You are on page 1of 26

On Godunov-Type Methods for Gas Dynamics Author(s): Bernd Einfeldt Reviewed work(s): Source: SIAM Journal on Numerical Analysis,

Vol. 25, No. 2 (Apr., 1988), pp. 294-318 Published by: Society for Industrial and Applied Mathematics Stable URL: http://www.jstor.org/stable/2157317 . Accessed: 14/07/2012 13:17
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize, preserve and extend access to SIAM Journal on Numerical Analysis.

http://www.jstor.org

SIAM J. NUMER. ANAL.


Vol. 25, No. 2, April 1988

1988 SocietyforIndustrial and Applied Mathematics


004

ON GODUNOV-TYPE METHODS

FOR GAS DYNAMICS*

BERND EINFELDTt Abstract.In thispaper we describea new approximateRiemannsolverforcompressible flow.In gas to for contrast previousRiemannsolvers,wherea numericalapproximation the pressureand the velocity at the contactdiscontinuity computed,we derivea numerical is approximation the largestand smallest for in we signalvelocity the Riemannproblem.Havingobtainedthenumerical signalvelocities, use theoretical results Harten,Lax and van Leer to obtainthe fullapproximation. by a A stability conditionfor the numericalsignal velocitiesis derived.We also demonstrate relation betweenthe signal velocitiesand the dissipationcontainedin the corresponding method. Godunov-type of The computation signal velocitiesfora general(convex) equation of stateis discussed. Numerical resultsforthe one- and two-dimensional compressible dynamicsequations are also given. gas Key words. Riemann solver, Godunov-type methods,hyperbolicconservationlaws, gas dynamics, equation of state AMS(MOS) subjectclassifications. 65M05, 65M10, 76M05

1. Introduction. thelastseveralyearsGodunov-type In methods havebeen applied for flow. practice, In thesemethods successfully thecalculationofinviscid compressible withvery are characterized their by robustness their and usefulness computing in flows complicatedshock structures. Godunov [7] used the nonlinearRiemannproblemas a "buildingblock" forhis of treatment weak and strongshock numericalmethod.This allows a self-operating shocks nearlyoptimally waves. The numericalsolutionrepresents thin; a monotone is cells without oscillations typical. profile overone to three computational nonphysical of Fromthetheoreticalpointof view,Godunov's methodis an extension theclassical Courant-Isaacson-Rees scheme [10]. The underlying physicalpictureof Godunov's of of methodis usefulforthe interpretation certainschemesand in construction new ones. The recentinterest Godunov-type in methodswas engendered van Leer, who by realizedtheimportance Godunov's methodand invented second-order of a extension. Further developments along thislineweremade byColella and Woodward[1], Colella and Winkler of [2] and Woodward,Colella, Fryxell [25]. A comparison some of these methodswithmoreclassical methodscan be foundin [26]. Godunov-type The disadvantageof Godunov's method and its higher-order extensionis the of difficulty solvingthe nonlinearRiemannproblemexactly, especiallyformaterials withcomplexequationsof state.The exact solutionof the Riemannproblemrequires an iterative whichleads to relatively numerical procedure, complexand time-consuming codes. Since computationalefficiency a major requirement applied numerical is for the methods, thishas restricted extensive applicationsof Godunov-type methods. To overcome have thisdrawback, severalapproximations theRiemannproblem to been developed.For ideal (polytropic) equationsof statethereare,by now,particular approximate"Riemann solvers" available; among themare the methodsdeveloped by Osherand Solomon [15] and Roe [18] and Pandolfi[17]. These linearapproximain tionsare also of interest thefieldof aerodynamics wheretheyprovidea foundation
* Receivedby theeditors is January 1987; accepted forpublication 5, April14, 1987.This research part 27 of a joint task conducted by the Sonderforschungsbereich "Wellen fokussierung" supportedby the Deutsche Forschungsgemeinschaft. und Praktische Mathematik RWTH Aachen,Templegraben 5100Aachen, der fiir t Institut Geometrie 55, West Germany. 294

ON GODUNOV-TYPE

METHODS

295

of forthe construction more elaborate schemes [19], [20]. More analyticaleffort is requiredif a generalequation of stateis considered.So far,onlytwo Riemannsolvers have been developedin thiscase. One iterative methodis by Colella and Glaz [3] and a second explicitmethodis by Dukowicz [5]. It is interesting note thatDukowicz to a derived Riemannsolverbyestablishing relationship theartificial his to shockviscosity methods. In this paper we describea new approximateRiemannsolverforcompressible gas flow.In contrast previousRiemannsolvers,wherea numericalapproximation to and forpressure velocity thecontact at is we discontinuity computed, derivea numerical for approximation the largestand smallestsignal velocityin the Riemannproblem. Having obtainedthe numericalsignalvelocitieswe use theoretical resultsby Harten, Lax and van Leer [10] to obtain the full approximation. This new Riemann solver keeps the computing time extremely low, while retaining essentialproperties the of Godunov's method and provides,therefore, attractive an alternative which to on construct extensions. higher-order The organization thispaperis as follows:In ?? 2 and 3 we describe theoretical of the foundations the Riemannsolver.The stability dissipation the corresponding of and of Godunov-typemethodare discussed in ? 4. In ? 5 we show how to obtain simple for numerical approximations the signalvelocities.Numericalresults theone- and for two-dimensional compressiblegas dynamicsequations are given in ? 6. Section 7 containsour conclusion. A comparison withGodunov's method[7] and Roe's scheme[18] is also contained in ?? 4 and 6. This makes thisarticle, a certainextent, reviewpaper. to a 2. Preliminaries. A. Conservation equation. We consider the (Euler) equations for an inviscid flow.The conservation formof these equations in one Cartesianspace compressible variable is
(2.1)
a-+

at

a &9x

0,

wherethe conservedquantitiesv and fluxes are: f p


(2.2) =

~~m
f(V) m2/P+P

m
e

m/p(e +p)

m Here, p is the density, = pu is the momentum unitlengthand e = pE +?pU iS per the total energy unit length.The physicalvariables u and E are the velocityand per theinternal energy unitmass. The pressure is relatedto the conserved per p quantities the through equation of state
(2.3) p= p7(,
?)

