P. 1
WENO_Tri

WENO_Tri

|Views: 0|Likes:
Published by Chung-Yueh Wang

More info:

Published by: Chung-Yueh Wang on Aug 25, 2013
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

08/25/2013

pdf

text

original

Journal of Computational Physics 150, 97–127 (1999) Article ID jcph.1998.6165, available online at http://www.

idealibrary.com on

Weighted Essentially Non-oscillatory Schemes on Triangular Meshes1
Changqing Hu∗ and Chi-Wang Shu†
Division of Applied Mathematics, Brown University, Providence, Rhode Island 02912 E-mail: ∗ hu@cfm.brown.edu and †shu@cfm.brown.edu Received June 16, 1998; revised November 24, 1998

In this paper we construct high-order weighted essentially non-oscillatory schemes on two-dimensional unstructured meshes (triangles) in the finite volume formulation. We present third-order schemes using a combination of linear polynomials and fourthorder schemes using a combination of quadratic polynomials. Numerical examples are shown to demonstrate the accuracies and robustness of the methods for shock calculations. c 1999 Academic Press Key Words: weighted essentially non-oscillatory schemes; unstructured mesh; high-order accuracy; shock calculations.

1. INTRODUCTION

ENO (essentially non-oscillatory) schemes (Harten et al. [16], Shu and Osher [28, 29]) have been successfully applied to solve hyperbolic conservation laws and other convection dominated problems, for example in simulating shock turbulence interactions (Shu and Osher [29], Shu et al. [30], and Adams and Shariff [2]), in the direct simulation of compressible turbulence (Shu et al. [30], Walsteijn [35], and Ladeinde et al. [20]), in solving the relativistic hydrodynamics equations (Dolezal and Wong [8]), in shock vortex interactions and other gas dynamics problems (Casper and Atkins [6] and Erlebacher et al. [10]), in incompressible flow calculations (E and Shu [9] and Harabetian et al. [13]), in solving the viscoelasticity equations with fading memory (Shu and Zeng [31]), in semiconductor device simulation (Fatemi et al. [11] and Jerome and Shu [17, 18]), and in image processing and level set methods (Osher and Sethian [24], Sethian [26], and Siddiqi et al. [32]). The original ENO paper by Harten et al. [16] was for a one-dimensional finite volume formulation. Later,
Research was supported in part by ARO Grant DAAG55-97-1-0318; NSF Grants DMS-9804985, ECS9627849, and INT-9601084; NASA Langley Grant NAG-1-2070 and Contract NAS1-97046 while the second author was in residence at ICASE, NASA Langley Research Center, Hampton, VA 23681-2199, and AFOSR Grant F49620-96-1-0150. 97
0021-9991/99 $30.00 Copyright c 1999 by Academic Press All rights of reproduction in any form reserved.
1

98

HU AND SHU

this finite volume formulation of ENO schemes was extended to two-dimensional structured meshes by Harten [14] and by Casper [5], and to unstructured triangular meshes by Abgrall [1], Harten and Chakravarthy [15], and Sonar [34]. Finite volume ENO schemes based on a staggered grid and Lax–Friedrichs formulation were given in Bianco et al. [4]. Although finite difference versions of ENO schemes [28, 29] are more efficient for multidimensional calculations, finite volume schemes have the advantage of easy handling of complicated geometry by arbitrary triangulations. Weighted ENO (WENO) schemes were developed later to improve upon ENO schemes, in Liu et al. [23] and Jiang and Shu [19]. Advantages of WENO schemes over ENO include the smoothness of numerical fluxes, better steady-state convergence, and better accuracy using the same stencils. Levy et al. [22] designed one-dimensional finite volume WENO schemes based on a staggered grid and Lax–Friedrichs formulation. For a review of ENO and WENO schemes, see [27]. Recently, Friedrich [12] constructed WENO schemes on unstructured meshes using a covolume formulation as in Abgrall [1]. The WENO schemes in [12] only achieve the same order of accuracy as the corresponding ENO schemes when the same set of stencils is considered. This is not optimal, as was known in Jiang and Shu [19] for structured meshes. In this paper, we present higher order WENO schemes on triangular meshes when using the same set of ENO stencils. We will construct third-order schemes using a combination of two-dimensional linear polynomials and fourth-order schemes using a combination of two-dimensional quadratic polynomials. We will first sketch the procedure to construct the high-order linear schemes. The formulation at this stage is important to accommodate nonlinear WENO weights later. We then describe the third- and fourth-order WENO schemes. Numerical examples will be given, to demonstrate the accuracy and resolution of the constructed schemes. Concluding remarks are included at the end.
2. FINITE VOLUME FORMULATION

In this paper we solve the two-dimensional conservation law ∂ f (u ) ∂ g (u ) ∂u + + =0 ∂t ∂x ∂y (2.1)

using the finite volume formulation. Computational control volumes are simply triangles. Taking the triangle i as our control volume, we formulate the semi-discrete finite volume scheme of Eq. (2.1) as 1 d ¯ i (t ) + u dt | i| F · n ds = 0,
i

(2.2)

¯ i (t ) is the cell average of u on the cell i , F = ( f , g )T , and n is the outward unit where u normal of the triangle boundary ∂ i . The line integral in (2.2) is discretized by a q -point Gaussian integration formula,
q ∂

F · n ds ≈ |∂
i

i| j =1

ω j F (u (G j , t )) · n ,

(2.3)

WENO SCHEMES ON TRIANGULAR MESHES

99

and F (u (G j , t )) · n is replaced by a numerical flux. The simple Lax–Friedrichs flux is used in all our numerical experiments, which is given by F (u (G j , t )) · n ≈ 1 [( F (u − (G j , t )) + F (u + (G j , t ))) · n − α(u + (G j , t ) − u − (G j , t ))], 2 (2.4)

where α is taken as an upper bound for the eigenvalues of the Jacobian in the n direction, and u − and u + are the values of u inside the triangle and outside the triangle (inside the neighboring triangle) at the Gaussian point. Since we are constructing schemes up to fourth-order accuracy, two-point Gaussian √ 3 + and q = 2 is used, which has G 1 = c P1 + (1 − c) P2 , G 2 = c P2 + (1 − c) P1 , c = 1 2 6 1 ω1 = ω2 = 2 for the line with endpoints P1 and P2 .
3. RECONSTRUCTION AND LINEAR SCHEMES

Let P k denote the set of two-dimensional polynomials of degree less than or equal to k . The reconstruction problem, from cell averages to point values, is as follows: given a smooth function u , and a triangulation with triangles { 0 , 1 , . . . , N }, we would like to construct, for each triangle i , a polynomial p (x , y ) in P k that has the same mean value as u on i , and is a (k + 1)th-order approximation to u on the cell i . The mean value of a function u (x , y ) on a cell i is defined as ¯i ≡ u 1 |
i|
i

u (x , y ) d x d y .

(3.1)

k + 2) In order to determine K = (k + 1)( degrees of freedom in a k th degree polynomial p , 2 we need to use the information of at least K triangles. In addition to i itself, we may take its K − 1 neighboring cells, and we rename these K triangles as Si = { 1 , 2 , . . . , K }, ¯ j on j Si is called a stencil for the triangle i . If we require that p has the mean value u for all 1 ≤ j ≤ K , we will get a K × K linear system. If this linear system has a unique solution, Si is called an admissible stencil. Of course, in practice, we also have to worry about any ill-conditioned linear system even if it is invertible. For linear polynomials k = 1, a stencil formed by i and two of its neighbors is admissible for most triangulations.

