Numerical Simulation of 2D Square Driven Cavity Using Fourth Order Compact Finite Di erence Schemes

Jun Zhangy Department of Computer Science, University of Kentucky, 773 Anderson Hall, Lexington, KY 40506-0046, USA April 6, 2000
Abstract
Fourth order compact nite di erence schemes are employed with multigrid techniques to simulate the two dimensional square driven cavity ow with small to large Reynolds numbers. The governing Navier-Stokes equation is linearized in streamfunction and vorticity formulation. The fourth order compact approximation schemes are coupled with fourth order approximations for velocities and vorticity boundaries. Numerical solutions are obtained for square driven cavity ow at high Reynolds numbers and are compared with solutions obtained by other researchers using other approximation methods.
Key words:

Incompressible Navier-Stokes equations, square driven cavity problem, convection di usion equation, multigrid method, fourth order compact discretization.

Mathematics Subject Classi cation: 65N06, 65N22, 65F10.

1 Introduction
The Navier-Stokes equations have been used to model uid dynamics phenomena describing ows of an incompressible viscous uid. These equations are highly nonlinear and are very di cult to solve, especially when the approximate solutions are required to have a high accuracy. A Navier-Stokes equation may be linearized in streamfunction and vorticity formulation 6, 18]. Computing accurate solution of a convection di usion equation thus becomes an issue. We begin our discussion with some stable and accurate numerical methods for solving the 2D convection di usion equation with high Reynolds numbers. The general convection di usion equation is of the form @ 2 U (x; y) + @ 2 U (x; y) + p(x; y) @U (x; y) + q(x; y) @U (x; y) = f (x; y); (x; y) 2 ; (1)

@x2

@y2

@x

@y

This paper has been accepted for publication in Computer and Mathematics with Applications. E-mail: jzhang@cs.uky.edu. URL: http://www.cs.uky.edu/ jzhang

where p(x; y) and q(x; y) are continuous functions of variables x and y. The magnitude of p and q simulates the Reynolds numbers (denoted by Re) which determines the convection strength of a ow. is a convex 2D domain and @ is the boundary of . This equation appears in many transport problems. In the current situation it results from numerical solution of the steady state incompressible Navier-Stokes equations 6]. Suppose Eq. (1) is discretized by some nite di erence scheme and a linear system

A u =f
h h

(2)
h

is obtained, where h is the uniform grid spacing of the discretized space . The linear system (2) is usually of very large dimension. For such large sparse linear systems, the use of direct solution methods based on Gaussian elimination incurs prohibitive memory and CPU cost and iterative solution methods are usually considered. The performance of classical iterative methods such as Jacobi and SOR is sensitive to the meshsize h, the type of boundary conditions and other factors. The coe cient matrix A is nonsymmetric and far from diagonally dominant if Re is large 22]. This property adds further di culty to classical iterative methods. Eq. (1) may be discretized by the central di erence scheme (CDS), the resulting linear system (2) is a 5 point formula with a truncation error of order h2 . In the case of CDS, classical iterative methods for solving the resulting linear system may not converge when the convective terms dominate and the cell Reynolds number (Re) is greater than a certain constant. For this reason, the upwind di erence scheme (UPS) has been used for many years despite it is only rst order accurate. Due to its importance in practical applications, various attempts have been made to solve the convection di usion equation and the incompressible Navier-Stokes equations with iterative (especially multigrid) methods 3, 16, 17, 18]. For example, de Zeeuw 20] developed a black-box multigrid solver with some matrix-dependent prolongations and restrictions for solving convection dominated problems. This and other methods have demonstrated the e ciency of multigrid techniques in solving convection dominated problems discretized by UPS. Recently, there has been increasing interest in developing fourth order compact schemes for solving Eq. (1) and the incompressible Navier-Stokes equations with large Reynolds numbers, see, e.g., 2, 4, 6, 9, 10, 14]. These schemes are somewhat similar and numerical results reported by these researchers are not very di erent. There are at least three advantages shared by these schemes:
h

1. Unconditional stability: although the coe cient matrices are no longer diagonally dominant for large Re, the schemes have been shown numerically stable for large Reynolds numbers 4, 8, 23]; 2. High accuracy: it has been shown that these schemes do produce numerical solution of fourth order accuracy for the convection di usion equation 4, 8, 21, 23] and of high accuracy for the Navier-Stokes equations with small to medium Re 1, 6, 10]; 2

