutions and residuals ob-
tained with the element and edge versions of τ
dof
and
the iteration update.
4.2 Reflected Shock
This problem consists of three regions (R1, R2 and
R3) separated by an oblique shock and its reflection
from a wall, as shown in Fig. (7). Prescribing the
following Mach 2.9 inflow data in the first region on
the left (R1), and requiring the incident shock to be at
an angle of 29

, leads to the following exact solution
at the other two regions (R2 and R3):
R1
_
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
_
M =2.9
ρ =1.0
u =2.9
v =0.0
p =0.714286
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
0 0.2 0.4 0.6 0.8 1
edge
element
exact
(a) Density profile at x = 0.9.
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0 100 200 300 400 500 600 700 800 900
edge
element
(b) Evolution of density residual.
Fig.6: Oblique shock – solutions and residuals ob-
tained with the element and edge versions of τ
dof
and
the time-step update.
R2
_
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
_
M = 2.3781
ρ = 1.7
u = 2.61934
v =−0.50632
p = 1.52819
R3
_
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
_
M =1.94235
ρ =2.68728
u =2.40140
v =0.0
p =2.93407
. (32)
The computational domain is a rectangle with 0 ≤
x ≤ 4.1 and 0 ≤ y ≤ 1. We prescribe the density,
velocities and pressure at the left and top boundaries,
8 L. CATABRIGA, A.L.G.A. COUTINHO, T.E. TEZDUYAR
ο
x
y
y=0.25
1
2
3
M=2.9
M=2.3781
M=1.94235
29
ο
23.28
Fig.7: Reflected shock – problem description.
the slip condition with v = 0 is imposed at the bottom
boundary, and no boundary condition is imposed at
the outflow (right) boundary. We use an unstructured
mesh with 1,837 nodes and 3,429 elements.
(a) With τ
g
and iteration update.
(b) With τ
g
and time-step update.
(c) With τ
dof
and iteration update.
(d) With τ
dof
and time-step update.
Fig.8: Reflected shock – density contours obtained
with the edge versions of τ
g
and τ
dof
.
Figures (8)-(11) and Tables (3) and (4) show the
performances of the element and edge versions of τ
g
and τ
dof
.
5 CONCLUDING REMARKS
We highlighted, for the SUPG formulation of invis-
cid compressible flows, stabilization parameters de-
fined based the element-edge matrices and the degree-
(a) With τ
g
and iteration update.
(b) With τ
g
and time-step update.
(c) With τ
dof
and iteration update.
(d) With τ
dof
and time-step update.
Fig.9: Reflected shock – density contours obtained
with the element versions of τ
g
and τ
dof
.
Table 3: Reflected shock – computational costs (in
number of time steps, GMRES iterations and CPU
seconds) for the element and edge versions of τ
g
.
Iteration update
Data structure N
GMRES
N
steps
time(sec)
Edge 10,237 787 500
Element 9,127 844 948
Time-step update
Data structure N
GMRES
N
steps
time(sec)
Edge 11,161 846 544
Element 9,913 893 984
of-freedom submatrices of those element-edge matri-
ces. These definitions are expressed in terms of the
ratios of the norms of the matrices, and take auto-
matically into account the flow field, the local length
scales, and the time step size. We carried out a num-
ber of test computations involving supersonic flows
and shocks. By inspecting the solution quality and
SUPG PARAMETERS COMPUTED FROM LOCAL MATRICES 9
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
0 100 200 300 400 500 600 700 800 900
Edge
Element
(a) With iteration update.
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0 100 200 300 400 500 600 700 800 900
Edge
Element
(b) With time-step update.
Fig.10: Reflected shock – evolution of the density
residuals obtained with the element and edge versions
of τ
g
.
convergence history, we compared these stabilization
parameters with the ones defined based on the ele-
ment matrices. In both cases the formulation includes
a shock-capturing parameter. Also by inspecting the
solution quality and convergence history, we investi-
gated the performance difference between updating
the stabilization and shock-capturing parameters at
time level n and at (every nonlinear iteration of) time
level n + 1. The formulation also involves activat-
ing an algorithmic feature that is based on freezing
the shock-capturing parameter at its current value
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
0 100 200 300 400 500 600 700 800 900
Edge
Element
(a) With iteration update.
1e-11
1e-10
1e-09
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0 100 200 300 400 500 600 700 800 900
Edge
Element
(b) With time-step update.
Fig.11: Reflected shock – evolution of the density
residuals obtained with the element and edge versions
of τ
dof
.
when a convergence stagnation is detected. We ob-
serve that in all cases the solution qualities are very
comparable. In terms of computational efficiency, τ
definitions based on the element matrices are respond-
ing better to using degree-of-freedom submatrices (in
place of full matrices) than the definitions based on
the edge matrices. The edge-based implementation is
computationally more efficient than the element-based
implementation. However, this advantage is less pro-
nounced for the degree-of-freedom submatrices than it
is for the full matrices.
10 L. CATABRIGA, A.L.G.A. COUTINHO, T.E. TEZDUYAR
Table 4: Reflected shock – computational costs (in
number of time steps, GMRES iterations and CPU
seconds) for the element and edge versions of τ
dof
.
Iteration update
Data structure N
GMRES
N
steps
time(sec)
Edge 9,689 867 544
Element 9,138 845 584
Time-step update
Data structure N
GMRES
N
steps
time(sec)
Edge 9,133 862 528
Element 9,278 817 579
REFERENCES
[1] A.N. Brooks and T.J.R. Hughes. Streamline
upwind/Petrov-Galerkin formulations for convec-
tion dominated flows with particular emphasis
on the incompressible Navier-Stokes equations.