3.1. Third-Order Reconstruction To construct a third-order linear scheme (a scheme is called linear if it is linear when applied to a linear equation with constant coefficients) as a starting point for the WENO procedure, we need a quadratic polynomial reconstruction. Notice that, as a linear scheme, the stencil of this quadratic polynomial depends not on the solution, but only on the local geometry of the mesh. It seems that one robust way is the least-square reconstruction suggested by Barth and Frederickson [3]. For the control volume of triangle 0 (see Fig. 3.1), let i , j , k be its three neighbors, and ia , ib be the two neighbors (other than 0 ) of 2 i , and so on. We determine the quadratic polynomial p by requiring that it have the same cell average as u on 0 , and also it matches the cell averages of u on the triangles in the set {
i, ia , ib , j, ja , jb , k, ka , kb },

k . This looks like an overdetermined system. u ¯ kb }. (3. i .3) where p 2 is defined before using the least-square procedure. jb). ia ). for arbitrary choices of cell averages ¯ 0. j . ib). ja ). which are constants depending only on the local geometry of the mesh. u ¯ jb . p2 (on triangles: 0. 3. u ¯ j. ib might be the same as ja .) may coincide. A typical stencil. ja . p5 (on triangles: 0. i . kb). We first build the following nine linear polynomials by agreeing with the cell averages of u on the following stencils: p1 (on triangles: 0. A key step in building a high-order WENO scheme based on lower order polynomials is carried out in the following. .2) satisfies R (x G . . p3 (on triangles: 0. u ¯i. k . which leads u to 10 linear equations for the nine weights γs . u ¯ ia . k . 9 R (x . i ). j ). (3. We want to construct several linear polynomials whose weighted average will give the same result as the quadratic reconstruction p 2 at each quadrature point (the weights are different for different quadrature points). i .1. p8 (on triangles: 0. such that the linear polynomial obtained from a linear combination of these ps . we want to find the linear weights γs .4) Since both the left side and the right side of the equality (3. u ¯ ib . j . however. for the equality to hold for arbitrary u ¯ ’s to be identically zero (when all terms are moved to one side of the equality).100 HU AND SHU FIG. Notice that some of the neighbors’ neighbors ( ia . ib . For each quadrature point (x G . For example. j . p7 (on triangles: 0. y G ). . ka ). p6 (on triangles: 0. in a least-square sense (as this is an overdetermined system). k ). u ¯k.4) one must have all 10 coefficients of the (3. does not affect the least-square procedure to determine p2 .3) are linear in the cell averages ¯ ’s in (3. y ) = s =1 γs ps (x . u ¯ ja .4). u ¯ ka . . This. p4 (on triangles: 0. {u (3. y G ). y ). and p9 (on triangles: 0. y G ) = p 2 (x G .

.3) is valid for all three cases u = 1.7) in the cases of u = 1.7) holds in these three cases then it holds for arbitrary ¯ ’s. as all three expressions in (3. I1 . Thus we can eliminate 2 equations from the 10.3). it turns out that. 2. 2 2 6 1 1 γ2 = . The first effort we would like to make is to use this degree of freedom to obtain a set of nonnegative γs . The same argument can be applied in the current two-dimensional case. Let us denote I j .5) and (3. resulting in a rank 8 system with one degree of freedom in the solution for γs .5) The quadratic polynomial p 2 . u = x .6) ¯ ’s. which is important for the WENO procedure to be developed later for shock calculations. 2 6 3 1 5 γ1 + γ2 = . u 2 2 (3. and u = y under only one 2 constraint on γs . The two linear polynomials ps . as three equal-sized consecutive intervals. resulting in a solvable 2 × 2 system for γs . The crucial observation is that (3. give the following two second-order approximations to the value of u at the point x3/2 (the boundary of I1 and I2 ): 3 1 ¯0 + u ¯ 1. and I2 . where p1 agrees with u on cell averages in the intervals I0 and I1 . Although there are 10 linear equations for the nine weights γs resulting from equality (3.7) (3. j = 0. 1.WENO SCHEMES ON TRIANGULAR MESHES 101 but is in fact underdetermined of rank 8. namely 9 s =1 γs = 1. 2 3 for the two unknowns γ1 and γ2 . Before explaining this. This leads to the three equations for arbitrary u 1 1 − γ1 = − . 3 The reason can be understood if we ask for the validity of the equality (3. and u = x 2 with only two conditions: γ1 + γ2 = 1 and another one obtained when u = x 2 . Hence the equality (3. gives the following third-order approximation to the value of u at the point x3/2 : 5 1 1 ¯0 + u ¯1 + u ¯ 1. this is impossible. − u 2 2 1 1 ¯1 + u ¯ 2. Clearly if (3. again because ps ( x ) and p ( x ) all reproduce linear functions exactly. In practice. Unfortunately. and u = x 2 . for many triangulations. we first look at a simpler but illustrative one-dimensional example. which agrees with u on cell averages in the intervals I0 . we obtain the solution γs for s ≥ 2 with γ1 as the degree of freedom.7) is valid for all the three cases u = 1. It looks like an overdetermined system but is in fact rank 2 and has a unique solution γ1 = 1 . we should notice that equality (3.6) reproduce linear functions exactly. and p2 agrees with u on cell averages in the intervals I1 and I2 . allowing for one degree of freedom in the choice of the nine γs .7) holds for both u = 1 and u = x as long choices of u as γ1 + γ2 = 1. 3 γ2 = 2 . u = x . u = x . − u 6 6 3 We would like to find γs such that 3 1 ¯0 + u ¯1 γ1 − u 2 2 + γ2 1 1 ¯1 + u ¯2 u 2 2 5 1 1 ¯0 + u ¯1 + u ¯1 =− u 6 6 3 (3. Some grouping is needed and will be discussed later.

ka . We construct the cubic polynomial p3 by requiring that its cell average agree with that of u on each triangle in the 10-triangle stencil shown in Fig. Again. ja . the key step in building a high-order WENO scheme based on lower order polynomials is carried out in the following. we can now obtain the third. j . y . It seems that for most triangulations this reconstruction is possible. ia ). ka . y ) = s =1 γs qs (x . k . (3. q2 (on triangles: 0. kb).8) satisfies Q (x G . jb).9) ¯ ’s.102 3. We can thus eliminate 5 equations from the 10. Fourth-Order Reconstruction HU AND SHU To construct a fourth-order linear scheme as a starting point for the WENO procedure.3. jb. we need a cubic polynomial reconstruction. A crucial observation is again that (3. (3. k . ka ). i . jb. . ia . 3. resulting in a rank 5 system with one degree of freedom in the solution for γs . j . which has 10 degrees of freedom. but it is in fact underdetermined with rank 5. u + (G j . We would like to construct several quadratic polynomials whose weighted average will give the same result as the cubic reconstruction p3 at each quadrature point (the weights are different for different quadrature points). we would like to find the linear weights such that the linear combination of these qs .9) is valid for all the six cases u = 1. y G ) in (3. y G ) in (3.4) with the reconstructed values R (x G . The following six quadratic polynomials are constructed by having the same cell averages as u on the corresponding triangles: q1 (on triangles: 0.2. In practice. ka . For each quadrature point (x G . ib.2) or Q (x G . ja ). j . y G ) (3. ja . thus allowing a solution for γs with one degree of freedom. i . k . x 2 . q5 (on triangles: 0. q6 (on triangles: 0. This looks like a grossly overdetermined coefficients of the 10 cell averages u system. because 3 qs (x ) and p (x ) all reproduce quadratic functions exactly. 3.1. Accuracy Test for the Linear Schemes From the third. y 2 under just one constraint on the γs .1). the first effort we would like to make is to use this degree of freedom to obtain a set of non-negative γs . i . for all u As before. namely 9 s =1 γs = 1.1) by replacing u − (G j . which are the ¯ ’s in (3. q4 (on triangles: 0. we obtain the solution γs for s ≥ 2 with γ1 as the degree of freedom. 6 Q (x . kb.8). t ) in (2. y G ) = p3 (x G .and the fourth-order linear schemes for (2. 3. ja . Similarly. kb are distinct in the stencil (see Fig. ib. ia . x y . respectively. ib. i . x . jb. j . q3 (on triangles: 0.9) results in 10 linear equations for the six unknowns γs . y G ).and the fourth-order reconstructions. This has been performed for the mostly near-uniform meshes used in the numerical examples. We thus only consider the case where ia . kb. y ). t ) is replaced with the reconstructed values in the neighboring cell i . k . ib).4). Again. which is important for the WENO procedure to be developed later for shock calculations.