volume.In the case of an ideal equation of statewe have wherer= I/p is the specific
(2.4) p = (y-1)-.

In thispaper we concentrate an ideal equationof stateand indicatea possible on extensionto general equations of state. General convex equations of state will be consideredin a future paper.

296

BERND EINFELDT

B. The Riemannproblem.The Riemannproblemis an initialvalue problemfor (2.1) withthe initialdata V v, forx<O0 forx > 0. tivr

[4] A detaileddiscussionof the Riemannproblemis givenby Courantand Friedrichs and in [22], [6] foran ideal equation of state. The solutionto the Riemannproblemconsistsof fourconstantstatesseparated a and forward facing waveis either rarefaction waves.The backward elementary bythree (Fig. 1). The exact or shock wave. The centerwave is always a contactdiscontinuity method described in [3]. The solution to the solution is obtained by an iterative about the physical incorporatesa large amount of information Riemann problem and strength forexample,the directions thatis modeled by (2.1). It includes, system states. of the emergefrom nonlinearinteraction two constant of the disturbances that For thatreason,the solutionof the Riemannproblemis a natural"buildingblock" schemes. differencing of forthe construction upstream of C. Godunov'smethod.We dividethe timeinto intervals lengthr and let l\be The solutionis to be evaluated at time t' = nr,where n is a the spatial increment. .Let v' approxiinteger thespatialincrements = il\,i = 0, ?1, ?2, at nonnegative x, matethe cell average

-f ~2V(X
+1/2

x1

tn)

dx

x~1/2

(i + 1/2)1\. as is In Godunov'supstream differencing scheme[7], thefluid described a sequence of cell-averagedconservedquantitiesp', mn, ei. The cell averages are advanced in at The a solving Riemannproblem each cell interface. averagesvi= n1=( (pin+1 timebyfirst at the nexttimelevel tn? = n + r are obtainedby averagingover a cell mi ein+) ((i-/2)\,(+ 1/2)1\); i.e.,
where Xi+1/2 =

(2.5)
interface x,+1/2=

v,=

whereci+1l/2(x-xi+l/2/(t(2.6)

(i + 1/2)1\at time tn (Fig. 2). Equation (2.5) can be rewrittenas


Vn+ =

dx +

0))

is the solution to -1/2(X/T)

to

dx

theW+l/2(X/T) Riemann problemat the cell

i+112-

i-112)

