ijnmf_srinivasan_2011
TRANSCRIPT
-
8/3/2019 ijnmf_srinivasan_2011
1/25
INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDSInt. J. Numer. Meth. Fluids (2011)Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/fld.2701
An upwinded state approximate Riemann solver
B. Srinivasan1, *, , A. Jameson2 and S. Krishnamoorthy1
1Department of Applied Mechanics, Indian Institute of Technology Delhi, Hauz Khas, New Delhi-110016, India2Department of Aeronautics and Astronautics, Stanford University, Durand Building, 496 Lomita Mall, Stanford,
CA 94305-4035, USA
SUMMARY
Stability is achieved in most approximate Riemann solvers through flux upwinding, where the flux at theinterface is arrived at by adding a dissipative term to the average of the left and right flux. Motivated by theexistence of a collapsed interface state in the gas-kinetic BhatnagarGrossKrook (BGK) method, an alter-native approach to upwinding is attempted here; an interface state is arrived at by taking an upwinded average
of left and right states, and then the flux is calculated as a function of this collapsed interface state. Thisso called state-upwinding approach gives rise to a new scheme called the linearized Riemann solver forthe Euler and NavierStokes equations. The scheme is shown to be closely associated with the Roe scheme.It is, however, computationally less expensive and gives qualitatively comparable results over a wide rangeof problems. Most importantly, this scheme is found to preserve stationary contacts while not exhibiting thecarbuncle phenomenon which plagues the Roe and other contact-preserving schemes. The scheme is there-fore motivated as a new starting point to analyze the origin of the carbuncle phenomenon. Copyright 2011John Wiley & Sons, Ltd.
Received 17 March 2011; Revised 7 September 2011; Accepted 18 September 2011
KEY WORDS: finite volume; hyperbolic; hypersonic; compressible flow; euler flow; NavierStokes;transonic
1. INTRODUCTION
Finite volume schemes have been widely used to derive computational schemes for flow problems,
because of their desirable property of maintaining conservation at the level of discretization. Various
finite volume schemes arise depending on how the flux at the interface between two cells is formu-
lated. A primary use of finite volume schemes has been in solving problems with stationary or
moving discontinuitieswhere the property of being conservative is especially useful in maintain-
ing the right discontinuity speed. Such discontinuities often arise naturally in hyperbolic partial
differential equations describing conservation laws such as the Euler equations.
In addition to conservation, problems involving shocks are also found to require additional dis-
sipative terms for stabilization and in the absence of any explicit shock tracking for shock
capturing. These stabilization or artificial diffusion terms have been motivated and derived in var-
ious ways some schemes add these terms explicitly and others use some feature of the physics ofthe problem equivalent to adding some stabilization terms. In the latter category, one of the most
popular class of finite volume methods employed in CFD problems is the one that employs Riemann
solvers. First suggested by Godunov [1], this method involves a piecewise constant discretization
of the fluid properties under examination and then calculating the evolution by means of solving a
series of Riemann problems. In order to avoid the expense of an exact Riemann solver, Roe pio-
neered the use of cheaper, approximate Riemann solvers [2]. Roes method, which is still widely
*Correspondence to: B. Srinivasan, Department of Applied Mechanics, Indian Institute of Technology Delhi, Hauz Khas,New Delhi-110016, India.
E-mail: [email protected]
Copyright 2011 John Wiley & Sons, Ltd.
-
8/3/2019 ijnmf_srinivasan_2011
2/25
-
8/3/2019 ijnmf_srinivasan_2011
3/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
3. Apply the upwinding constraint on the Riemann invariants from the left and right state as
Jinterfacei D
JLi , ci > 0
JRi , ci < 0.(1)
4. Recompute the primitive (or conservative) variables at the interface, W.WL, WR/ of the
system from the characteristic variables J
interface
i obtained from the previous step.5. Compute the interface flux F.WL, WR/D F .W
/.
2.1.1. Scalar advection. For the scalar linear advection equation ut C cux D 0, upwinding can beachieved by evaluating
ujC1=2 D
uL , c > 0
uR , c < 0.(2)
Here, the states uL and uR represent, respectively the left and right estimates for the interfacestate. These estimates can be obtained from some local interpolation defined in the cells. Then, the
interface flux can be evaluated as f .ujC1=2/D cujC1=2. For constant interpolants within each celland c > 0, this results in the familiar upwind scheme
Duj
DtC c
uj uj1
xjC1=2 xj1=2D 0. (3)
2.2. System of equations
The aforementioned method can be extended to the linear hyperbolic system of equations
Wt CAWx D 0. (4)
State upwinding can be achieved by writing the system in characteristic form and upwinding the
Riemann invariants along the respective characteristics. For example, the linear system of equations
ut C cvx D 0,
vt C cux D 0,(5)
have the characteristic speeds c and Riemann invariants u v. State upwinding can now beperformed using
uC v D
uLC vL , c > 0
uR C vR , c < 0,(6)
u v D
uR vR , c > 0
uL vL , c < 0,(7)
where u and v are the desired upwinded-state variables. These two equations can now be solved forujC1=2 and vjC1=2, and the flux can be evaluated as
f .WjC1=2/D
"cvjC1=2
cujC1=2
#. (8)
It is worth noting that the final result is equivalent to the Roe and exact Riemann solvers for
the linear system of equations. For nonlinear systems in which Riemann invariants can be found, a
state upwinding procedure via Riemann invariants can be used. For example, for the 1-D shallow
water equations, "u
gh
#t
C
"u2
2C gh
ugh
#x
D 0. (9)
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
4/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
The characteristic speeds are u c and Riemann invariants are u 2c
u 2c D
uL 2cL , u0 c0 > 0
uR 2cR , u0 c0 < 0,(10)
where the state W0 is evaluated from some suitable average ofWL and WR. The conservative fluxes
can now be evaluated as in the linear case. This upwinded-state scheme for the nonlinear system isnow no longer equivalent to the Roe scheme.