94 4. the accuracy in each rectangle.93 2. We first use uniform triangular meshes which are obtained by adding one diagonal line . Linear Schemes P 1 (third order) h 2/5 1/5 1/10 1/20 1/40 1/80 L 1 error 1. where h is the length of the rectangles. shown in Fig.01 4.10) with the initial condition u 0 (x .07E-08 Order — 3.2. For the temporal discretization. Uniform Meshes.02 4.76E-05 7. at t = 2. There are more triangles here than in the uniform mesh cases with the same h . we use t = ( x )4/3 to achieve fourthorder accuracy in time.79E-01 4. The accuracy result is shown in Table 3. Uniform mesh with h = 2 5 for accuracy test. but only for the examples of the accuracy test.11E-04 5.80E-01 2. For the fourth-order scheme. 3.99 3. where h is just an average length. shown in Fig. −2 ≤ x ≤ 2. TABLE 3.81E-02 3.1 Accuracy for 2D Linear Equation.17E-02 1.68 2.2 for the coarsest case h = 2 5 results are shown for both the third-order scheme (from the combination of linear polynomials) and the fourth-order scheme (from the combination of quadratic polynomials).41E-03 8.37E-02 5.02 4.0.12E-07 1. EXAMPLE 3.2.01 L ∞ error 2.03 4.22E-04 9.43E-06 2.1. The refinement of the meshes is done in a uniform way.13E-05 Order — 2.02 4.34E-07 2.00 3. −2 ≤ y ≤ 2. We remark that h here is only a rough indicator of the mesh size. 3.72E-03 7. (3.95 2. In Table 3.60E-04 5. and periodic 2 boundary conditions.00 L ∞ error 2. the third-order TVD Runge–Kutta scheme of Shu and Osher [28] is used.32E-08 P 2 (fourth order) Order — 3. The errors presented are those of the cell averages of u .02 4.05E-05 1.99 3.1.57E-05 3.00 3. 3.39E-06 3.00 L 1 error 1.21E-06 Order — 2. We then use non-uniform meshes.68 2.01 .72E-05 5.94 4.65E-03 4.3 for the coarsest case h = h 0 = 1.WENO SCHEMES ON TRIANGULAR MESHES 103 FIG.40E-02 9. namely by cutting each triangle into four smaller similar ones. y ) = sin( π (x + y )). Two-dimensional linear equation u t + u x + u y = 0.

02 (3.73E-02 1.42E-03 8.7 sin( π (x + y )). shown in Fig. See [27] for a description of this problem. EXAMPLE 3.61 4.78E-05 4. y ) = 0. −2 ≤ y ≤ 2. We consider the compressible Euler equations of gas dynamics ξt + f (ξ )x + g (ξ ) y = 0.1.40E-05 1.81E-02 2.03 L ∞ error 1.74 2.03 4.01 4.22E-05 Order — 2. EXAMPLE 3.95E-08 Order — 3.16E-07 1.11) with the initial condition u 0 (x . as in this nonlinear case it is easier to obtain the exact solution of the PDE in point values than in cell averages.90E-04 2.80 2. The accuracy result is shown in Table 3.5/π 2 when the solution is still smooth.36E-03 3.94 2.98 2.09 4. In Table 3.4. Linear Schemes P 1 (third order) h h 0 /2 h 0 /4 h 0 /8 h 0 /16 h 0 /32 h 0 /64 L 1 error 1.32E-05 5.75E-09 P 2 (fourth order) Order — 4. 3. the accuracy results are shown for both the third-order the coarsest case h = 2 5 scheme and the fourth-order scheme. 3. shown in Fig.09 3.10 4.76E-08 4. TABLE 3.1.95E-03 2.3. −2 ≤ x ≤ 2. Non-uniform Meshes.94 2. at t = 0.19E-04 9. The errors presented are those of the point values at the six quadrature points of each triangle. We first use the same uniform triangular meshes as in Example 3.3.25E-01 3.104 HU AND SHU FIG. Non-uniform mesh with h = 1 for accuracy test.75E-06 Order — 2. y (3.71 4.39E-03 7. Two-dimensional Burgers’ equation ut + u2 2 + x u2 2 = 0. Two-dimensional vortex evolution problem for the Euler equations.2 Accuracy for 2D Linear Equation.59 2.12) . 3.99 2. and 2 periodic boundary conditions. We then use the same non-uniform meshes as in Example 3.99 L ∞ error 2.09E-06 3.21E-05 1.74E-02 5.91 2.21E-01 1.2 for .2.3 + 0.95 L 1 error 4.06 4.3 for the coarsest case.29E-06 7.3.00E-04 3.

00 L 1 error 8.34E-05 1.69 2.00 4. r 2 = x ¯2 + y ¯ 2 .81E-04 1.00 L ∞ error 1. y ¯ ) = (x − 5.68 3. 2 δ S = 0.16E-02 1. E is the total energy.98 4.83E-09 P 2 (fourth order) Order — 3.63E-07 Order — 2.67E-02 3.72 .08E-04 3.99 4.94 3. p = 1.69E-03 2.3 Accuracy for 2D Burgers’ Equation.57E-05 4.99 3.57E-08 Order — 3. ρv.00E-05 4.98 4.18E-06 8. to the mean flow. u ( E + p )).52E-03 1.58 3.57E-07 9.70E-03 1. 1). ρ u v. p is the pressure.38E-05 2.01 3.96E-07 Order — 2.99 2.99 2.99 3.79 3. x ¯) e0.74 2. = 5. y − 5). and (u . E ).77 3.00 4.87 2.60E-04 5.00 L ∞ error 2. Non-uniform Meshes. an isentropic vortex (perturbations in (u .17E-03 1.23E-03 2.83 3. We add. Here ρ is the density.97E-05 2.88E-02 2. ρ u v.93 2. v) and the temperature T = p /ρ .90E-05 1. g (ξ ) = (ρv.75E-09 P 2 (fourth order) Order — 3. v( E + p )). (u .87E-04 1. Linear Schemes P 1 (third order) h 2/5 1/5 1/10 1/20 1/40 1/80 L 1 error 2. f (ξ ) = (ρ u .95E-01 1.00 L ∞ error 7.5(1−r ) (− y 2 2 e1−r . v) = (1.82 L 1 error 3.12 3.93 2.4 Accuracy for 2D Burgers’ Equation.48E-06 5. γ −1 2 with γ = 1.23E-02 1.89 2.57E-08 4. no perturbation in the entropy S = p /ρ γ ) (δ u .00 3.86 2. The mean flow is ρ = 1. v) is the velocity. ¯. ρ u .92 2. ρ u 2 + p .75E-05 7.65E-03 4.99 3.00E-06 7.87 3.20E-06 7.63E-03 6.51E-06 1.97 2.01 3.74 3.91E-04 2.69E-02 2.97E-06 Order — 2. and E= 1 p + ρ(u 2 + v 2 ). ρv 2 + p . and the vortex strength where (x TABLE 3.16E-04 7.62E-07 2.00 where ξ = (ρ. Linear Schemes P 1 (third order) h h 0 /2 h 0 /4 h 0 /8 h 0 /16 h 0 /32 h 0 /64 L 1 error 1.99 L ∞ error 7.4.18E-02 1.WENO SCHEMES ON TRIANGULAR MESHES 105 TABLE 3. Uniform Meshes.22E-04 3.75E-02 1.89E-08 Order — 3.37E-06 4.92 3.84E-04 3. δv) = 2π (γ − 1) δT = − 8γ π 2 ¯.96E-03 2.26E-06 Order — 2.