3. Easy boundary treatment: since the computational stencil involves only the nearest 9 grid points, the schemes are of compact type and no special formula is needed for approximating grid points near the boundaries of a regularly structured domain. However, until recently, computational advantages of these compact schemes have not been fully investigated. For example, it is not clear if these schemes can be used to solve incompressible Navier-Stokes equations of very large Re because of the limitations of available computer power and of the di culty of solving the resulting linear system (lack of diagonal dominance) with traditional SOR type iterative methods. To fully investigate the potential of using the fourth order compact schemes for solving Navier-Stokes equations with large Re, nontraditional iterative methods seem necessary. One class of promising methods are multigrid techniques which have been successfully used with the rst and second order discretization schemes for solving problems in computational uid dynamics (including the driven cavity problem) (see, e.g., 3, 17, 19]). A preliminary investigation on combining the fourth order compact schemes with multigrid techniques was made by Altas and Burrage 1] for di usion dominated ow problems. Multigrid solution and accelerated multigrid solution methods with the fourth order compact schemes for solving convection dominated problems are relatively new 8, 23, 24] The current paper is to use multigrid techniques to evaluate the fourth order compact schemes in solving steady state incompressible Navier-Stokes equations for large Re, exempli ed by solving the square driven cavity problem. This follows from Gupta's work 6] using SOR iterative method with these schemes to solve driven cavity problem for Re 2; 000 and Zhang's work 21] employing the fourth order compact scheme and multigrid techniques to compute high accuracy solution of the convection di usion equation with very large Reynolds numbers. We present the 9 point compact nite di erence discretization scheme for Eq. (1) and discretization schemes for the streamfunction and vorticity formulation of the incompressible Navier-Stokes equations in Section 2. The multigrid solvers for the convection di usion (and the Poisson) equation and for the incompressible Navier-Stokes equations are outlined in Section 3. In Section 4, we solve the driven cavity problem and compare our numerical results with those obtained by other researchers using other methods. Concluding remarks are given in Section 5.

2 Fourth Order Finite Di erence Schemes

The approximate value of a function U (x; y) at a mesh point (x; y) is denoted by U0 . The approximate values at its eight immediate neighboring points are denoted by U ; i = 1; 2; : : : ; 8. The computational stencil of the 9 compact grid points is shown as
i

0u u u 1 6 2 5 B A: @ u3 u0 u1 C
u7 u4 u8
3

The discretized values of p ; q and f ; i = 0; 1; : : : ; 4; have their obvious meanings. The compact nite di erence formula for a reference point (x; y) involves the nearest eight neighboring mesh points with the mesh spacing h and is given by (for detailed procedure to derive the nite di erence scheme, see 4]):
i i i

8 X
i

=0

2 h3 p (f f ) + q (f f )]; 8 f + f + f + f + f ] + U =h 0 1 2 3 4 0 2 4 2 4 0 1 3
i i

(3)

where the coe cients ; i = 0; 1 : : : ; 8; are given as

1

h 2 = 4+ h 4 4p0 + 3p1 p3 + p2 + p4 ] + 8 4p0 + p0 (p1 p3 ) + q0 (p2 p4 )]; h h2 2 2 = 4 + 4 4q0 + 3q2 q4 + q1 + q3 ] + 8 4q0 + p0 (q1 q3 ) + q0 (q2 q4 )]; h 2 = 4 h 4 4p0 p1 + 3p3 + p2 + p4 ] + 8 4p0 p0 (p1 p3 ) q0 (p2 p4 )]; h h2 2 4 = 4 4 4q0 q2 + 3q4 + q1 + q3 ] + 8 4q0 p0 (q1 q3 ) q0 (q2 q4 )];
3 5 2

h h = 1+ h 2 (p0 + q0 ) + 8 (q1 q3 + p2 p4 ) + 4 p0 q0 ; h h h2 6 = 1 2 (p0 q0 ) 8 (q1 q3 + p2 p4 ) 4 p0 q0 ; h h = 1 h 2 (p0 + q0 ) + 8 (q1 q3 + p2 p4 ) + 4 p0 q0 ; h h2 h 8 = 1 + 2 (p0 q0 ) 8 (q1 q3 + p2 p4 ) 4 p0 q0 ; 2 20 + h2 (p2 0 = 0 + q0 ) + h(p1 p3 ) + h(q2 q4 )]:
7 2