2.3. State upwinding for Euler equations
We can now attempt to apply the aforementioned method to the 1-D Euler equations. These can be
written in characteristic form as @
@t
1
c2@p
@t
C u
@
@x
1
c2@p
@x
D 0, (11)
@u
@tC
1
c
@p
@t C .uC c/
@u
@xC
1
c
@p
@xD 0, (12)
@u
@t
1
c
@p
@t
C .u c/
@u
@x
1
c
@p
@x
D 0, (13)
where c Dq
p
. These can be written in the equivalent form
D
Dt
1
c2Dp
DtD 0, (14)
DCu
DtC
1
c
DCp
DtD 0, (15)
Du
Dt
1
c
Dp
DtD 0, (16)
where
D
Dt
@
@tC u
@
@x, (17)
DC
Dt
@
@tC .uC c/
@
@x, (18)
D
Dt
@
@t C .u c/
@
@x , (19)
are the directional derivatives along the respective characteristic lines.
2.3.1. Riemann invariants for Euler equations. The characteristic equations previously written are
not in invariant form, that is, one cannot directly determine from the aforementioned equations
what the conserved quantity along each characteristic is. Equation (14) can be multiplied by the
integrating factor 1
to obtain the invariant form
Dv0
DtD 0, (20)
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
5/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
where v0 D log.p/, the entropy, is the invariant along this characteristic. In the isentropic
case, the first characteristic equation becomes redundant and the other two equations can be directly
integrated to obtain the invariants
v1 D uC2c
1, (21)
v2 D u2c
1. (22)
Perhaps, because of this special case, it is a common practice to represent the invariants in the
general, anisentropic case as
v1 D uC
Zdp
c, (23)
v2 D u
Zdp
c. (24)
However, the term uC Rdpc
cannot be evaluated as an indefinite integral in the general case as u is
a state function, whereasRdp
c is path dependent.This, of course, does not preclude the possibility of there being an integrating factor for these
characteristics. However, the existence of an integrating factor for differential forms with three
independent variables is not guaranteed [11]. In fact, as will be shown here, an integrating fac-
tor does not exist for either of these characteristics in general, nonbarotropic case when p and areindependent variables.
We begin by assuming that there exists an integrating factor for the third characteristic equation;
that is, the differential form du dpc
. Then, using c2 D p
, we obtain the differential form
cdu 2cdc c2d.
This has an integrating factor if and only if
du 2dc cd
has an integrating factor. Now, the Pfaffian u1.x, y, /dx C u2.x, y, /dy C u3.x, y, /d has anintegrating factor if and only if [11]
!u r !u D 0. (25)
We get for the present Pfaffian
!u r !u Dc2, (26)
and hence, no integrating factor exists along this characteristic. Note that it is not that an analytical
expression cannot be found for the invariants or that they are not well-known properties of the fluid;
there are simply no invariants (in the sense of properties preserved along the characteristics) in the
general case when p and are independent.An alternate interpretation of the characteristics [12] is that they are the path lines along whichsmall perturbations of flow variables travel. For example, for the scalar wave equation utCcux D 0,
perturbations u of the mean flow u0.x, t/, satisfyD.u/Dt
= 0, that is, u D const . along xct D 0.Similarly, for the Euler equations, one has
D
Dt
p
c2
D 0, (27)
Note that this substitution only makes the existence of three independent variables explicit and does not add any furtherconstraints to the original differential form.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
6/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
DC
Dt
uC
p
c
D 0, (28)
D
Dt
u
p
c
D 0. (29)
Using the aforementioned expressions, we can now find the relations for the numerical flux for
the LRS solver.
3. LINEARIZED RIEMANN STATE SCHEME
The interpretation in the previous section can now be utilized to derive an upwinded-state scheme.
For example, for the advection equation, we define the mean state
u0 DuLC uR
2. (30)
Unless the flow is extremely well resolved by the grid, in a general region of the flow, the left and
right state and the aforementioned average are all different from each other. Then, we may interpret
the situation at the interface as follows The left cell tries to change this average, interface state u0
to its current state, that is, uL, whereas the right cell similarly attempts to change it to uR. That is,the left cell effects a perturbation uL D uL u0 over the average state, whereas the right effectsthe perturbation uR D uR u0. Then, an upwinded average state at the interface may be obtainedas u D u0C u
, where
u D
uL , c > 0
uR , c < 0,(31)
where the perturbation has been upwinded using the characteristic direction. As can be quickly seen,
this results in exactly the same scheme as the one obtained via Equation (2).