6 for the non-uniform meshes.28 2.77 4.57 3.53E-02 4. TABLE 3.31 The computational domain is taken as [0.59 2.40 4.34 3.57 2. while the dissipation is quite noticeable for the long-time results of the third-order scheme.98 4.27 2.31E-03 1.84 3.56E-08 P 2 (fourth order) Order — 2.62E-02 1.10 L ∞ error 8.31E-03 2.11 L ∞ error 7.67E-03 2.106 HU AND SHU TABLE 3.48E-06 1.86 2. 10] × [0.55 2.10 2. We show the line cut through the center of the vortex for the density ρ . 3.25E-04 Order — 1.65E-02 6. and periodic boundary conditions are used.5 for the uniform meshes and Table 3.95E-04 1.98 L ∞ error 2.36E-04 5.89E-02 1.21E-04 2.67 L 1 error 7. and 10 time periods. Non-uniform Meshes.26E-03 7.89 2. The meshes are the same as those in Example 3. 5. We first compute the solution to t = 2.44E-01 2.and fourth-order schemes.5 is the result by the fourth-order scheme.77E-06 Order — 1. Figure 3.44 2.16E-02 1.4 is the result by the third-order scheme at t = 0 and after 1.81E-02 7.43 4.54 2.57 3.31E-03 1.10E-04 7.37E-09 P 2 (fourth order) Order — 2.28 3. Linear Schemes P 1 (third order) h h 0 /2 h 0 /4 h 0 /8 h 0 /16 h 0 /32 h 0 /64 L 1 error 1.86 2. The reconstruction procedure is applied to each component of the solution ξ .23E-05 Order — 1.97 L 1 error 5. (uniform) and compute the long-time evolution of the We then fix the mesh at h = 1 8 vortex. and Fig.86E-04 3.10E-07 6. The accuracy results are shown in Table 3.17E-05 4.12 4. The mesh chosen is on the borderline of resolution for the fourth-order scheme and does not give enough resolution for long-time simulation of the third-order scheme.70E-06 2. It is clear that the exact solution of the Euler equation with the above initial and boundary conditions is just the passive convection of the vortex with the mean velocity.18E-03 8.21 2.83 3.47E-02 4.98E-01 1.36E-05 2.6 Accuracy for 2D Euler Equation of Smooth Vortex Evolution.37 2.40E-05 2.5 Accuracy for 2D Euler Equation of Smooth Vortex Evolution. It is easy to see the difference between the third.0 for the accuracy test.61E-02 1. The errors presented are those of the cell averages of ρ .05 2.26 4.94E-05 5.99E-07 Order — 2.07E-06 Order — 1.85 4.79E-03 7.69 . The fourth-order scheme gives almost no dissipation even after 10 periods. Linear Schemes P 1 (third order) h 1 1 /2 1 /4 1 /8 1/16 1/32 L 1 error 1.60E-01 1.39 2. 10].13 4.21E-01 2.26 2.98E-05 3.73E-06 5.03E-03 5.54 4.66E-03 6.1 suitably scaled for the new spatial domain.96 L ∞ error 2.68E-07 1.00E-03 1.62 3.32E-06 1. Uniform Meshes.44E-04 8.25E-07 Order — 2.74E-03 1.

by consistency. we need non-negative weights. We want to group these nine linear polynomials into three groups.1. 3. .5. γ ˜s p FIG. 2D vortex evolution: third-order linear scheme. In the case of the third-order scheme. and thus we need non-negative linear weights to start with. WENO RECONSTRUCTION AND WENO SCHEMES In this section.2). grouping of polynomials is also used to achieve positivity. In (3. y ) = s =1 s =1 ˜s (x . we will introduce non-linear weights to make the resulting schemes suitable for shock computations. 4. this will be explored to produce positive linear weights. To ensure stability near shocks.WENO SCHEMES ON TRIANGULAR MESHES 107 FIG. As there is one degree of freedom in the choice of linear weights for both the third-order and fourth-order cases described in the previous section. Positivity of Linear Weights for the Third-Order Scheme We will use the same notation as in the previous section. 4. y ). 3. 9 3 γs ps (x .4. 9 s =1 γs = 1. 2D vortex evolution: fourth-order linear scheme.

so that when shocks are present. y ) to be reasonably separated. j ). kb). k . i . y ) = R s =1 ˜s (x . Group 1 contains p3 (0. p γ ˜1 = γ2 + γ4 + γ5 . We can do the same thing for the other two sides( j . ja ). ka ). and p7 (0. and p9 (0. k ). in a joint project with R. We do not need to store ˜ 1 itself. p8 (0. = a1 u (4. k . 2 . and p7 (0. Group 3 contains p1 (0. with the combination coefficient γ ˜1 = γ3 + γ4 + γ5 . ˜ 1 = (γ2 p2 + γ4 p4 + γ5 p5 )/(γ2 + γ4 + γ5 ). ˜ 3 = (γ1 p1 + γ8 p8 + γ9 p9 )/(γ1 + γ8 + γ9 ). j . k . 3. and p5 (0. ˜ (x . jb). cases when it does give some negative coefficients with very small magnitudes. but it does not seem to affect the stability of WENO schemes built upon them. y ) γ ˜s p (4. Recently we have learned. i . p The resulting linear polynomial 3 γ ˜3 = γ1 + γ8 + γ9 . ja ). k ). such that ˜1 x G . jb). The grouping we will introduce in the following works for most triangulations. ib). Biswas and J. p γ ˜2 = γ3 + γ6 + γ7 . not linear polynomials p all stencils will contain the shock under normal situations. p8 (0. ia ). Djomehri on using adaptive methods with WENO schemes. 3. y G p 1 1 ¯ 0 + a2 u ¯ i + a3 u ¯ k + a4 u ¯ ia + a5 u ¯ ib . it is the five constants ai which depend on the local geometry only. with each p positive coefficients γ ˜s ≥ 0. i . y ) being still a linear polynomial and a second-order approximation to u . p6 (0.1). i ). ˜ 2 = (γ3 p3 + γ6 p6 + γ7 p7 )/(γ3 + γ6 + γ7 ). j . We remark that for practical implementation. j ). Group 3 contains p1 (0. and p9 (0.2 For the first quadrature point on side i (G 1 in Fig. with combination coefficient γ ˜3 = γ1 + γ6 + γ7 . i . i . kb). ka ). We also require the stencils corresponding to the three new ˜s (x . i ). k . p4 (0. y ) in (3. k . p4 (0. New grouping techniques to overcome this are under investigation. Group 2 contains p2 (0. through the solution of a group of three linear inequalities for γ1 . j . j .1). Group 1 contains p2 (0.2) that have to be precomputed and stored once the mesh is generated. that this grouping may fail near adaptively refined regions where triangle sizes change very abruptly. There are. k . with the combination coefficient γ ˜2 = γ2 + γ8 + γ9 . p6 (0. Group 2 contains p3 (0. and p5 (0. however. any information about the polynomial p For the second quadrature point on side i (G 2 in Fig.2) and in most cases the coefficients γ ˜s can be made non-negative by suitably choosing the value of the degree of freedom γ1 . j .1) is identical to R (x . ia ). j . i . k ). ib).108 HU AND SHU ˜s (x .