t backward facing wave

contact di sconti nui ty / forward facing wave

in FIG., 1. Thesolution the space. problem physical of Riemann

ON GODUNOV-TYPE

METHODS

297

Vit1

n
i

~~~~~I
(i-1/2)A (i+1/2)A

v2n
+1 t

FIG. 2. Godunov's method physical in space.

wherethe numericalflux
(2.7)

g'?1/2
gn?1/2

is givenby
ng(V, v+?) =f(ca)?l/2(0)).

More details can be foundin the paper by Harten,Lax and van Leer [10]. 3. The Riemannsolver.The solution 0i+Il2((x-xi+l//2)(tproblem v,+f(v)> = 0,
(3.1)( 0) = { 0) V(X' vj+I Cn)) of the Riemann

forx < Xi+


forx> Xi+/2

at thecell interface Xi+l/2,depends onlyon thestatevi,vi+land theratio (x-xi+l/2)/t. n. we For convenience, assume that tn=0 and omitthe subscript Furthermore,

(3.2)

(ix'

vj

for ai+1/2t, x'<

and ai+1/2, ai+1/2 are the smallest and largestphysical signal where x'=x-XiX1/2 discussed by Harten, The HLL-Riemann solver,theoretically velocity,respectively. Lax and van Leer [10], extracts information about the signal velocitiesfromthe the fullRiemannproblem.It consistsof threeconstantstates,i.e., vi (3.3)
(oi+l/2(X/

t) =

Vi+1/2

I Vi+
t bill/2

forx' < b'i+112t, for~~~~~~~~~ forb'+l/2t < x' < < forb + ,/2t x'.

b1 1/2t,

bi+1/2 vi+1/2

I~~~~~
i+1 x

solution(3.3) of theRiemannproblem. FIG. 3. Thie approximate

298

BERND

EINFELDT

approximations the smallestand to are Here, bi+1/2 and bl+1/2 Lipschitz-continuous largestphysicalsignalvelocities(Fig. 3). The averagestate Vi+1/2 is definedsuch that law, i.e., formof the conservation withthe integral the Riemannsolveris consistent

(3.4)

+A/2

-J/2

= w?i+,/2(x'/t) dx'

+ A/2(vi

Vi+?) - Tf(vi+?) + Tf(Vi),

> Riemannsolver Having computedthe approximate for1\/2 r max {lbi'+121, I b?i'/21}. to we at thecell interfaces, obtainthecell averagesat thenexttimelevel similar (2.5):
(3.5) VI = 1/1\ {
2 A/

W14/2(X
0

/r)

dx'? 1/0

.0.-A/2

Wi/2(X /T) dx'.

in form This again can be rewritten the conservation

(3.6)
gi+l/2(vi, vj+?)=f(vi)

v = v?A (go

1/2-g0-1/2)9

withthe numericalfluxfunction for0<bWi+l/2


- bi+/2f(Vi+I) bi+1/2bi+l/l

b1?l/2f(v,)

(V

-V )

(3.7)

- bi+l/ i12i bi+112

? b

-1bbl/2 +1

forbl/2 < 0 < br/2 =f(vi+?) forb+


1/2<

0.

This again can be combinedinto a singleformula: (3.7')


g?12

=b,+ 12f(Vi)-bb?l/2f(vi+l)?
?/2i+

i+b1/2

?+1/2bT?1/2( i+1/2 i+12

velocities bounds the for physical signal a'i+112,

(3.3) of the = where b+l/2 = max (0, br+/2) and bi- l/2 min(0, bl+1/2). The definition to from one givenin [10]. In contrast the originalversion, the Riemannsolverdiffers are bi?1/2 lowerand upper signalvelocitiesb'+1/2, we do notassumethatthenumerical signalvelocitieswill be derivedin the nextsection. Suppose thatthe solutionof the Riemannproblemconsistsof a singleshock. In thiscase the exact solutionis
(3.8)

a i+ 1/2 Stability for numerical bounds the

u(x, t) =V

vi+I

forx'< st, for st <x',

that of wheres is thespeed ofthepropagation theshock.Suppose further thealgorithm forcalculatingthe numericalsignalvelocitiesis such thatit furnishes s fora backwardfacing shock, bl= (3.9a) shock. facing bh?112 s fora forward (3.9b) the It thenfollowsfrom equalityof Vi+112, withthe mean value of the exact solution, that(3.3) is the exact solution(3.8). The Let vi and vi+lbe connectedby a singlecontactdiscontinuity. exact solution velocityui= uj+I= u. is again (3.8) withthe shock speed replacedby the characteristic In general,

(3.10)

bi+112 < u

<

bi+112.

rough and Thus,the Riemannsolver(3.3) spreadsthe contactdiscontinuity is a rather to approximation the exact solution.

ON GODUNOV-TYPE

METHODS

299

Next we show how to modify Riemannsolver(3.3) to obtaina moreaccurate the approximation a contactdiscontinuity. of The information the contactdiscontinuity buriedin the averagestateVi+1/2. of is A contactwave is a weak solutionof (2.1) of the form (3.11) v(x, t) = cW(x, t)(1, u,2u ) whereu is the constant of recover velocity propagationof the wave. We therefore the information about the contactdiscontinuity, modifying average state Vi+,/2 in by the (3.3) by

(3.12)
Here (3.13a) and

x'/ vi for t<


W?i+1/2 = Vi+l/2+ +i+l/2(X

bi+112,
-

Vi+1 for bl+1/2<


R

i+?1/2t)7 1+1/2Ri+l/2