For the Euler equations, one may similarly upwind along the characteristic directions, starting
from Equation (27). It is easiest to find an expression for the upwinded state starting from primitive
variables (we have not found any qualitative difference in using conservative variables for defining
the upwinded state)
W D
264
u
p
375 . (32)
We can define the upwinded state as
W DW0C W, (33)
where
W0 DWLCWR
2(34)
and W is obtained by solving
p
c20D
8 0
R pR
c2, u0 < 0,
(35)
uCp
0c0D
8 0
uRC pR
0c0, u0C c0 < 0
(36)
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
7/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
u p
0c0D
8 0
uR pR
0c0, u0 c0 < 0.
(37)
The aforementioned process is equivalent to solving the linearized Riemann problem with the states
WL and WR for the state W. This particular upwinded-state scheme has hence been called the
LRS scheme here. It can be shown with a small amount of algebraic manipulation that the solutionto the aforementioned system can be written in the form
W DR0 sgn.0/ R10
WR WL
2
, (38)
where R0 is the matrix of Eigen vectors
RD
2664
1 2c
2c
0 1=2 1=2
0 c2
c2
3775
, (39)
evaluated at the state W0. Similarly,
R1 D
2664
1 0 1c2
0 1 1c
0 1 1c
3775 , (40)
and
D
264
u 0 0
0 uC c 0
0 0 u c
375 . (41)
It can be seen from Equation (38) that W D sgn.u0/WL when ju0j> c0, that is, as with the Roe
scheme, the upwinding is sharp when the flow is locally supersonic. As will be seen, this property
is essential for maintaining stationary shocks.
Significant computational savings can be made by noting that for ju0j< c0, the following simpleexpressions for the effective perturbations result:
p D 0c0uL, (42)
u DpL
0c0, (43)
D sgn.u0/
L
pL
c20
C
p
c20. (44)
In addition to the fact that the LRS uses only one flux evaluation per interface, whereas the Roe
(like all upwinded flux schemes) uses two; the existence of such simple expressions makes the LRS
scheme cheaper than the Roe scheme. As will be seen, this comes without any loss of quality of
shock or contact resolution. In fact, as will be seen later, there seems to be an actual advantage over
Roes scheme when it comes to handling the carbuncle phenomenon.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
8/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
4. SOME PROPERTIES OF THE LINEARIZED RIEMANN SOLVER SCHEME
4.1. Relation with Roes scheme
The form (38) is a reminiscent of the following way of expressing Roes scheme
F .WRoejC1=2/DF .WL/CF .WR/
2
jAjRoeWR WL
2
. (45)
To see the exact correspondence between Roe and LRS, we note that
WR DWLCWR
2
WL WR
2, (46)
WL DWLCWR
2C
WL WR
2. (47)
Hence,
F .WR/D F
WLCWR
2
A0
WL WR
2
CH.O.T, (48)
F .WL/D FWLCWR2
CA0 WL WR2
CH.O.T, (49)where H.O.T represents higher order terms. Similarly, for the LRS scheme,
F .WLRSjC1=2/D F
WLCWR
2
A0 sgn.A0/
WL WR
2CH.O.T. (50)
Combining these, we get
F .WLRSjC1=2/DF .WL/CF .WR/
2 jAj0
WR WL
2CH.O.T. (51)
The LRS and Roe scheme are then identical up until the lowest order dissipative term except for
the state at which jAj is evaluated. Surprisingly enough, this difference causes not only the expectedquantitative differences but qualitative differences as well, as will be seen in Section 7.
4.2. Treatment of discontinuities
Because of the way that the Roe state is constructed, that is,
fR fL D ARoe.WR WL/, (52)
it supports a stationary shock as well as a stationary contact. We try and investigate here LRSs
behavior in these situations.
Stationary shock: For a stationary shock, we have
c20 D pLC pR
LC R. (53)
It can be shown on the substitution of the RankineHugoniot jump conditions that for a station-
ary shock juLCuR
2 j > c0, if uL and uR satisfy the shock jump conditions and c0 is computed asmentioned earlier. Hence,
WjC1=2 D
WL u0 > 0
WR u0 < 0,(54)
)
F .WiC1=2/D
F .WL/ u0 > 0
F .WR/ u0 < 0.(55)
and because F .WL/D F .WR/ for a shock, the shock is maintained as is.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
9/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
Stationary contact:
Here,
pR D pL D p, (56)
uL D uR D 0, (57)
L R. (58)
Hence,
F .WR/D F .WL/D
0BB@
LuL
Lu2LCpL
LuLHL
1CCAD
0@ 0p
0
1A . (59)
Now, because
uL D pL D 0,
we have,
u D 0,
and
p D 0.
Then, from Equation (44),
D sgn.u0/
L
pL
C20
C 0
uL
C0(60)
D sgn.u0/L, (61)
)
WiC1=2 D
0C L 0 p
T , (62)
)
F .WiC1=2/D
24 0p
0
35 . (63)
Hence, the stationary contact is maintained.
One-point Shock : A more important property of the Roe scheme is its ability to hold an ideal
one point shock (see [13] for a proof). We see if something can be said about the LRS scheme in
this situation.