we define the following measurement for smoothness. larger gives better accuracy for smooth problems but may generate small oscillations for shocks. We do not need to store any information about the ˜ 1 itself.3. Smaller is more friendly to shocks.e.. The numerical results are not very sensitive to the choice of in a range from 10−2 to 10−6 . 2) then |α | = 3 and D α p (x .e. The difficulty in doing a grouping here similar to the third-order case is that the stencil to each group would then be quite large.4) ˜s in (4. y ) (i.8) for the fourth-order case)..4) would then replace the linear weights γ j to form a WENO reconstruction. ( + Si )2 (4. The nonlinear weights ω j in (4. Our objective is to use this degree of freedom to obtain non-negative linear weights. that have to be precomputed and for the smoothness measurements (4.2. We have found out through our numerical experiments that positivity of linear weights can only be achieved for nearly uniform meshes. Positivity of Linear Weights for the Fourth-Order Scheme From Section 3.5) S = (b 1 u ˜ 1 in (4. and hence shocks may be included in all stencils. such that ¯ 0 + b2 u ¯ i + b3 u ¯ k + b4 u ¯ ia + b5 u ¯ ib )2 + (c1 u ¯ 0 + c2 u ¯ i + c3 u ¯ k + c4 u ¯ ia + c5 u ¯ ib )2 (4. which depend on the local geometry only.2. For a polynomial p (x . The nonlinear weights are then defined as ωj = ω ˜j . In this paper only such meshes are used for the fourth-order WENO scheme. and is a small positive number which we take as = 10−3 for all the numerical experiments in this paper. For example. We emphasize that the smoothness measurements (4. ˜i i ω ω ˜i = γi .3) are quadratic functions of the cell averages in the stencil. (4. y ))2 d x d y . Si is the measurement ˜s in (4. This again amounts to solving a group of six linear inequalities for γ1 . Good grouping strategies are still under investigation. polynomial p .8) for the fourth-order case). the γ for the third-order case and the γs in (3.3) where α is a multi-index and D is the derivative operator. y ) = ∂ p3 (x .1). 4.3) of p stored once the mesh is generated.WENO SCHEMES ON TRIANGULAR MESHES 109 4. for example. when α = (1. y )/∂ x ∂ y 2 . y ) with degree up to k .1) where γi is the i th coefficient in the linear combination of polynomials (i. S= 1≤|α |≤k | ||α|−1 ( D α p (x . In general. Smoothness Indicators and Nonlinear Weights We finally come to the point of smooth indicators and nonlinear weights. the p the qs in (3. it is the 10 constants bi and ci .1) for the third-order case and of smoothness of the i th polynomial pi (x . there is a degree of freedom (which was chosen to be γ1 ) in the determination of the linear weights γs in (3.8). For this we follow exactly Jiang and Shu [19].

the constant coefficients in the linear combinations inside each of the squares of the smoothness indicator S in (4. per triangle. ∗ For the third-order case. using the prestored constants.1). One is to do so component by component. the number of multiplications is 306 and the number of additions is 234. each with 5 multiplications and 4 additions in (4.3). for each polynomial qs in (3. We will use componentwise methods for all numerical examples with the third-order scheme. —The total storage requirement per triangle: 270 numbers for the third-order case and 342 numbers for the fourth-order case. Denote by R the matrix of right eigenvectors and L the matrix of left eigenvectors of A.2). We will give a brief description in the following. There are 5 × 3 × 6 = 90 such constants per triangle for the third-order case (5 for ˜s in (4.8). This is easy to implement and cost effective.5. the constant coefficients in the combination of all the quadratic terms of the smoothness indicator S in (4. for each of the 6 quadrature points). u polynomials qs in (3.4. It reconstructions. Let us take one side of the triangle which has the outward unit normal (n x .8)).1). 4. • Start the time iteration..2). Examples of these constants are given in (4. This is because for each quadrature point there are 3 ˜s in (4. With the reconstructed point values v . such as u 0.3). For the fourth-order case. compute the lower order reconstructions for each quadrature point and the nonlinear weights. to each component of the vector v = L u .110 4. Let A be some average Jacobian at one quadrature point. we define our reconstructed point values u by u = R v . There are 21 × 6 = 126 such constants per triangle (21 quadratic ¯ i . each of the 3 grouped linear polynomials p and 6 × 6 × 6 = 216 such constants per triangle for the fourth-order case (6 for each of the ˜ s in (3. Extension to the Euler Systems HU AND SHU There are two ways to extend the previous results to systems. Examples of these constants are given in (4. 6 quadratic polynomials q —For the third-order case. out of the 6 u ¯2 ¯ 0u ¯ ’s in the stencil. i. . for each of the 6 quadrature points). for each stencil. for for the third-order case (10 for each of the 3 grouped linear polynomials p each of the 6 quadrature points). the Roe’s mean matrix [25] is used. For each time stage: —For each triangle. A = nx ∂f ∂g + ny .6) For Euler systems. for each of the 6 quadratic terms. for each stencil. . • Compute and store the following mesh-dependent constants: —The constant coefficients in the linear combinations of cell averages to recover values of u at Gaussian points. n y ). Then the scalar triangular WENO scheme can be applied to each of the characteristic fields.1). Algorithm Flowchart and Parallel Implementation We summarize the practical aspects and operation and storage costs of implementing the method: • Generate a triangular mesh.e. . There are 10 × 3 × 6 = 180 such constants per triangle ˜s in (4. and it seems to work well for the third-order scheme. ∂u ∂u (4. Another extension method is by the characteristic decomposition. for the 3 p .5).8). .