Results of numerical experiments in 4] show that SOR iterative method with this scheme converges for any values of p(x; y) and q(x; y) tested. When Re 0, Eq. (1) reduces to the Poisson equation, and Eq. (3) reduces to the well known (simpler) Mehrstellen formula 5]: 1 h2 8f + f + f + f + f ]: (4) 4 U1 + U2 + U3 + U4 ] + U5 + U6 + U7 + U8 20U0 = 2 0 1 2 3 4 Multigrid applications of the Mehrstellen formula have been investigated by Scha er 13], Gupta, Kouatchou and Zhang 7]. The Navier-Stokes equations representing the two dimensional steady ow of an incompressible viscous uid are given in streamfunction and vorticity formulation as follows 3, 6, 18]:
@

@x

(x; y); ( ) + 2 ( ) Re u(x; y) ( ) + v(x; y) ( ) i = 0; 2 2 u(x; y) = ( ) ; v(x; y) = ( ) :

@

@x2

( )+
x;y @

x;y

x;y

@y 2

( )
x;y

x;y

x;y

@y

@x

@y

x;y

x;y

@y

@x

(5) (6) (7)

Here is the streamfunction, the vorticity; u and v are the velocities; Re is the nondimensional Reynolds number 6]. The streamfunction (5) is a Poisson equation and the fourth order approximation is given by Eq. (4) with U = and f = . The vorticity equation (6) is a special case of the convection di usion equation (1) and the fourth order approximation in this case may be obtained with U = , f = 0 and p(x; y) = Re u(x; y), q(x; y) = Re v(x; y) in Eq. (3). The velocities u, v at a grid point (x; y) are calculated from the discrete approximation of Eq. (7). It has been shown 6] that it is bene cial for both convergence and accuracy to use the fourth order approximations for the velocities. In particular, Gupta 5] derived some high accuracy compact approximations for the gradients of the solution of the Poisson equation. Since the streamfunction equation (5) is a Poisson equation in , high accuracy approximations for the gradient and can be obtained from 5], and the corresponding fourth order compact approximations for the velocities are given as (also see 6]):
x y

u0 = ( v0 = (

2 3

4 )=3h + ( 5 1 )=3h ( 5

6 6

7 8 )=12h + h( 2 7 + 8 )=12h + h( 3

4 )=12; 1 )=12:

(8)

3 Solution Method
In this paper, we are mainly concerned with the accuracy of the computed solution, not the computational e ciency of the solution method that we used. We just outline our solution method in this section and point out relevant references in this direction 8, 18, 19, 20, 21, 23, 27]. Our multigrid solution method for the convection di usion equation (1) (the Poisson equation (4) is a special case) is as follows: 1. Start on the ne grid with some initial guess and perform 1 red-black Gauss-Seidel relaxation sweeps. 2. Compute residuals on the ne grid points which are common to the coarse grid and multiply the residuals by a scaling factor and inject the scaled residuals to the coarse grid 23, 25, 26]. 3. Perform multigrid cycles on the coarse grid. 4. Interpolate the coarse grid correction to the ne grid by bilinear interpolation and add the correction to the ne grid solution. 5. Perform 2 red-black Gauss-Seidel relaxation sweeps on the ne grid. If = 1, the multigrid algorithm is called a V cycle. It is a W cycle for = 2. 1 and 2 are the numbers of presmoothing and postsmoothing sweeps. 5

For the streamfunction equation (5), a V(1,1) cycle algorithm is su cient to obtain accurate solution with acceptable convergence and is more cost e ective than a W(1,1) cycle algorithm. The smoother for the streamfunction is the red-black Gauss-Seidel without a parameter. For the vorticity equation (6), we use a W(1,1) cycle algorithm. The smoother is the red-black Gauss-Seidel with a damping parameter !i 2 (0; 1) (all damping factors were found by numerical experiments). The Navier-Stokes equations (5) { (7) may be solved by a nested inner-outer iteration procedure (see 6]) with di erent multigrid cycling algorithms being applied to the streamfunction equation (5) and the vorticity equation (6). While it is sometimes advantageous to solve Eqs. (5) { (7) simultaneously to maintain the physical coupling between them 3, 18], the relaxation methods (such as the coupled strongly implicitly (CSI) procedure 3] and incomplete LU decomposition 18]) used to accomplish this coupling are expensive and di cult to implement, comparing with the point Gauss-Seidel relaxation. Since our goal is to evaluate the usefulness of the fourth order compact schemes for solving Navier-Stokes equations with large Re, we used the simplest point Gauss-Seidel relaxation.