Consider the situation
WjD
8 is,
(64)
with WL and WR satisfying the normal shock jump conditions (i.e F .WL/ D F .WR/). Then, tomaintain this one point structure, we need
1. F .WL/D F .Wis1=2/ for equilibrium at is 1.2. F .Wis1=2/D F .WisC1=2/ for equilibrium at is .3. F .WisC1=2/D F .WisC 32
/ for equilibrium at is C 1.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
10/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
Condition .1/ would be met for any WA with uL > uA > uR, because, then u
RCuL
2< c0,
uLCuA
2< c0. Hence, Wis1=2 DWL and F .WL/D F .Wis1=2/.
Now, if condition .2/ is met, condition .3/ is automatically satisfied because F .WisC32/ D
F .WR/ D F .WL/ D F .Wis1=2/. We require that F .WL/ D F .WisC1=2/. It can be shown thatF .W1/ D F .W2/ supports only the two states given by the normal shock jump conditions. Hence,we need WisC1=2 DWL or WisC1=2 DWR. That is,
WisC1=2 DWACWR
2
1
2sgn.AisC1=2/.WR WA/. (65)
Consider WisC1=2 DWR. This is not possible because it would require that
sgn.AisC1=2/DI, (66)
where I is the identity matrix. This is not possible because from condition(1), uL > uA > uR > 0and hence, at least one eigenvalue ofAisC1=2 is positive.
Now, consider WisC1=2 D WL. This can be split into two further cases. If W0 DWACWR
2is such
that u0 > c0, then,
WisC1=2 DWACWR
2 1=2
WR WA
2DWA WL, (67)
because we want an intermediate state and not a zero-point shock. Hence, a supersonic intermediate
point is ruled out.
IfW0 DWACWR
2is such that u0 < c0, then, from Equations (42) and (43), we have
p DpACpR
2C 1=2.uA uR/0c0 (68)
u DuAC uR
2
C 1=2pA pR
0c0. (69)
Because we require that p D pL, this leads to
pL pACpR
2D 1=2.uA uR/0c0. (70)
Because we also need pL < pA < pR, we are led to
pL pACpR
2< 0. (71)
However, the RHS is > 0, because uAuR < 0. Hence, it is not possible to maintain a nonoscilla-tory one point shock with the LRS scheme. It was not found possible to prove analytically anything
further about what shock structure the LRS scheme does support. In practice, we have found that
for transonic flow, the LRS scheme maintains a shock structure very close to a one point shock, and
any post-shock oscillations are not significant.
It is worth noting here that it may be possible to maintain the one point shock by choosing the
intermediate state to be something different from the arithmetic average. As a quick examination
of the proofs for the zero-point shock and contacts show, the arithmetic average is not by itself
necessary in order to maintain them. However, we have not yet found any simple way to define an
intermediate state that satisfies the one point shock criterion. We leave this, for now, as an avenue
for future work.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
11/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
5. EXTENSIONS TO HIGHER DIMENSIONS
The LRS scheme can easily be extended to higher dimensions with little additional cost if one
solves, as is usual with other approximate Riemann solvers, the one-dimensional problem normal
to the face. In three dimensions, the Eigen values are u, u, u, uC c and u c, and the characteristicequations in face normal coordinates are
D
Dt
1
c20
D
DtD 0, (72)
DCu
DtC
1
0c0
DCp
DtD 0, (73)
Du
Dt
1
0c0
Dp
DtD 0, (74)
Dv
DtD 0, (75)
DwDt
D 0, (76)
with DDt
, DC
Dt, D
Dtdefined as before. Notice that the equations are identical to the one-dimensional
case except for the two new decoupled equations. Hence, one can find the interface flux by simply
solving the one-dimensional problem with the additional equationsv w
D sgn.u0/
vL wL
. (77)
If one wishes to avoid the costs associated with local rotation to the face normal coordinates, the
standard methodology of using An DAnxCBnyCC n could also be used where .nx , ny , n/ arethe face normals. Then, it can be shown that
Winterface DWLCWR
2C W, (78)
where W can be found by using the following:
1. For supersonic flow, ju0j> c0' u v w p
TD sgn.u0/
L uL vL wL pL
T. (79)
2. For subsonic flow, ju0j< c0
qn DpL
0c0(80)
p D 0c0.nxuLC nyv
LC nwL/D 0c0.nxq
L/ (81)
D sgn.u0/
pL
pL
c20
C
p
c20(82)
u D sgn.u0/.uL.n2y C n
2/ .nxnyv
LC nxnwL//C nxqn (83)
v D sgn.u0/.nxnyuL vL.n2x C n
2/C nxnw
L/C nyqn (84)
w D sgn.u0/.nxnuL vLnynC w
L.n2x C n2y//C nqn. (85)
Finally, Finterface D F .Winterface/ as earlier.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
12/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
6. NUMERICAL RESULTS
The aforementioned scheme was implemented into structured and unstructured Euler solvers
for a number of cases and proved to be a robust and accurate solver. Some of the results are
presented herein.
6.1. One-dimensional results
The LRS scheme was used to simulate two classic one-dimensional test cases Sods shocktube
test and the Woodward and Colella 1-D blast case.
1. Sods test case: In this case, a shock was placed in the middle of a computational domain
extending between x D 5, 5. The initial condition was specified as follows:
WL D .L, uL, pL/D .1, 0, 1/,
WR D .R, uR, pR/D .0.125, 0, 0.1/.