and 594 multiplications/divisions.10). and 6 flux evaluations for the fourth-order scheme. Two-dimensional Burgers’ equation as defined in Example 3.4) at each quadrature point.2 for the non-uniform meshes. per triangle.9 µs. NUMERICAL EXAMPLES We will implement the third.1 for the uniform meshes and in Table 5. (3. are 390 multiplications/divisions. This is about six times the CPU time needed for a linear scheme of the same order of accuracy implemented in a most economical way.3) for the whole triangle. the same accuracy tests as in the linear weights case are given for linear.1.WENO SCHEMES ON TRIANGULAR MESHES 111 ˜s in (4. As an explicit method. Most of the numerical examples with large meshes in the next section are obtained with the SP-2 parallel computer using 16 processors at the Center for Fluid Mechanics of Brown University. Eq. each with 6 multiplications and 5 additions.8). 375 additions/subtractions.3). —Compute the numerical flux (2. Accuracy Test for Triangular WENO Schemes We use the same examples as in Section 3. In our implementation on a SUN Ultra1 workstation. we continue the the above calculation to t = 5/π 2 when discontinuities develop. some non-smooth problems for two-dimensional Euler equations are tested. This is because for each quadrature point there are 6 reconstructions. Figure 5.1. The parallel efficiency is over 90% when 16 processors are used.4 for the non-uniform meshes. We can see that the correct orders of accuracy are obtained by the third. 285 additions/subtractions.2.11).and fourth-order WENO methods. then form the residue and forward in time. each with 12 also involves 3 evaluations of the smooth indicator (4.4). 5. Two-dimensional linear equation as defined in Example 3. The accuracy results are shown in Table 5.2. Eq. the third-order WENO scheme takes 33. including the WENO reconstruction and flux evaluation (2.1. the WENO schemes constructed above are easily implemented on an IBM SP-2 parallel computer. —The total costs per triangle per residue evaluation. for the 6 qs in (3. and 6 flux evaluations for the third-order scheme.6 µs per triangle per residue evaluation. each with 42 multiplications and 20 additions. 5.and fourth-order triangular WENO schemes constructed in the previous section. Next. The accuracy results are shown in Table 5. for the 3 p multiplications and 9 additions in (4.and fourth-order WENO methods.5). There are also 6 evaluations of the smooth indicator (4. for the 6 qs in (3. We can see that the correct orders of accuracy are again obtained by the third. EXAMPLE 5. ∗ For the fourth-order case.3 for the uniform meshes and in Table 5.and fourth-order WENO schemes developed in the previous sections to some two-dimensional test problems. Burgers’ equation and the smooth vortex problem. while the fourth-order WENO scheme takes 35. EXAMPLE 5. First. (3. To demonstrate the application for shock computations.8). the number of multiplications is 468 and the number of additions is 300.3 to test the accuracy of third.1 is the result for h = 1/20 of .1).

42 4.73 1.85E-05 1.40E-05 Order — 1.97 3.18E-04 6. Uniform Meshes.98 3.46E-04 7.00 4.18 3.45E-08 4.13 TABLE 5.77 3.94E-02 2.58 2.71 4.21E-06 Order — 1.22 3.25 L ∞ error 5.50E-06 1.31 4.34 4.28E-01 2. Non-uniform Meshes.23E-06 Order — 2.19 1.39E-03 3.37E-06 3.92E-09 P 2 (fourth order) Order — 4.18E-04 7.112 HU AND SHU TABLE 5.77 2.88E-04 2.76E-02 4.92E-04 1.376-02 2.35 4.50 4.94 3.07E-03 1.26 3.16E-03 1.44 4.87 3.63E-06 9. WENO Schemes P 1 (third order) h h 0 /2 h 0 /4 h 0 /8 h 0 /16 h 0 /32 h 0 /64 L 1 error 2.85 3.42 4.43E-02 2.15E-05 Order — 1.74 3.00 L ∞ error 2.84 3.00 L 1 error 8.71E-07 2.85E-03 5.08E-05 1.2 Accuracy for 2D Linear Equation.94 4.34E-06 2.38 4.82E-07 Order — 2.34E-08 P 2 (fourth order) Order — 2.00 L ∞ error 2.42E-06 4.36E-07 2.62 3.34E-05 5.16E-02 1.93E-01 6.106-02 1.72 .38E-02 1.74E-03 1.90E-04 2.44 3.08E-07 Order — 2.82E-06 8.41 3.80E-02 3.19 L ∞ error 6.12E-08 Order — 3.81 4.67E-09 P 2 (fourth order) Order — 3.47E-02 1.49 4.95E-05 4.99 4. WENO Schemes P 1 (third order) h 2/5 1/5 1/10 1/20 1/40 1/80 L 1 error 2.99 4.96E-06 Order — 2.32E-01 7.26 4.79E-04 5.49 4.16 L ∞ error 4.57E-07 9.64E-03 6.59 2. WENO Schemes P 1 (third order) h h 0 /2 h 0 /4 h 0 /8 h 0 /16 h 0 /32 h 0 /64 L 1 error 2. WENO Schemes P 1 (third order) h 2/5 1/5 1/10 1/20 1/40 1/80 L 1 error 2.73E-03 1.29 3.22 4.87E-05 4.35E-05 9.18E-02 1.05E-06 Order — 1.01E-02 3.99 4.17 4.01 3.06 4.90E-04 1.73 3.11E-06 3.63E-03 6.38 2.07 L ∞ error 8.77E-02 8.73 3.76 3.35 2.38 3.66E-01 8.51 3.71 1.41E-02 3.16 1.00E-08 Order — 4.25 4.97E-04 7.23 4.43E-04 6.14 L ∞ error 9.36E-05 2.10 L ∞ error 2. Uniform Meshes.16E-02 8.68E-03 1.64 2.85E-04 4.99E-07 7.1 Accuracy for 2D Linear Equation.44E-04 8.60 3.32 4.00 TABLE 5.89E-08 Order — 3.14E-02 6.65 2.25 4.80E-03 8.11E-02 2.97 3.56E-08 Order — 3.07 TABLE 5.83E-09 P 2 (fourth order) Order — 3.61E-03 1.05E-04 3.83 3.63E-07 2.18E-03 2.12E-05 7.47 3.80 L 1 error 4.36 4.3 Accuracy for 2D Burgers’ Equation.24 4.44 L 1 error 1.77E-03 4.55E-04 2.24E-03 1.35 4.65E-02 2.94E-05 2.79E-01 8.95E-08 4.30E-07 1.4 Accuracy for 2D Burgers’ Equation.18E-06 7.30E-01 1.20E-02 2. Non-uniform Meshes.35 L 1 error 1.12E-06 5.53E-02 2.83E-04 3.

The accuracy results are shown in Table 5.12) in a domain of [−1. 2D Burgers’ equation: t = 5/π 2 .2 is the result for h = h 0 /16 of a non-uniform mesh. and periodic boundary condition is used in the y -direction. Figure 5. The pictures shown FIG.5.2. 5. 5. The velocity in the y -direction is zero. a uniform mesh.1.3.6 for the non-uniform meshes.3.5. We can see that the shock transitions are sharp and non-oscillatory. non-uniform mesh.3 and 5.WENO SCHEMES ON TRIANGULAR MESHES 113 FIG. Figures 5.2] with a triangulation of 101 vertices in the x -direction and 11 vertices in the y -direction.4 are the results for the long-time evolution of the vortex. uniform mesh. . 1] × [0. Riemann Problems of Euler Equations The two-dimensional triangular WENO methods are applied to one-dimensional shock tube problems. 3. EXAMPLE 5. We consider the solution of the Euler equations (3.and fourth-order WENO methods. Part of the mesh is shown in Fig. These results are similar to those obtained with the linear schemes in Figs. 0. 5. 2D Burgers’ equation: t = 5/π 2 . Two-dimensional vortex evolution problem as defined in Example 3.5 for the uniform meshes and in Table 5. We can see that the correct orders of accuracy are again obtained by the third. 5.4 and 3.2.