4 Solution of Square Driven Cavity

The steady ow of an incompressible viscous uid in a square cavity = (0; 1) (0; 1) has been used for many years as the model problem by researchers to test their new numerical schemes and solution methods 1, 3, 6, 17, 18], although there is no analytical solution available and there are singularities at two of its corners. Highly accurate benchmark solutions are available in the literature. In particular, Ghia et al. 3] used multigrid techniques and grid points of 256 256 to compute numerical solutions for 100 Re 10; 000. Their solutions have been considered very accurate not only because of the small grid spacing employed, but also as a result of the fact that and were computed directly from the streamfunction and vorticity formulation (5) { (7). The ow is induced by the sliding motion of the top wall (y = 1) from right to left and is described by the Navier-Stokes equations (5){(7). The boundary conditions are those of no slip: on the stationary walls u = 0 and v = 0; on the sliding wall u = 1 and v = 0 (see Fig. 1). In order to solve the driven cavity problem, we replace the Navier-Stokes equations (5){(6) by the nite di erence approximations given in Eqs. (3) and (4) respectively. The velocities, de ned in Eq. (7), are calculated by using the fourth order approximations (8). The unit square is covered by a grid of uniform meshsize h = 1=N . The discrete approximations (3) and (4) are written at each of the (N 1)2 interior grid points. Zero values are prescribed for on the boundary. The usual approximations for vorticity on the boundary are the Jensen's formulas (see 6, 12]), which have a local truncation error of second order. Fourth order approximations could also be de ned to obtain boundary values of . In particular, some fourth order approximations analogous to the Jensen's

y
(0,1)

u = 1; v = 0
Primary Vortex VC

(1,1)

v=0 u=0
UVC DVC

v=0 u=0

(0,0)

u = 0; v = 0
Figure 1: Driven cavity problem.

(1,0)

- x

> 2 4 20 0 )=7h2 ; > = 2 2 4 20 0 )=7h ; > 2 3 20 0 )=7h2 ; > 2 2 1 2 3 20 0 )=7h : ; (9) Some discussions about the accuracy and performance of numerical wall boundary conditions for steady 2D incompressible stream-function vorticity can be found in 15]. An inner-outer iteration procedure is employed to obtain the numerical solutions (see 6]). At each outer iteration, the linear system from the discrete streamfunction equation (5) is solved by a multigrid V(1,1) cycle algorithm. We then compute the velocities using the fourth order formulas (8) and evaluate the boundary conditions using the fourth order formulas (9). After that, we solve the discrete vorticity equation (6) using a multigrid W cycle algorithm with a relaxation parameter !i 2 (0; 1). The outer iteration processes for the streamfunction equation and the vorticity equation are also damped after each iteration using di erent damping factors !s 2 (0; 2) and !v 2 (0; 2) to give the new iterates. Once again these damping factors were found by numerical experiments. At each inner iteration of the streamfunction equation, one or two multigrid V(1,1) cycles are applied; at each inner iteration of the vorticity equation, one multigrid W(1,1) cycle is applied. Since the streamfunction equation converges very fast and the solution of a Poison equation is much cheaper to compute than that of a convection di usion equation, the major cost of each iteration step is dominated by solving the vorticity equation. We
2 2 1

formulas were obtained by Altas and Burrage 1]: On stationary wall x = 0 : 3 = ( 6h2 0 + h2 1 2 On stationary wall x = 1 : 1 = ( 6h2 0 + h2 3 2 On stationary wall y = 0 : 4 = ( 6h2 0 + h2 2 2 On sliding wall y = 1 : 2 = (24h 6h2 0 + h2 4

have found that, at each outer iteration step, there is no need to solve each inner iteration to a higher accuracy. One or two multigrid cycles are enough and cost e ective. The inner-outer iteration process for our multigrid solver may be summarized as following: Set initial guess as 0 for all values except boundary values of the known velocities. For k = 0; 1; 2;