The boundaries are fixed at initial values and Figure 1 plots vs. x at t D 1.8.2. Woodward Collela 1D blast case: First discussed by Woodward and Colella [14], this is a fairly
severe test case involving strong shocks and good test of any schemes robustness. The initial
conditions prescribed are
WL D .L, uL, pL/D .1, 0, 1000/
for the leftmost one-tenth of the computational domain,
WR D .R, uR, pR/D .1, 0, 100/
for the rightmost one-tenth of the computational domain, and
WMD .M, uM, pM/D .1, 0, 0.01/
in between the two extremes. Reflecting boundary conditions are prescribed on the walls of
the domain. Figure 2 displays the result when this case is implemented using 500 points in the
computational domain.
It can be seen that the current scheme performs extremely well on these cases. Note that we have
not used any tunable parameters in the two cases previously mentioned. It is important to mention
here, however, that like Roes scheme, LRS does require some additional tunable dissipation in cer-
tain cases which generate expansion shocks in the Roe scheme. However, as seen in the following
results, in transonic 2D and 3D cases, even this additional dissipation is typically unnecessary.
5 0 50
0.2
0.4
0.6
0.8
1
x
Sods Test Case (t=1.8)
Theoretical Solution
LRS
Figure 1. Sods shocktube test using linearized Riemann solver.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
13/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
Figure 2. 1-D blast case using linearized Riemann solver.
6.2. Two dimensional results
The LRS scheme was implemented into the following two transonic solvers
1. Flo82 : This is a structured solver, which along with the LU-SGS (Lower-Upper Symmetric
Gauss-Seidel) scheme [15] and multigrid leads to excellent convergence. Figure 3 shows the
pressure contours for the standard case of a National Advisory Committee for Aeronautics
(NACA) 0012 airfoil at a free stream Mach number M1 D 0.8 and angle of attack D 1.25with this solver.
Figure 3. Pressure contours for a NACA 0012 airfoil with M1 D 0.8 and D 1.25 with the linearizedRiemann solver scheme. The solver used was FLO82-SGS.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
14/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
Figure 4 compares the Cp distribution on the surface of the airfoil. The LRS schemecompares well with the Convective Upwind Split-Pressure (CUSP) scheme[13]. (CUSP has
been used here mainly to compare the convergence rate, which in our experience, has been
extremely good for CUSP. The final flow field is very similar for CUSP and Roe.). The struc-
ture of the shock also seems to be close to a one point, ideal shock structure. Figure 5 has the
convergence history for the two schemes. It can be seen that while the CUSP converges much
faster than the LRS scheme, the performance of the latter is still good. In particular, there is nostalling due to insufficient dissipation at the sonic point which can happen for schemes such
as the LRS where there is no dissipation at the sonic point. This is perhaps because of the
inherent extra dissipation that results due to upwinding normal to the face (instead of using
truly multidimensional upwinding).
2. Flo76 : This is an unstructured solver which uses a cell-centered triangular mesh along with
a multistage time-integration scheme integrated with multigrid in order to accelerate conver-
gence. Figure 6 shows the pressure contours for the aforementioned case with LRS. Figure 7
compares the Cp distribution for the LRS scheme with the BGK scheme for the same case.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 51.5
1
0.5
0
0.5
1
1.5
X
Cp
LRS scheme with SGS
LRSCUSP
Figure 4. Surface Cp distribution for the linearized Riemann solver and CUSP schemes for a NationalAdvisory Committee for Aeronautics 0012 airfoil with M1 D 0.8. The solver used was FLO82-SGS.
Figure 5. Convergence history for L1 norm of the density residual for the SGS scheme.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
15/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
X
Y
0 2 4 6 8
-2
0
2
4
Figure 6. Pressure contours for a National Advisory Committee for Aeronautics 0012 airfoil withM1 D 0.8 and D 1.25 with the linearized Riemann solver scheme. The solver used was FLO76.
Figure 7. Surface Cp distribution for the linearized Riemann solver and CUSP schemes for a NationalAdvisory Committee for Aeronautics 0012 airfoil with M1 D 0.8 and D 1.25. The solver used was
FLO76.
6.3. Three-dimensional results
The LRS scheme was also tested for a couple of three-dimensional configurations on FLO3XX
[16].
1. An ONERA M6 wing with M1 D 0.84 and D 3.06. The computation was done on 316, 000node tetrahedral mesh. The pressure contours on the surface wing are shown in Figure 8. The
surface Cp plots for the LRS and CUSP schemes are compared at the cross-sections x=c D 0.2and x=c D 0.4 in Figures 9 and 10.
2. A complete aircraft configuration : The Falcon business jet on a 356, 000 tetrahedral mesh.Density contours are plotted in Figure 11, and a comparison of the multigrid convergence
history with CUSP is given in Figure 12.
7. QUALITATIVE DIFFERENCES WITH THE ROE SCHEME
As can be seen from the results in the preceding section, LRS is a competitive approximate Riemann
solver, producing qualitatively similar results to Roes scheme on a number of numerical test
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
16/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
Figure 8. Pressure contours with the linearized Riemann solver for the ONERA M6 wing at M1 D 0.84and D 3.06.
1 2 3 4 5 6 7 8 9 10 111
0.5
0
0.5
1
1.5
X
Cp
Cp
distribution at 20% chord
LRSCUSP
Figure 9. Surface Cp distribution at x=c D 0.2 for the ONERA M6 wing at M1 D 0.84 and D 3.06.Comparison of the linearized Riemann solver and CUSP schemes.