40 is shown in the first three plots in Fig. 5. 0.32E-04 1.78E-03 6. 0. 3. The second test case is the Riemann problem proposed by Lax [21]: (ρ L .38 3.50 1. and also a less oscillatory result from the characteristic version of the fourth-order scheme over the component version.89 3.69 5.30E-02 2.79E-04 6.44 L ∞ error 3.80E-03 2. u L . Uniform Meshes.19 L 1 error 1.69 3.5.47E-04 8.5 Accuracy for 2D Euler Equation of Smooth Vortex Evolution. We can observe a better resolution of the fourth-order scheme over the third-order one.64E-03 6.27E-01 6.6. (ρ R .80 4.21 2.85 L 1 error 1.66E-04 Order — 0.91E-04 1.84E-03 8.28E-02 3. WENO Schemes P 1 (third order) h 1 1 /2 1 /4 1 /8 1/16 1/32 L 1 error 1.66E-08 P 2 (fourth order) Order — 2.45E-02 3. p R ) if x ≤ 0 if x > 0.49 3.84E-02 2.1).86 . p L ) = (0.571). p R ) = (0.83E-09 P 2 (fourth order) Order — 2.31 4.42 3. WENO Schemes P 1 (third order) h h 0 /2 h 0 /4 h 0 /8 h 0 /16 h 0 /32 h 0 /64 L 1 error 2.10 2.114 HU AND SHU TABLE 5.125. (ρ R .698.57E-03 5.92E-06 2. u R .01E-02 2. Density at t = 0.26E-04 1.96E-07 Order — 2. 0.14E-03 5. p ) = (ρ L .48 4.71 1. p L ) (ρ R .12E-02 1. u . The first test case is Sod’s problem [33].95E-01 2.14E-01 3.6 Accuracy for 2D Euler Equation of Smooth Vortex Evolution.86 2. Density at t = 0.95E-06 1.09E-05 3. u R .43E-02 6.72 4. 5.07 5. u L .38 3.62 L ∞ error 2.19E-04 Order — 0.96E-04 4. The initial data are (ρ L .74E-05 7. p L ) = (1.10E-06 Order — 0. 0. u R .02E-02 5.87E-02 1.18 2.12E-04 1.09 4.55 1. 0.66 below are obtained by extracting the data along the central cut line for 101 equally spaced points.91 3.05E-01 4.05E-02 8.16E-05 Order — 0.50E-03 1.70 L ∞ error 2.03E-07 7.6.09E-01 6.86 4. 0.73 1.76 4.33E-01 2.73 1.06 1.85E-02 3.78E-07 Order — 2.74 2.50 L ∞ error 2.445.72 3. TABLE 5. p R ) = (0.64 2.97E-05 6.91 4. We consider the following Riemann-type initial conditions: (ρ.37E-02 3.26 is shown in the last three plots in Fig.76E-07 1. u L .89 1.528).28 4.29E-03 1. 1). Non-uniform Meshes.

6 length units from the left end of the tunnel. 5.3. instead we adopt the same technique as the one used in [7]. The corner of the step is a singularity point.4. Initially. We use the third-order scheme for this problem. the triangle size away from the corner is roughly equal to a rect1 . Four meshes have been used. The step is 0.7. we do not modify our method near the corner. For the angular element case of x = y = 40 second mesh.WENO SCHEMES ON TRIANGULAR MESHES 115 FIG. 2D vortex evolution: fourth-order WENO scheme. a right-going Mach 3 flow is used. .12) in a wind tunnel of 1 length unit wide and 3 length units long. see Fig. 5. namely refining the mesh near the corner and using the same scheme in the whole domain.2 length units high and is located 0. while it is one-quarter that near the corner. 5. 5. For first mesh. In this paper. A Mach 3 Wind Tunnel with a Step This problem is from [36]. We solve the Euler equations (3. 2D vortex evolution: third-order WENO scheme. Reflective boundary conditions are applied along the walls of the tunnel. and inflow and outflow boundary conditions are used at the entrance and the exit. but it is 1 away one-eighth that near the corner.3. the triangle size away from the corner is the same as in the first mesh. [36] uses an assumption of nearly steady flow in the region near the corner to fix this singularity. The third mesh has a triangle size of x = y = 80 FIG.

Mesh for the Riemann problems. Riemann problems of Euler equations: density.6. 5. .5. 5.116 HU AND SHU FIG. FIG.

from the corner. Triangulations for the forward step problem: part near the corner. and it is one-half that near the corner. It is clear that with more triangles near the corner the artifacts from the singularity decrease significantly. The last mesh has a triangle size 1 of x = y = 160 away from the corner.0.WENO SCHEMES ON TRIANGULAR MESHES 117 FIG.8 is the contour picture for the density at time t = 4. 5. Figure 5.7. . and it is one-quarter that near the corner.

118 HU AND SHU FIG.15. .32 to 6.8. 5. Forward step problem: 30 contours from 0.

CONCLUDING REMARKS We have presented the development of third. The reflective boundary condition is used at the wall.and fourth-order WENO schemes based on linear and quadratic polynomials for 2D triangle meshes. the exact postshock while for the rest of the bottom boundary (the part from x = 0 to x = 1 6 condition is imposed.9) is also from [36]. 1]) and a blow-up region around the double Mach stems.11). Double Mach Reflection This problem (Fig. 2] × [0. x = y = 100 roughly equal to rectangular element cases of x = y = 50 1 1 (Fig. We refer to [7] for similar results obtained with discontinuous Galerkin methods.10). 5.14 and 5. 4] × [0. Triangulation for the double Mach reflection. 5. We test both the third.2. we use uniform triangular 1 1 and x = y = 400 . Accuracy and stability issues are considered in the design of the schemes and verified numerically. All pictures are the density contours with 30 equally spaced contour lines from 1. 1]. For the third-order scheme. 5.12) in a computational domain of [0. The results shown are at t = 0.9.and the fourth-order schemes. 5. Numerical examples show the improvement of resolution using high-order schemes. 6.13). and the characteristic version is much less oscillatory than the component version for the fourth-order scheme.12 and 5. A reflecting wall lies at the bottom of the domain . they are 1 1 (Fig. For the cases of x = y = 200 of whole region ([0. sizes. the flow values are set to describe the exact motion of the Mach 10 shock. and x = y = 400 (Figs. starting from x = 1 6 6 ◦ making a 60 angle with the x -axis.15 provide further examples for the fourth order. Four triangle sizes are used. 3] × [0. ). 5.4. We solve the Euler equations (3. We can clearly see that the fourth-order scheme captures the complicated flow structure under the triple Mach stem much better than the third-order scheme.10–5.15).WENO SCHEMES ON TRIANGULAR MESHES 119 FIG. x = y = 200 (Figs. we use both uniform triangular mesh (equilateral triangles) and locally refined triangular mesh (the refined region has the above triangle 1 locally). Fig.5. . y = 0. Initially a right-moving Mach 10 shock is located at x = 1 . At the top boundary. 1] of such a mesh of x = y = 50 Figures 5. 5.5 to 21. 5.9 shows the region [0. 5. respectively. we present both the picture mesh only.

5.120 HU AND SHU FIG. t = 0. Double Mach reflection: h = 1 50 .10. .2.

Double Mach reflection: h = 1 100 . .11. 5.WENO SCHEMES ON TRIANGULAR MESHES 121 FIG. t = 0.2.