Computer Methods in Applied Mechanics and En-
gineering, 32:199–259, 1982.
[2] L. Catabriga and A.L.G.A. Coutinho. Implicit
SUPG solution of Euler equations using edge-
based data structures. Computer Methods in Ap-
plied Mechanics and Engineering, 191:3477–3490,
2002.
[3] L. Catabriga and A.L.G.A. Coutinho. Improv-
ing convergence to steady-state of implicitSUPG
solution of Euler equations. Communications in
Numerical Methods in Engineering, 18:345–353,
2002.
[4] L. Catabriga, A.L.G.A. Coutinho, and T.E. Tez-
duyar. Finite element SUPG parameters com-
puted from local matrices for compressible flows.
In Proceedings of the 9th Brazilian Congress of
Engineering and Thermal Sciences, Caxambu,
Brazil, 2002.
[5] L. Catabriga, A.L.G.A. Coutinho, and T.E. Tez-
duyar. Finite element SUPG parameters com-
puted from local dof-matrices for compressible
flows. In Proceedings of the 24th Iberian Latin-
American Congress on Computational Methods in
Engineering, Ouro Preto, Brazil, 2003.
[6] L. Catabriga, A.L.G.A. Coutinho, and T.E. Tez-
duyar. Finite element SUPG parameters com-
puted from local edge matrices for compressible
flows. In Proceedings of the 17th International
Congress of Mechanical Engineering, Sao Paulo,
Brazil, 2003.
[7] J. Donea. A Taylor-Galerkin method for convec-
tive transport problems. International Journal for
Numerical Methods in Engineering, 20:101–120,
1984.
[8] T.J.R. Hughes and A.N. Brooks. A multi-
dimensional upwind scheme with no crosswind
diffusion. In T.J.R. Hughes, editor, Finite Ele-
ment Methods for Convection Dominated Flows,
AMD-Vol.34, pages 19–35. ASME, New York,
1979.
[9] T.J.R. Hughes, L.P. Franca, and M. Balestra. A
new finite element formulation for computational
fluid dynamics: V. Circumventing the Babuˇska–
Brezzi condition: A stable Petrov–Galerkin for-
mulation of the Stokes problem accommodating
equal-order interpolations. Computer Methods
in Applied Mechanics and Engineering, 59:85–99,
1986.
[10] T.J.R. Hughes, L.P. Franca, and M. Mallet. A
new finite element formulation for computational
fluid dynamics: VI. Convergence analysis of the
generalized SUPG formulation for linear time-
dependent multi-dimensional advective-diffusive
systems. Computer Methods in Applied Mechan-
ics and Engineering, 63:97–112, 1987.
[11] T.J.R Hughes and T.E. Tezduyar. Finite ele-
ment methods for first-order hyperbolic systems
with particular emphasis on the compressible Eu-
ler equations. Computer Methods in Applied Me-
chanics and Engineering, 45:217–284, 1984.
[12] C. Johnson, U. Navert, and J. Pitk¨aranta. Finite
element methods for linear hyperbolic problems.
Computer Methods in Applied Mechanics and En-
gineering, 45:285–312, 1984.
[13] G.J. Le Beau and T.E. Tezduyar. Finite element
computation of compressible flows with the SUPG
SUPG PARAMETERS COMPUTED FROM LOCAL MATRICES 11
formulation. In Advances in Finite Element Anal-
ysis in Fluid Dynamics, FED-Vol.123, pages 21–
27, New York, 1991. ASME.
[14] T.E. Tezduyar. Stabilized finite element formula-
tions for incompressible flow computations. Ad-
vances in Applied Mechanics, 28:1–44, 1991.
[15] T.E. Tezduyar. Adaptive determination of the
finite element stabilization parameters. In Pro-
ceedings of the ECCOMAS Computational Fluid
Dynamics Conference 2001 (CD-ROM), Swansea,
Wales, United Kingdom, 2001.
[16] T.E. Tezduyar. Calculation of the stabilization
parameters in SUPG and PSPG formulations. In
Proceedings of the First South-American Congress
on Computational Mechanics (CD-ROM), Santa
Fe–Parana, Argentina, 2002.
[17] T.E. Tezduyar. Calculation of the stabilization
parameters in finite element formulations of flow
problems. to appear in Journal of Computational
Methods in Sciences and Engineering, 2003.
[18] T.E. Tezduyar. Computation of moving bound-
aries and interfaces and stabilization parameters.
International Journal for Numerical Methods in
Fluids, 43:555–575, 2003.
[19] T.E. Tezduyar and T.J.R. Hughes. Develop-
ment of time-accurate finite element techniques
for first-order hyperbolic systems with particu-
lar emphasis on the compressible Euler equa-
tions. NASA Technical Report NASA-CR-
204772, NASA, 1982.
[20] T.E. Tezduyar and T.J.R. Hughes. Finite element
formulations for convection dominated flows with
particular emphasis on the compressible Euler
equations. In Proceedings of AIAA 21st Aerospace
Sciences Meeting, AIAA Paper 83-0125, Reno,
Nevada, 1983.
[21] T.E. Tezduyar and Y. Osawa. Finite element sta-
bilization parameters computed from element ma-
trices and vectors. Computer Methods in Applied
Mechanics and Engineering, 190:411–430, 2000.
[22] T.E. Tezduyar and Y.J. Park. Discontinuity cap-
turing finite element formulations for nonlinear
convection-diffusion-reaction problems. Computer
Methods in Applied Mechanics and Engineering,
59:307–325, 1986.