3 4 5 6 7 8 9 10 11 121
0.5
0
0.5
1
1.5
X
Cp
Cp
distribution at 40% chord
LRSCUSP
Figure 10. Surface Cp distribution at x=c D 0.4 for the ONERA M6 wing at M1 D 0.84 and D 3.06.Comparison of the linearized Riemann solver and CUSP schemes.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
17/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
Figure 11. Falcon business jet. Nondimensional density contours from 0.6 to 1.1.
0 50 100 150 200 250 300 350 400 450 50010
10
109
108
107
106
105
104
103
102
101
100
Work
Log(Error)
Convergence history for the Falcon business jet
LRSCUSP
Figure 12. Convergence history of the density residual for the Falcon business jet. Comparison of thelinearized Riemann solver and CUSP schemes.
cases. Further, both Roe and LRS behave identically on stationary contacts, stationary shocks,
and both produce expansion shocks in the absence of additional dissipation. This, combined with
Equation (51), would lead one to expect that the LRS scheme would at best offer a cheaper alterna-
tive to Roes scheme, because the two are identical up to the first order dissipation term. A simple test
case reveals that the higher order terms can cause surprising qualitative differences between the two.
7.1. A shock stability test
The test case that reveals this difference is that of a planar stationary shock subject to two dimen-
sional perturbations. Roes scheme is known to be unstable for this case beyond a certain Mach
number. In fact, an analysis in [6] shows that for schemes of the form
F .WL, WR/D1
2.FLCFR/
1
2AD.WL, WR/.WL WR/, (86)
strict stability on this problem is incompatible with exact resolution of contact discontinuities. Roes
scheme, which is unstable to small perturbations, is of the aforementioned form, whereas LRS is
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
18/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
Figure 13. Roes scheme for shock subject to a two-dimensional perturbations with Ms D 5.0 , t D 3.3.The situation shown here is on its way to instability.
not. Because LRS and Roe are effectively equivalent in the first-order terms, one would expect thatLRS would also be unstable. We explore here if this is actually true in practice.
The problem being considered in the following is a simpler version of a problem proposed by
Quirk [5] as a diagnostic tool for the hypersonic carbuncle phenomenon. Dumbser et al. [17] have
done a linearized matrix stability analysis for this problem for Roes scheme. They used a uniform
25 25 grid with a normal steady shock as the initial unperturbed flow. The initial condition wasperturbed by a uniform random perturbation of relative order 106 102 in every cell.
Roes scheme exhibits three different ranges of behavior when subjected to perturbations
1. It was found to be stable for Mach numbers lower than 1.5.
2. It was found to be quasi stable for intermediate Mach numbers.
3. For high Mach numbers, the shock was unstable to mean flow perturbations, and for suffi-
ciently large Mach numbers, even round-off errors were found to be sufficient to set off the
instability. Figure 13 shows the solution for Ms D 5 on its way to instability.
The LRS scheme on the other hand, was found to be stable even for an upstream Mach number
of100 even with perturbations of relative order as high as 102 of mean flow (higher perturbationswere not tested). Figures 14 and 15 are representative solutions. The LRS scheme seems to be stable
for this problem.
7.1.1. Hypersonic blunt body. The aforementioned shock stability case is used as a diagnostic tool
for determining if a given scheme would exhibit the carbuncle phenomenon [7] for the hypersonic
blunt body test case. The carbuncle phenomenon is an incorrect but entropy satisfying solution to
a hypersonic flow past a blunt body. This occurs usually when the solution has nearly grid-aligned
shocks, computed with schemes which have low dissipation. The origins for this are unknown and
it is still a topic of active research (see [17] for a recent review ).
Roes scheme exhibits this phenomenon. The simplest fix for this phenomenon for Roes schemeis to add an empirically determined large dissipation to the linear vorticity mode via Hartens fix
[18] applied to the eigenvalue D u corresponding to the linear vorticity mode. Apart from beingphysically and mathematically unjustified, [5], this is a somewhat undesirable fix in that it might
affect computations of viscous shear layers adversely. Further, because of its empirical nature, the
amount of dissipation that needs to be added for a physically valid solution cannot be determined a
priori. This makes the fix unreliable for new test cases.
This high Mach number is used here only to indicate the robustness of the underlying Euler solver even though the Eulerand Navier-Stokes equations are no longer valid for this Mach number.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
19/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
Figure 14. Linearized Riemann solver scheme for shock subject to two-dimensional perturbations withMs D 2.0 , t D 10. The perturbations show no growth.
Figure 15. Linearized Riemann solver scheme for shock subject to two-dimensional perturbations withMs D 100.0 , t D 100. The shock is stable to perturbations.
Because the LRS scheme was stable for the planar shock, it was tested to see if it exhibited the car-
buncle. The test case chosen was one which was known to produce the carbuncle for Roes scheme.
A blunt body with the free stream Mach number M1 D 6 on a 160 64 grid was chosen.Figures 16 and 17 show the carbuncle phenomenon for this configuration with the use of Roes
scheme without Hartens fix on the D u mode (the fix was still applied to the other modes for con-vergence). As can be seen, the solution is not a detached shock as desired. It should be emphasized
here that this is a completely convergent solution, albeit the physically incorrect one.