Double Mach reflection: h = 1 200 .12. t = 0.122 HU AND SHU FIG. 5.2. .

.2 (blow-up). Double Mach reflection: h = 1 200 .13. 5.WENO SCHEMES ON TRIANGULAR MESHES 123 FIG. t = 0.

Double Mach reflection: h = 1 400 . 5.14. . t = 0.2.124 HU AND SHU FIG.

WENO SCHEMES ON TRIANGULAR MESHES 125 FIG. . t = 0. Double Mach reflection: h = 1 400 .2 (blow-up).15. 5.

A. 39 (1994). 91-76 (1991). 3. Comput. Abgrall. 10. J. Phys. Shu. Saint-Etienne. and C. Comput. P. 15. Shu. Y. J. J. Friedrichs. P. 127. 115. Harten. X. Comput..-W..-W. Bianco. Uniformly high order essentially non-oscillatory schemes. and W. 2829 (1992). E. 266 (1995). E. D. and G. G. J. 202 (1996). Chan. On essentially non-oscillatory schemes on unstructured meshes: Analysis and implementation. III. 141. 18. Frederickson. Solution of the hydrodynamic device model using high order non-oscillatory shock capturing algorithms. 110. High order central schemes for hyperbolic systems of conservation laws. Computer-Aided Design Integrated Circuits Syst. Comput. 7. SIAM J. Central WENO schemes for hyperbolic systems of conservation laws. Shu. Casper. Atkins. 16. Ladeinde. J. 199 (1998). Liu. Comput. 27 (1996). Cole. G. Weighted essentially non-oscillatory schemes. Preliminary results on the extension of ENO schemes to two dimensional problems. 22. Adams and K. Lax. 127. F. Anal. White (SpringerVerlag. Phys. Phys. J. O. B. Sethian. 917 (1995). Harten. Interaction of a shock with a longitudinal vortex.-W. 30. AIAA Paper No. Shu. Phys. Jerome and C. W. Phys. Osher.-W. Computer-Aided Design Integrated Circuits Syst. in IMA Volumes in Mathematics and Its Applications. and S. 200 (1994). AIAA J. Jiang and C. Fronts propagating with curvature-dependent speed: Algorithms based on Hamilton– Jacobi formulation. Harten and S. High order solution of the Euler equations on unstructured grids using quadratic reconstruction. G. Vol. to appear. 337. . Pure Appl. Cai. 106. 15 (1996). B. Shariff. Multi-dimensional ENO Schemes for General Geometries. Berlin/New York. Modelling Numer. Energy models for one-carrier transport in semiconductor devices. Wong. Erlebacher. Phys. IEEE Trans. ICASE Report No. IEEE Trans. Shu. 59. O’Brien. Comput. S. J. S. 79. Jerome. J.126 HU AND SHU REFERENCES 1. edited by W. Fatemi. 7. Russo. Jerome and C. X. R. The Runge–Kutta discontinuous Galerkin method for conservation laws. S. Osher. Advection by polytropic compressible turbulence. 2848 (1995). 12 (1988). 21. Fluids 7. in Proceedings of the International Conference on Hyperbolic Problems. Comput. Hussaini. Cockburn and C. J. 5. 23. Commun. Comput. 24. 13. A. A high-resolution hybrid compact-ENO scheme for shock-turbulence interaction problems. Efficient implementation of weighted ENO schemes. Phys. Sci. 11. 114. Phys. Puppo. G. 144. Dolezal and S. 14. and G. Casper and H. 14. Coughran. J. Phys. J. to appear. Phys. 185. Liu. Shu. D. Math. A. 45 (1994). Osher. Phys. 129 (1997). J. 231 (1987). Barth and P. 19. E. Finite-volume implementation of high-order essentially non-oscillatory schemes in two dimensions. N. 1994). 159 (1954). Phys. V. 20. 232 (1991). and J.-W. Weak solutions of nonlinear hyperbolic equations and their numerical computation. J. A finite-volume high-order ENO scheme for two dimensional hyperbolic systems. Puppo. Fluid Mech. Phys. 194 (1998). Comput. 9. J. and S. A numerical resolution study of high order essentially non-oscillatory schemes applied to incompressible flow. 126. Russo. Comput. 62 (1993). T. p. A. S.-D. An Eulerian approach for vortex motion using a level set regularization procedure. Chakravarthy. 6. 2.-W. J. Relativistic hydrodynamics and essentially non-oscillatory shock capturing schemes. J. J. 17. Engquist. and C. Osher. 4. 120. 90-0013. F. 1986. Comput. J. Harabetian. 10. Chakravarthy. 12. 71. Osher and J. 8. Transport effects and characteristic modes in the modeling and simulation of submicron devices. Shu. E and C. Math. and T. Levy. Lloyd. J.-W. Comput. Multidimensional systems. Comput. Weighted essentially non-oscillatory schemes for the interpolation of mean values on unstructured grids.

Geometric shock-capturing ENO schemes for subpixel interpolation. 357 (1981). 33. Phys. 34. 59. Fluid Dynamics. A. G. D. Editor. Quarteroni. edited by A. Press. 140.-W. and difference schemes. High order ENO schemes applied to twoand three-dimensional compressible flow. to appear). 9. T. 1996). P. New York. Comput. Appl. Phys. Woodward and P. Shu. 29. CIME subseries (Springer-Verlag. Lecture Notes in Mathematics. Comput. Whitaker. Quart. Comput. Methods Appl. On the construction of essentially non-oscillatory finite volume approximations to hyperbolic conservation laws on general triangulations: Polynomial recovery. Efficient implementation of essentially non-oscillatory shock capturing schemes. Osher. 129 (1994). 43. Sonar. Walsteijn. Mech. Sod. Berlin/New York. 83. C. 45 (1992). 157 (1997). Kimia. Appl. J.-W. Phys. Math. Shu and S. 27. Siddiqi. High order essentially non-oscillatory scheme for viscoelasticity with fading memory. Zeng. 459 (1997). and Material Science. Phys. J. 31. Essentially non-oscillatory and weighted essentially non-oscillatory schemes for hyperbolic conservation laws. C. The numerical simulation of two-dimensional fluid flow with strong shocks. in Advanced Numerical Approximation of Nonlinear Hyperbolic Equations. and S. ICASE Report 97-65. Level Set Methods: Evolving Interfaces in Geometry. Engrg. Computer Vision. 36. Colella. . Sethian. J. and C. J. C. Comput. 54. K. T. 55. F. 439 (1988). Shu. Osher. Math. C. 35. Phys. B. Erlebacher. 28. 32. Robust numerical methods for 2D turbulence. Osher. J. G. Efficient implementation of essentially non-oscillatory shock capturing schemes. 1 (1978).-W. Shu and S. 26. 77. II.-W. Approximate Riemann solvers. J. 30. Shu and Y. accuracy and stencil selection.-W. Numer. A survey of several finite difference methods for systems of nonlinear hyperbolic conservation laws. computation and curve evolution.WENO SCHEMES ON TRIANGULAR MESHES 127 25. Roe. 115 (1984). Cambridge Monographs on Applied and Computational Mathematics (Cambridge Univ. Zang. 278 (1997). J. Shu. Comput. parameter vectors. Phys. Vision Graphics Image Processing Graphical Models Image Process. Comput. 27. Comput. P. 32 (1989). C. Comput.-W. 114. L.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->