; do

Step 1: Compute the right-hand side of the streamfunction; Step 2: Solve approximately the streamfunction equation (5) by performing one or two V(1,1) cycles; Step 3: Compute the di erence between the current and the previous values of the streamfunction; Step 4: Damp the values of the streamfunction using a damping parameter !s 2 (0; 2); Step 6: Compute the vorticity boundary values by using the fourth order boundary approximations; Step 7: Solve approximately the vorticity equation (6) by performing one W(1,1) cycle and by using a relaxation parameter !i 2 (0; 1); Step 8: Compute the di erence between the current and the previous values of the vorticity; Step 9: Damp the values of the vorticity by using a parameter !v 2 (0; 2); Step 10: Check convergence, if both di erences of the current and previous values of the streamfunction and vorticity computed from Steps 3 and 8 are less than a prescribed tolerance, then stop; otherwise go to Step 1 and begin next outer iteration. We solved the driven cavity problem for 100 Re 7; 500. For each Re, we give location coordinates of the primary vortex and the corresponding streamfunction and vorticity values. The problem with same Re was solved several times using di erent meshsizes to investigate what is the coarsest meshsize producing acceptable solution for a given Re. The solutions of small Re were primarily used to build con dence and to verify that our method produced accurate solutions. The iterations were started with zero initial data and were terminated when the maximum di erence between successive approximations of both and were smaller than 10 4 . The computations were carried out on an SGI workstation using Fortran 77 programming language in double precision.

1/16 1/32 1/64 1/128

0.102901 0.103354 0.103501 0.103511

3.307272 3.282395 3.108643 3.168745

Location Di . in (0.375000, 0.750000) 0.59% (0.375000, 0.750000) 0.15% (0.390625, 0.734375) 0.01% (0.382813, 0.734375) 0.00%

Di . in 4.37% 3.59% 1.90% 0.00%

Table 1: Values and locations of the primary vortex for Re = 100 using di erent discretizations.

4.1 Comparison with benchmark solutions

Unless otherwise indicated explicitly, we compare our results with those obtained by Ghia et al. 3] as the benchmark solutions. Our problem was set up slightly di erent from that of Ghia et al. 3] and the u velocity at the top wall (y = 1) is di erent due to the fact that the ow in our problem is induced by the sliding motion of the top wall from right to left 6]. It was induced from left to right in the model problem solved by Ghia et al. in 3]. The computed value at a grid point (x; y) listed in our tables should be compared with that at the point (1 x; y) in tables of 3]. We rst solved the model problem for Re = 100 with h = 1=16; 1=32; 1=64; 1=128 and compare our results with those of Ghia et al. with h = 1=128. The numerical values are given in Table 1. It can be seen that, if a 5% departure from the benchmark solution is acceptable as the engineering accuracy, our method with h = 1=16 can produce acceptable results, while Ghia et al. used much ner discretization (h = 1=128). Note that when the mesh is re ned, the accuracy of our solution is increased rapidly. When Re increases to 1; 000, results in Table 2 show that ner mesh is needed to produce accurate solution. However, we can see that with only a quarter of the number of equations used by Ghia et al., our method produced high accuracy solutions. When Re 2; 000, Gupta 6] reported slow convergence when the fourth order compact schemes were used with SOR iteration. It is shown in Table 3 that our multigrid algorithm can compute very accurate solution for Re = 3; 200 with h = 1=128. With Re = 5; 000, Ghia et al. had to use h = 1=256 to compute accurate solution, we found (see Table 4) that our fourth order multigrid method can compute solution to comparable accuracy using h = 1=128. For Re = 7; 500, with h = 1=128, our method was still able to yield solution accurate enough (5% error) to compare with the benchmark solution of Ghia et al. 3] using h = 1=256.

4.2 Comparison of high accuracy solutions

In Section 4.1, the di erences for the vorticity values are usually larger than that of the streamfunction values for large Re, as indicated by data in Table 3 for Re = 3; 200 and in 9

1/16 1/32 1/64 1/128

0.091208 0.106723 0.117386 0.118806

3.073332 2.026724 2.056608 2.066777

Location Di . in (0.437500, 0.750000) 23.36% (0.468750, 0.593750) 13.32% (0.468750, 0.562500) 1.36% (0.468750, 0.562500) 0.17%

Di . in 48.57% 2.02% 0.58% 0.09%

Table 2: Values and locations of the primary vortex for Re = 1; 000 using di erent discretizations.

1/64 1/128

0.109094 0.120157

1.893534 1.948934

Location Di . in (0.484375, 0.578125) 9.95% (0.484375, 0.539063) 0.18%

Di . in 2.93% 4.01%

Table 3: Values and locations of the primary vortex for Re = 3; 200 using di erent discretizations.

1/128

0.118121

1.906214

Location Di . in (0.484375, 0.539063) 0.71%

Di . in 2.48%

Table 4: Values and locations of the primary vortex for Re = 5; 000 using h = 128.

1/128

0.114338

1.875897

Location Di . in (0.492188, 0.554688) 4.28%

Di . in 0.21%

Table 5: Values and locations of the primary vortex for Re = 7; 500 using h = 1=128.