The converged LRS results for the same problem (without any fix on any mode) are shown in
Figures 18 and 19. The solutions are indeed carbuncle free through a post shock overshoot that
is seen. This is probably because of the LRS not being able to support an ideal one-point shock
structure. It has been argued in the literature (see [8] for example) that numerical shock structure is
intimately tied to the carbuncle phenomenon.
Ismail et al. [9] suggest a one-dimensional carbuncle problem which tests a one-dimensional shock
with an artificially introduced intermediate state. They find that almost no scheme leads to an ulti-
mately stable shock structure for all possible intermediate states, but the better scheme performs on
this test, the better it performs on the carbuncle. Figure 20 shows LRS performance on this test case
for a MD 6 shock. The remarkable similarity of LRS behavior on this problem to its performanceon the hypersonic problem reinforces the belief that shock structure might be the key to finding the
reason behind the carbuncle phenomenon. An appropriate way to fix the overshoot, while keeping
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
20/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
X
Y
-2 -1 0 1 2
-2
0
2
4
Figure 16. Roe scheme for a hypersonic blunt body with M1 D 6. Convergent scheme exhibiting thecarbuncle solution. Isothermal contours are plotted.
2 1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 10.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
X
Cp
Figure 17. Cp along the centerline for the Roe scheme. Hypersonic blunt body with M1 D 6.
accuracy intact could not yet to be found. We did find, however, that using a Roe average instead of
an arithmetic average for the intermediate state does alleviate (but not eliminate) the problem.
Further insight into this overshoot can be gained by looking at the corresponding entropy plots.
Figures 22 and 24 show that the situation is seemingly reversed with entropy viz. The entropy in
CUSP shows an overshoot, whereas LRS shows a (nearly) smooth entropy profile. Figures 21 and
23 show entropy contours over the blunt body for LRS and CUSP respectively. The local entropy
overshoot, however, maybe interpreted as an approximation to the entropy profile for a viscous shock
profile; as is known [19], the viscous shock entropy profile is not monotonic and shows a maximum
within the shock. This then, might give a non ad hoc method for correcting the pressure overshoot
in LRS via adding an appropriate viscous term to Equation 35. We are currently investigating
this approach.
On the pessimistic end, it is entirely possible, the fact that LRS does not show that the carbuncle
might be intimately tied to the fact that it exhibits an overshoot. However, its excellent performance
on the other one-dimensional tests in Section 6.1 leads us to believe that this is probably not the case.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
21/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
Figure 18. Linearized Riemann solver scheme for a hypersonic blunt body with M1 D 6. Convergentscheme exhibiting no carbuncle. Isothermal contours are plotted.
2 1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 12.5
2
1.5
1
0.5
0
X
Cp
Figure 19. Cp across stagnation line using linearized Riemann solver scheme for a hypersonic blunt bodywith M1 D 6. Slight overshoot is observed across the shock.
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.51.4
1.2
1
0.8
0.6
0.4
0.2
0
0.2
x
p
1D "carbuncle" using LRS (M0= 6, = 0.7)
Figure 20. 1-D carbuncle using linearized Riemann solver at MD 6. Note the similarity with the behaviorfor hypersonic flow over a blunt body.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
22/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
Figure 21. Linearized Riemann solver scheme for a hypersonic blunt body with M1 D 6. Contours of
constant entropy are plotted.
X
S
-2 -1.9 -1.8 -1.7 -1.6 -1.5 -1.4 -1.3 -1.2 -1.1 -10
0.5
1
1.5
2
2.5
3
3.5
Figure 22. Entropy profile along the centerline for the linearized Riemann solver scheme (M1 D 6).
Figures 25 and 26 show that the CUSP scheme does not exhibit the carbuncle phenomenon either,
even when no additional dissipation is added to any mode.
Further, it gives an extremely clean and sharp shock and does not exhibit the overshoot that the
LRS scheme exhibits. The sharp shock resolution here is much better than the shock profiles typ-
ically obtained by the use of other schemes not exhibiting the carbuncle such as the HLL: Harten
Lax Van-Leer.
The LRS and CUSP schemes were also tested for freestream Mach numbers as high as M1 D 50and were found to be carbuncle free. The LRS, however, had problems in convergence for high
Mach numbers while the CUSP was very well behaved. This very robust behavior of CUSP for the
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
23/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
Figure 23. CUSP scheme for a hypersonic blunt body with M1 D 6. Contours of constant entropyare plotted.
X
S
-2 -1.9 -1.8 -1.7 -1.6 -1.5 -1.4 -1.3 -1.2 -1.1 -10
0.5
1
1.5
2
2.5
3
3.5
Figure 24. Entropy profile along the centerline for the CUSP scheme. (M1 D 6).
hypersonic case seems to have been unreported in the literature. We therefore think that, because of
its excellent performance on the hypersonic problem, CUSP offers one of the best practical schemes
for such flows. However, as with all Riemann solvers, CUSP does have its lacunae CUSP, unlike
LRS, does not resolve steady contacts exactly.