forb1+12 < x'/t <b

X'/ t.

1+1/2= 2

(I, 9 1

i+112

2U i+112)

(3.13b)
with
(3.13c)
li1/2 = (l-

N7+1/2=

i+112(Vi+l-

Vi)

21 2

2 ji+112

-1)

_2?,/-2 Ci+l/2 ~C1l/2

whichwill be specifiedin the nextsection.Observethat Ril+1/2 8i+1/2 is a parameter is the second eigenvector theJacobianmatrix and that7i+1/2 is an approximate of df value for the projectionfrom vi+1 vi onto this eigenvector.Ui+1/2 and Ci+1/2 are of numericalapproximations the velocityand the sound speed at the contactdisconWe tinuity. define
(3.14a) (3.14a)

+ br+1/2 + ~ ~~~~~~~ 2 bi+1/2 Ui+/2 2

The approximatesolution (3.12) is thus completely definedby the numericalsignal of velocities.The modification the average state Vi+1/2 does not change the integral in the (3.4). Therefore Riemannsolver(3.12) remains conservation form. The numerical fluxof the corresponding methodis defined Gundov-type by
hi+?12 =f(vi)
1 'o
--

W i+l/2(X'/17)

dx' +-vi

(3.15)
i+= gi=12-

'J -A/2

b+ 1 1/b-

2Ri+1/2q 11224i+1/2i+2/2

wheregi+1/2 is the numericalflux(3.7). 4. Dissipationand stability. The resolutionof a shock wave or contactdisconis scheme.In tinuity largelydetermined the numerical by dissipationof the difference for the thissectionwe willstudy dissipation theGodunov-type methods described here. The numerical fluxfunction in (3.7) may also be written the form

(4.1)

gi+112 = 2[f(V

) +f(Vi+l)-Qi+112(Vi+l-VA,

300

BERND

EINFELDT

wherethe numerical viscosity-matrix by Qil/2 is defined (4.2)


Qi+1-2

i+112 bi+l i+112

Ai -/22br -+b'i+12
i+1/2

i+1/2

Here the matrixA?+/2 is a Roe-typelinearization whichhas real eigenvaluesai+l/2, a completeset of eigenvectors and satisfies (4.3) f(vi+?)-f(vi) =
A?+l/2(vi+l
-

vi).

The existenceof a Roe-typelinearization follows fromthe existenceof an entropy for function (2.1) (see [10]). A necessary is condition stability that viscosity for the matrix Qi+/2 has nonnegaThe tiveeigenvalues. eigenvalues Qi+?12 are of
?k

= b?l1/2(a 1+/2- bi1/2)


i+ 1/2

bi+1/2(b?l+/2-ai+/2)
i+ 1/2

(k 1, .,3) ' where a+l/2 denotesthe eigenvalueof Aj+1/2. Thus, we obtain a necessarystability conditionforthe signalvelocities

(4.5) ) (
whichis satisfied if (4.6)
b7+l/2<a +l/2

(J~~~~~~~~~~i+1/2-?

ba?k,b

2 fork= 1,**, 3.

If the stronginequalityholds in (4.6), then the scheme (3.6) is dissipativeof order two. Since g+1/2(v,v) =f(v), the schemeis accurateof orderone and fora linearflux function L2-stability followsfrom Kreiss theorem the [23, Thm. 3.3]. f, A stronger matrixQi+1/2is TVD-typeconditionforthe viscosity

(4-7)

i+112-1a +

/21

(k = 1,*

3).

For linearf or scalar nonlinear thisrequirement f impliesthata scheme(3.6) in conservation formis totalvariationdiminishing (TVD), underthe CFL restriction

(4.8)

<1 Ao-j+1/2

schemesare monotonicity whichguaran(see Harten[8]). TVD difference preserving, teesthattheschemedoes notgenerate A spuriousoscillations. shortcalculationshows thatthe stronger condition(4.7) holds if and only if the signal velocitiessatisfy the inequalities(4.6). Suppose thatthe solutionof the Riemannproblemconsistsof a singleshock. In the Rankinethis case the exact solution is (3.8), where the shock speed satisfies Hugoniotcondition (4.9) f(vi) -f(v?+1) = s(v, -v'+).

From (4.3) and (4.9) it followsthatthe shock speed s is an eigenvalueai+1/2 (k=1 for or 3) of the Roe linearization Aj+1/2. Therefore, a backwardfacingshockthe only dissipationis introducedthroughthe first eigenvalue 0J1+1/2 of the viscositymatrix
Qi+?12*

We have

(4.10)

OK1+1/2

Sl,

ON GODUNOV-TYPE

METHODS

301

where the equality sign holds for bi11/2 =s. This shows that the least amount of in numerical dissipationis introduced the method(3.6), (3.7) if
i+112= S =
+1/2-

A similarstatement holds fora forward facingshock. In the case of a stationary shock,we may have thatO- +1/2 = Isl = 0 fork = 1 or 3 and the numerical termin (4.1) vanishes.The lack of numericaldissipation viscosity allows the perfectresolutionof stationary shocks, i.e., (3.8) with s = 0 is a steady scheme.On theother solutionofthenumerical hand,theschememayadmita stationary To the nonphysical discontinuity. avoid thiswe maymodify numerical signalspeeds by (4.11a) (4.1lb)