Finally, we note that LRS does have a pressure difference term in the mass flux, as can be easily
from Equations (87) and (88) and yet it does not, strictly speaking, exhibit the carbuncle phe-
nomenon. This therefore seems to show that Lious conjecture [20] that it is this pressure difference
in the mass flux that causes the carbuncle is probably wrong.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
24/25
B. SRINIVASAN, A. JAMESON AND S. KRISHNAMOORTHY
Figure 25. CUSP scheme for a hypersonic blunt body with M1 D 6. Convergent scheme exhibiting nocarbuncle. Isothermal contours are plotted.
2 1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 12
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
X
Cp
Figure 26. Cp along the centerline for the CUSP scheme. A clean one point shock is captured. Hypersonicblunt body with M1 D 6.
8. CONCLUSIONS
Using a simple linearized model for the propagation of information along the characteristics, an
upwinded-state scheme, LRS, was derived to compute interface fluxes for the Euler and Navier
Stokes equations. This scheme is significantly cheaper compared with the popular Roe scheme andperforms similarly to Roe on a number of one-dimensional and multidimensional problems. Most
interestingly, it seems to be closely related mathematically to Roe in that they are identical to first-
order terms; yet, it is stable on a shock stability problem where the Roe scheme exhibits instability.
The scheme also seems to violate Lious conjecture on the carbuncle problem despite an unphysical
overshoot while not exhibiting the carbuncle.
We foresee the following applications for LRS
1. As an approximate Riemann solver wherever Roe scheme is used. LRS offers a cheaper,
less complicated alternative. This includes, for instance, incompressible flows using pseudo-
compressibility.
Copyright 2011 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Fluids (2011)
DOI: 10.1002/fld
-
8/3/2019 ijnmf_srinivasan_2011
25/25
AN UPWINDED STATE APPROXIMATE RIEMANN SOLVER
2. As a solver for the hypersonic blunt body problem where Roe cannot be applied in an
unmodified form.
3. As a starting point for analysis of the reasons for the carbuncle and shock stability phe-
nomenon. The fact that the LRS and Roe schemes exhibit vastly different behaviors on the
same problem when they are closely related mathematically might offer a new way to analyze
this difficult problem.
Finally, as an ancillary result, the CUSP scheme was found to be a very efficient and robust solver
for the hypersonic blunt body problem.
ACKNOWLEDGEMENTS
This work was partially funded by a grant from the Aeronautics Research and Development Board(Aerodynamics), DRDO, Government of India. Balaji Srinivasan would like to thank Mr. Vivek Hariharanfor his help in generating some of the results.
REFERENCES
1. Godunov SK. A difference scheme for numerical computation of discontinuous solution of hyperbolic equation.
Sbornik: Mathematics 1959; 47:271306.
2. Roe PL. Approximate Riemann solvers, parameter vectors, and difference schemes. Journal of ComputationalPhysics 1981; 43(2):357372.
3. Mandal JC, Deshpande SM. Kinetic flux vector splitting for Euler equations. Computers and Fluids 1994;
23(2):447478.
4. Xu K. VKI report 1998-03, 1998.
5. Quirk JJ. A contribution to the great Riemann solver debate. International Journal for Numerical Methods in Fluids
1994; 18:555574.
6. Gressier J, Moschetta J-M. Robustness versus accuracy in shockwave computations. International Journal for
Numerical Methods in Fluids 2000; 33:313332.
7. Peery K, Imlay S. Blunt body flow simulations. No. 88-2924, AIAA Conference, 1988.
8. Chauvat Y, Moschetta J-M, Gressier J. Shock wave numerical structure and the carbuncle phenomenon. International
Journal for Numerical Methods in Fluids 2004; 00:16.
9. Ismail F, Roe P, Nishikawa H. A Proposed Cure to the Carbuncle Phenomenon, 2009.
10. Harten A, Lax PD, van Leer B. On upstream differencing and Godunov-type schemes for hyperbolic conservation
laws. SIAM Review 1983; 25:3561.
11. Cantwell B. Introduction to Symmetry Analysis. Cambridge University Press, 2002.12. Landau LD, Lifshitz EM. Fluid Mechanics: Volume 6 (Course of Theoretical Physics). Butterworth-Heinemann,
1987.
13. Jameson A. Analysis and design of numerical schemes for gas dynamics, 2: artificial diffusion and discrete shock
structure. International Journal of Computational Fluid Dynamics 1995; 5(1):129.
14. Woodward P, Colella P. The numerical simulation of two-dimensional fluid flow with strong shocks. Journal of
Computational Physics 1984; 54:115173.
15. Jameson A, Caughey DA. How many steps are required to solve the euler equations of steady compressible flow:
in search of a fast solution algorithm. 15th AIAA Computational Fluid Dynamics Conference, June 11-14 2001;
2001-2673:20012673.
16. May G, Jameson A. AIAA Paper 2005-0318, 2005.
17. Dumbser M, Moschetta J-M, Gressier J. A matrix stability analysis of the carbuncle phenomenon. Journal of
Computational Physics 2004; 197:647670.
18. Harten A. High resolution schemes for hyperbolic conservation laws. Journal of Computational Physics 1983;
49(3):357393.
19. Morduchow M, Libby P. On a complete solution of the one-dimensional flow equations of a viscous, heat conducting,compressible gas. Journal of the Aeronautical Sciences 1949; 16:674684.
20. Liou MS. Probing numerical fluxes: mass flux, positivity and entropy satisfying property. AIAA Paper 97-2035, 1997.