b+1/2=

min(bi+1/2, A1),
(bl+1/2, A),

b?i /2=max

and whereA1= j - ci and A3= ui+I + ci+I are thefirst third eigenvaluesof theJacobian Since the entropy inequality df(vi), df(vi+l),respectively. (k = 1, 3) the of holds [22], we observethatthe modification (4.11) does not effect resolution a the shock wave. For a nonphysicaldiscontinuity inequalityin (4.12) is reversed.In this case the signalvelocities(4.11) introduceenough numericaldissipationto avoid the occurrenceof a nonadmissible discontinuity. The Now suppose thatv, and v,1 are connectedby a contactdiscontinuity. exact solution is again (3.8), withthe shock speed s replaced by the characteristic speed (4.12)
Ak(Ur)

< S < Ak(Ul)

U = Ul= Ur. From (4.3) and (4.9) it follows that u is equal to the second eigenvalue

matrix Qi+1/2- Since theviscosity

ai+1/2

the of Aj+I/2. The only dissipationis introduced through second eigenvalueof

(4.13a)
(4.13b)

_ bi+1/2 a +<112a+1/2,
ati2+112< ai+1/2 i+1/2

we have
0.4i+1=

l 1a2l/2

2fobb7
=

i++/2bi+1/2

=,

(4Ji+1/2>

la1l/2l

otherwise.

the contactdiscontinuity is Therefore, numericaldissipationfor a nearlystationary largerthan it is forGodunov's method. Now considerthe modified numerical fluxfunction (3.15). If we assume that (4.15)
-

Ui/2

bl?1/2?bi?1/2 = i 2 ai+112

and represent - vi in termsof the right of eigenvectors the Roe linearization, i.e., v,+1 if (4.16)
v+ vi =
k=I

i+1l/2Ri+l/2,

thenthe modified numerical matrix becomes viscosity (4.17)


Qi12-bi+

ii+li =1+2

b-1/2 +bl+ b+ t
12-

b-+,

i+ 1/2

2 Ai+1/2-

bl+112b i+
bi+ 12-

i+(2

b-+/

i_

i+2

-~~~~~~~~2
7i+1/2/

302 where (4.18)

BERND

EINFELDT

Bi+1/2= T+l/2 0

1 i012

= with T?l/2 (R!+1/2, R ?1/2, R ?1/2). If the parameter53i+1/2 is positive,then the last termon the rightof (4.17) is negativeand can be considered as an anti-diffusion term,which steepens contact condition(4.7) is valid discontinuities. 3i+1/2 should be chosen such thatthe stability fork= 2. If the numerical signalvelocitiesare the smallestand largesteigenvaluesof the Roe linearization, if i.e., = (4.19a) b 1/2 min(a 1+ /2, 0),

(4.19b)

bt+12=max

(a?+1/2,

0),
?li+1/2

thentheprojection(3.13) is exact and therefore Thi+1/2 =


(4.20)
8i+1/2 'r ji+1/2+lui+1/21'

In thiscase we choose

of Substitution (4.20) into (4.17) gives


3

(4.21)

Qi+11/2(vi+1 vi)

k=l

ai+1l/27

i+l/2Ri+l/2,

wherewe have used (4.15), (3.14) and 77i+1/2= '7i+1/2. The numericalflux(4.1), (4.21) is the same as thatof Roe's method[18]. Thus method(3.7) withan antidiffusion as Roe's schemecan be considered a Godunov-type in the lineardegenerate field. term We see that the dissipation of Roe's method (4.1), (4.21), the Godunov-type method (4.1), (4.2) (withthe signal velocities(4.19)) and Godunov's methodis the same across a singleshock wave. We therefore expectthatthesemethodswill show a the For similar for resolution shockdiscontinuities. a contactdiscontinuity dissipation of the methods of Roe and Godunov is the same, whereas the dissipationof the Godunov-typemethod (4.1), (4.2) is larger. The Godunov-typemethod therefore over moregridpoints. spreads a contactdiscontinuity Remark.Rememberthat Roe's method (4.1), (4.21) with the signal velocities as (4.19) can be rewritten
h1+112

b+1l/2f(v,) -bL+

I+ 112f(Vi+l)

(4.22)
* +

-i+

+l/2 - b-+/
(V1?1
Ci?1/2?j2l/j ji+112 + Ilui+1121 +1/2/

b++l,2b,+l,2 bi+l

whereCi+1/2,

used fluxfunction(4.1), (4.21) this formof the numericalfluxdoes not commonly and the decomposition(4.16). of requirethe full computation the right eigenvectors

Ui+1/2, 77i+1/2,

by Ri+112 are defined (3.13), (3.14). In

to contrast the

of 5. Computation thesignalvelocities.Harten,Lax and van Leer [10] leave open and for the question of how to computeapproximations the signal velocitiesb?+1/2 for bl+1/2.We now describetwo algorithms this essentialpart of the Godunov-type of method.In the last sectionwe saw thatthe mostaccurateresolution shockdiscontinuitiesare obtained if we choose the smallest and largesteigenvalues of a Roe

ON GODUNOV-TYPE

METHODS

303

linearization Ai,1/2 forthe signal velocities.For a y-law (ideal) equation of state a by special matrix Ai+1/2was constructed Roe [18]. The minimaland maximaleigenare values of thismatrix
(5.1a)
b1+1/2=
u1?i/2 ci?l/2,

(5.1b)

= b U+1?2 ui+1/2 + Ci+1/2,

and Hi+1/2 are the averagedvelocity with4+1/2 = (y - 1)(H1+1/2 2- +1/2). Here, Ui+1/2 by defined and averagedtotalenthalpy,

(5.2) (5.3)

Cii+1/2N- ,
-

/ + p-PHi+ I PHi
-l-

P?i

Hi+l/2

_ 4p Hi + v'p_

IHi+,

with piHi= ei +pi. For a more generalequation of state (at present)a Roe average and may also lead to a more complex matrixAi+1/2is not available in the literature algorithm. Next we indicatehow the Roe signal velocities(5.1) may possiblybe extended to moregeneralequations of state. as To do thiswe note that 4+1/2 in (5.1) can be rewritten

(5.4) (5 4)
with (5.5)

cl-2+/= C,1+/2-

V4c' '-P

+ 1+1I _'1+ NP

(Pi, Pi+l)(Ul+l

_Ui)2,

7y(Pi,Pi+I) =

(+ P??

P)2'

For most gases, the ratio of the specificheats y is a constantbetween 1 and 5/3. Therefore,

(5.6)

<T 72()Pi, 17T(pip?i+c)

9i+l) =

+ 2 (4 2 (v"F + p?)2

and we may approximate(5.4) by

(5.7) (5.7)
(5.8a) (5.8b)

d~~~+1/2- NPiPi+ +1 V/PC


bi+1/2=

Il~-i-

Pi?i vF7

2i?112 +

772(Pi P +)(Ui+

_Ui2

The signalvelocities
ui+1/2 -

di+1/2,

bl+1/2 = ui+1/2 + di+1/2

are also well definedfora generalconvex equation of state,since the sound speed ci from can be computeddirectly (5.9) Since (5.10)
di+1/2> c+ 1/2,

c2 =2( (PPe

) (P=

l/p).

(4.6) forany y-lawgas. We cannot requirement the thesesignalspeeds satisfy stability questionfora moregeneralequation of state. yetanswerthe corresponding

304

BERND

EINFELDT

Remark.Lax proved [22, Thm. 17.16] thatforweak shocks,the shock speed s is in by (up well approximated to second orderterms the shockstrength) the averageof both sides. This holds whichimpingeon the shock from speeds thetwo characteristic fora generalconvex equation of state. A shortcalculationshows that (5.1la) (5.1lb) + ui+1/2 2(Ui Ui+1),
di+112 2(Ci+ Ci+ )

in terms the shock strength. up to second-order weak shocks,thephysicalsignalvelocitiesare well approxiThus, forsufficiently matedby the numericalsignalvelocities(5.8), (5.7), (5.2), even fora generalconvex equation of state. 6. Numericalresults.We have used the numericalmethod (3.6), (3.7) together to withthe Roe-signalvelocities(5.1) and the signalvelocities(5.8). We will refer the first versionas theHLLR-methodand denotethesecondversionas theHELL-method. was used by Sod numericalexperiment The first problems. A. One-dimensional schemes.This test of [24] to compare a variety new and establishedfinite-difference laws (2.2). The initial problemforthe conservation problemis a standardshock-tube conditionsare x<0.5, (6.1) PI = 1.0
U= 0.0,

0.5<x,
Pr= 0.125
Ur= 0.0,

p,=

1.0,

Pr = 0.1.

In the calculationwe have used 100 cells withA = 0.01 and CFL 0.8. The numerical solutionis evaluatedafter50 timesteps. Figure4 showsthe resultsof the HLLR scheme.The cornersat the endpointsof statebetweenthecontactdiscontinuity wave are rounded.The constant therarefaction is and the shock has been realized. The contactdiscontinuity spread over sixteento of eighteen zones, whilethe transition the shock occupies onlytwo to fourzones. For comparison,Godunov's method is shown in Fig. 5. To compute the full methoddescribedin [6]. The plot is almost Riemannproblemwe used the iteration less wave is slightly cornerof the rarefaction equal to that of Fig. 4. Only the right less is roundedand the contactdiscontinuity slightly smeared. the Figure6 represents resultsof the Roe scheme.The plot is the same as in Fig. morerounded. wave is slightly cornerof the rarefaction 5, exceptthatthe right The resultsforthe HELL-method(not shown) are identicalto the plot of Fig. 4. The second testproblemwas used by Lax in [11]. The initialconditionsforthis Riemannproblemare x<0.5,
PI -0.445, (6.2)

0.5<x,
Pr= O..5
Ur = 0.0

ul =0.698,
pl-=3.528,

Pr = 0.571.

in withthisproblemare reported [9]. Othernumericalexperiments

ON GODUNOV-TYPE
HLLR - METHOD Density r

METHODS

305

rmx -

0.0000

HLLR -METHOD

prssr

vIeloctyu

\0

pmx pmn\o.10

.00

Fi.4

LRshmefrSdspolm

umax usmi

0.0000 0.10000

FI.4

HeL MEHLLOceeD oo'spolm

306

BERND
GOOUNOV - METHOD Density r

EINFELDT

>

~~~~~~~~~~
--\occ

rmax rmin

i 0000 O 1250

GOOUNOV - METHOD ve loci1ty u

g/D7-fi6o

l~~~~

/0

Gt
D:

umax
umin

=
=

0 9374
0 0000 O

GOOUNOV -METHOO
lpressure

'

X~~~~~

pmax
I

=
=

i ????
O iOOO

i
?pa

~~~~~~~pmin

FIG. 5. Godunov'sscheme for Sod's problem.

ON GODUNOV-TYPE
ROE - METHOD

METHODS

307

neocSi ty

uma umn

= =

0.32 000

..00

pnax

1.0000 O. Z50

pmnin =

FI.6Io'

ceefrSdspolm

308
HLLR Oensity r

BERND

EINFELDT

-METHOD

00o

00

rmax rmin

i O

2809 3425

HLLR ve locilty u

-METHOO

00

~~ ~ ~

ul

O00

pmax umin

= -

3 5455 570 0

FG7.TeHLLR

METHD

orLx'

poles

ON GODUNOV-TYPE

METHODS

309

GODUNOV -METHOD Dens


ityr

00O0 o 0o

rmwax = n rmni =

1.2849

0 .3427

GODUNOV -METHOD
velocity uJ

umax
TiTIn

=
=

i .5435|
-0 0000 0

GODUNOV - METHOD pressure p

o~~~~~~~~~~~~~~~~~~~~~

prnax in Om

= -

3.5243 0. 5710O

FIG. 8. Godunov's scheme for Lax 's problem.

310

BERND

EINFELDT

ROE - METHOO Density r

= 000

03430~~~~--P,0

rmax

1.2850

rmin i 3-50.430

RaEE velocity u

.5THOD -

-do

oo~~~~~~~~~~~~~~~
umax umin 1.5411 0.0000 o

ROE pressure p

THW

pmax pnmin

3.5243 0,57i0

FIG. 9. Roe's scheme for Lax's problem.

ON GODUNOV-TYPE

METHODS31

If)

No XC~ ._,u

0?

?o

?s CD ?ou

c oa}

N, 0

i)

I~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

312

BERND

EINFELDT

HLLR - METHOD Density r

FIG.> rarefaction wavewithsoni .rm =.1519

12

rmax

1.2050

HeLRr p ressure p

METHOD

00

00

<

= 1.CO ~~~~~pmax prnin O0.5969

FIG. 12 FIGS. 10-12. Godunov's, Roe's and the HLLR rarefaction wave with sonic point.

schemewiththe modified signal velocities (4.11) for a

In Figs. 7 and 8 we show theresults applying HLLR-methodand Godunov of the difference occursin the resolution the of schemeto thisproblem.The onlysignificant The HLLR-method requiresabout threemore points for the contactdiscontinuity. transition. the Figure 9 represents resultfor Roe's scheme. The plot is nearlyidenticalto thatof the Godunov scheme. Next we test for the resolutionof an entropyviolatingstationary shock by a the where statesv,and Vr connected a rarefaction are considering Riemannproblem by wave. Fixingthe state vr by Pr=1.205, UrO=.O, pr=10 and C=(3-y)/(y+1)cr, we determine state v, by the
Pi (6.3) Pr,
(Cr + C1),

U1
P1

P1 C2/y

ON GODUNOV-TYPE

METHODS

313

FIG. 13. Second orderextension Godunov'smethod theshock of problem. focusing for

The calculationsin Figs. 10-12 were performed and 100 cells with80 timesteps a for CFL number 0.8. Godunov's scheme(Fig. 10) breaksup theinitial of discontinuity. of The occurrence the"kink" in thesolutionis well known.Figure11 showsthatRoe's schemeadmitsan entropy shock.The same entropy violation occurs violating stationary forthe HLLR scheme. In Fig. 12 we presentthe resultsforthe HLLR scheme,with the modifiedsignal velocities(4.11). The plot is now nearlyidenticalto thatof the Godunov scheme.The "kink" in the solutionis even smaller.The modification (4.11) does not alter the numericalresultsforthe two precedingtest problems.A similar can fluxin theform modification be used forRoe's scheme,withthenumerical (4.22). We remark thatthe occurrence a nonphysical of is discontinuity onlypossible if thereis a sonic point in an expandingregion. of B. A two-dimensional problem.The testproblemis the focusing a plane shock methodswe have The Godunov-type wave in air (y = 1.4) by a parabolic reflector.

314

BERND

EINFELDT

extension the HLLR scheme theshock FIG. 14. Second order focusing problem. of for

describedare onlyfirst-order accuratein smoothpartsof thesolution.For theapplicaTo tion to "real world" problemsfirst-order accuracyis not sufficient. computethe of solutionof thefocusing a plane shockwave,we have incorporated numerical Roe's Riemann solver and the HLLR-Riemann solver in the second-orderextensionof Godunov's method, describedby Colella and Glaz in [3]. the The numerical calculationsweremade in two space dimensions extending by via The Mach number theincident of one-dimensional method a fractional method. step shock wave is 1.1. In the numerical is calculation,the shape of the reflector approxiboundaryconditions[2]. Because matedby a -stepwise wall to apply simplereflecting the consideredflowproblemis symmetrical, only the upper halfplaneof the flowis boundary by computed.The halfplanewas discretized a 90 x 90 grid.At the reflector and the axis of symmetry, reflection the boundaryconditionis applied. At the other is boundariesof thegridno reflection assumed.All calculationsweredone on a Cyber 175 computer.

ON GODUNOV-TYPE

METHODS

315

_< ___
2-

_~~~~~~~ _j

FIG.

15.

of extension Godunov's method. for history a second-order Pressure

of A more detailed description thistestproblemis givenby Olivierand Gr6nig calculationswiththe Random Choice Method and compared in [16]. Theyperformed data. Experiproblemwithexperimental the numericalresultsforthe shock focusing in in studiesof shockwave focusing waterare performed [21]. mentaland numerical the Figure13 represents resultsforthe PLM-method[3] withthe fullsolutionto method the Riemannproblem.The Riemannproblemwas computedby the iterative the from wall, after [24]. The pressurecontoursare shownin the processof reflection of fromthe wall, at the moment and afterfocusing. reflection Figure 14 shows the resultforthe PLM-methodwiththe HLLR-Riemannsolver. The pressurecontoursare verysimilar. can be obtained fromFigs. 15-17. The pressure More particularinformation are historiesclose to the reflector shown (see [16] for details of the location). No of was difference foundbetweenthe pressurehistories the PLM-methodwiththe full solutionto the Riemannproblem(Fig. 15) and withRoe's Riemannsolver(Fig. 16). the with HLLR-Riemann of histories thePLM-method the Figure17represents pressure are pressurehistory the same as in Fig. 15. The solver.The shock waves in the first thereis a slight In higher. the second pressurehistory in maximum Fig. 15 is slightly "kink" behind the shock wave, whichdoes not occur in Fig. 15. The peak is slightly for higher the HLLR-Riemann solverbut is more rounded. to 7. Discussionand conclusion.In thispaper we have shownthatit is sufficient signalvelocity physical for approximation the smallestand largest derivea numerical Riemannsolverforgas dynamics. in the Riemannproblemto obtain an efficient

316

BERND

EINFELDT

2.4

.8.
.0-

_______

________

___

.00

.20

.40

2.4==

_I

1.6

=_

.B
.-

.00

.20

.40

extension Roe's method. of FIG. 16. Pressure history a second-order for

A stability conditionforthe numericalsignal velocitieswas derivedin ? 4. Here a betweenthe signal velocitiesand the dissipation we also demonstrate relationship method. Godunov-type containedin the corresponding The computation the signal velocitiesforgeneral (convex) equation of state of was discussed in ? 5. The solver(3.3) is itssimplicity. approximation Thegreat advantageoftheRiemann while retainingessential featuresof reduces the programcomplexity substantially of Godunov's method,especiallythe accurate approximation shock waves. Furthera vectorizable versionis easily obtained. more, method(3.6), (3.7) is thatthe resolutionof a A drawbackof the Godunov-type moregridpointsas Godunov's method. requires contactdiscontinuity stationary nearly an can be overcomeby introducing antidiffusion We showed thatthis problem termin the lineardegeneratefield.If we choose the smallestand largesteigenvalues thentheGodunov-type oftheRoe linearization signalvelocities, [18] forthenumerical flux function (3.15), (4.20) becomesidentical numerical method(3.6) withthemodified of Thus we obtaineda new interpretation Roe's scheme[18]. withRoe's fluxfunction. The successfulapplicationto the shock focusingproblemshows the usefulness scheme. of the Riemannsolverin a higher-order Godunov-type of The the thanks Stosswellenlabor theRheinisch-WestAcknowledgments. author for fiilische TechnischeHochschuleAachen formakingavailable theplotsoftware the

ON GODUNOV-TYPE

METHODS

317

2.4

1 .6

2
.6-

.00

.20

.40

extension theHLLR of a second-order 17. historyfor FIG. Pressure

scheme.

the from Stosswellenare Special thanks to Dr. M. Sommerfeld problem. shockfocusing calculationand theprovision shockfocusing labor,forhis help in thetwo-dimensional PLM-code. of the two-dimensional

REFERENCES
[1] P. COLELLA

[2] [3] [4] [53 [6] [7] [8] [9] [10]

method (PPM) for gas-dynamical AND P. R. WOODWARD, The piecewise-parabolic J. simulations, Comput. Phys.,54 (1984), pp. 174-201. SIAM J.Sci. Statist. gas schemefor dynamics, Eulerianmuscle P. COLELLA, A direct Comput.,6 (1985), pp. 104-117. for problem realgases,J. for solution algorithms theRiemann P. COLELLA AND H. M. GLAZ, Efficient Comput. Phys.,59 (1985), pp. 264-289. New Flow and Shock Waves,Wiley Interscience, R. COURANT AND K. 0. FRIEDRICHS, Supersonic York, 1948. J. Godunov'smethod, Comput.Phys.,61 Riemannsolverfor non-iterative J. K. DUKOWICZ, A general, (1985), pp. 119-137. Computing, 39 (1987), pp. zur B. EINFELDT, Ein schneller Algorithmus Losungdes RiemannProblems, 77-86. solutions equations of scheme numerical of computation disco'ntinuous for S. K. GODUNOV, A difference Mat. Sb., 47 (1959), pp. 271-290. offluiddynamics, thisJournal, total-variation-stable A. HARTEN, On a class of highresolution schemes, finite-difference 21 (1984), pp. 1-23. laws, conservation J.Comput.Phys., 49(1983), pp.357-393. schemesfor hyperbolic -, Highresolution schemes for and A. HARTEN, P. D. LAX AND B. VAN LEER, On upstream differencing Godunov-type conservation hyperbolic laws,SIAM Rev., 25 (1983), pp. 35-61.

318

BERND

EINFELDT

Comm. computation, numerical equationsand their hyperbolic of [11] P. D. LAX, Weak solutions nonlinear J. convection, Comput. Phys.,23 (1977), pp. 276-299. schemesV. A second ordersequel to Godunov's conservative ' Towardsthe ultimate difference J. method, Comput. Phys.,32 (1979), pp. 101-136. and Roe, schemes Godunov,Engquist-Osher of the between upwind-differencing [14] , On therelation Comput.,5 (1984), pp. 1-20. SIAM J. Sci. Statist. laws, of systems conservation for schemes hyperbolic difference [15] S. OSHER AND F. SOLOMON, Upwind Math. Com., 38 (1982), pp. 339-374. shockfocussing choicemethod appliedto two-dimensional [16] H. OLIVIER AND H. GRONIG, The random J. and diffraction,Comput. Phys.,63 (1986), pp. 85-106. AIAAJ., 22 (1984), pp. 602-610. of to [17] M. PANDOLFI, A contribution numericalpredictionunsteadyflows, J. schemes, Comput. Phys., and vectors, difference parameter Riemannsolvers, [18] P. L. ROE, Approximate 53 (1981), pp. 357-372. LectureNotes in Physics141, Schemes, in [19] , The Use of theRiemannProblem Finite-Difference New York, 1981,pp. 354-359. Springer-Verlag, LectureNotesin AppliedMathematics to of - , Somecontributionsthemodeling discontinuousflows, [20] New York, 1985,pp. 163-193. 22, Springer-Verlag, in studiesofshockwavefocusing and [21] M. SOMMERFELD AND H. M. MULLER, Experimental numerical in Experiments Fluids, submitted. water, Equations,Springer-Verlag, New York, 1983. [22] J. SMOLLER, Shock Wavesand Reaction-Diffusion Press,Cambridge,1985. CambridgeUniversity Methodsin Fluid Dynamics, SOD, Numerical [23] G. A. laws, conservation hyperbolic of systems nonlinear methodsfor ~ , A survey severalfinite of difference [24] J. Comput. Phys.,27 (1978), pp. 1-31. method hybrid [25] P. WOODWARD, P. COLELLA, B. A. FRYXELL AND K. H. WINKLER, An implicit-explicit J. for Lagrangianhydrodynamics,Comput. Phys.,63 (1986), pp. 283-310. with of simulation two-dimensionalfluidflow strong [26] P. WOODWARD AND P. COLELLA, The numerical J. shocks, Comput.Phys.,54 (1984), pp. 115-173. IV. scheme, A new approachto numerical difference conservative [12] B. VAN LEER, Towardstheultimate Pure Appl. Math.,7 (1954), pp. 159-193.

[13]

You might also like