10

Authors Current author Current author Li, Tang & Fornberg 10] Nishida & Satofuka 11] Ghia, Ghia & Shin 3]

0:120157 0:121659 0:120529 0:121154 0:120377

1:948934 1:959487 1:94286 1:95078 1:98860

Meshsize Method Order 128 128 4th-order 256 256 4th-order 128 128 4th-order 128 128 6th-order 128 128 2nd-order

Table 6: Comparison of high accuracy solution for Re = 3; 200. Table 4 for Re = 5; 000. This may be caused by the inaccuracy of the benchmark solution for large Re. As indicated in 11] a higher order of accuracy in space is necessary for (time dependent) high Reynolds number simulations. In Table 6, we give some recently available high accuracy results from several higher order methods for Re = 3; 200 and compare the results with those obtained by our scheme with h = 1=128 and h = 1=256. It can be seen from Table 6 that Ghia et al.'s solution is relatively inaccurate compared with the high order, high accuracy solution recently available in the literature. In particular, the vorticity value is shown to have a relative error of 1:49% with respect to our very accurate solution computed by using our fourth order algorithm with h = 1=256. In contrast, our solution with h = 1=128 has only a relative error of 0:54% in vorticity value with respect to our very accurate solution, but has a relative di erence of 4:01% with respect to Ghia et al.'s solution. It can be observed that the solution from our fourth order compact algorithm is comparable with that of Li et al.'s fourth order compact scheme 10], the relative errors in both streamfunction and vorticity values are less than 0:32%. Furthermore, the solution from our fourth order method using h = 1=256 is comparable with Nishida and Satofuka's sixth-order spectral method 11] using h = 1=128 with a di erence of 0:45%. This comparison again supports the claim made by Nishida and Satofuka 11] that higher order algorithm is necessary (at least bene cial) for high Reynolds number computations. The streamlines and the vorticity contours with Re = 5; 000 and 7; 500 for h = 1=128 are presented in Figs. 2 and 3. These gures are graphically comparable with the well known gures obtained by Ghia et al. 3], taking into account the di erence in the de nition of the velocity (see Section 4.1).

4.3 Solution contours

5 Concluding Remarks
Fourth order compact discretization formulas have been used in conjunction with multigrid techniques to simulate the two dimensional square driven cavity ow and may be used to 11

solve more general steady state incompressible Navier-Stokes equation. For the driven cavity problem, our computed solutions up to Re = 7; 500 with coarser discretizations are compared well with the benchmark solutions obtained by other researchers using ner discretizations.

References
1] I. Altas and K. Burrage, A high accuracy defect-correction multigrid method for the steady incompressible Navier-Stokes equations, J. Comput. Phys. 114, 227{233 (1994). 2] S. C. R. Dennis and J. D. Hudson, Compact h4 nite-di erence approximations to operators of Navier-Stokes type, J. Comput. Phys. 85, 390{416 (1989). 3] U. Ghia, K. N. Ghia and C. T. Shin, High-Re solutions for incompressible ow using the Navier-Stokes equations and a multigrid method, J. Comput. Phys. 48, 387{411 (1982). 4] M. M. Gupta, R.P. Manohar, and J.W. Stephenson, A single cell high order scheme for the convection-di usion equation with variable coe cients, Int. J. Numer. Methods Fluids 4, 641{651 (1984). 5] M. M. Gupta, A fourth-order Poisson solver, J. Comput. Phys. 55, 166{172 (1984). 6] M. M. Gupta, High accuracy solutions of incompressible Navier-Stokes equations, J. Comput. Phys. 93, 343{359 (1991). 7] M. M. Gupta, J. Kouatchou and J. Zhang, Comparison of second and fourth order discretizations for the multigrid Poisson solver, J. Comput. Phys. 132 (2), 226{232 (1997). 8] M. M. Gupta, J. Kouatchou and J. Zhang, A compact multigrid solver for the convection-di usion equation, J. Comput. Phys. 132 (1), 123{129 (1997). 9] T. Y. Hou and B. T. R. Wetton, Stable fourth order stream-function methods for incompressible ows with boundaries, (preprint, 1993). 10] M. Li, T. Tang and B. Fornberg, A compact fourth-order nite di erence scheme for the steady incompressible Navier-Stokes equations, Int. J. Numer. Methods Fluids 20, 1137{1151 (1995). 11] H. Nishida and N. Satofuka, Higher-order solutions of square driven cavity ow using a variable-order multi-grid method, Int. J. Numer. Methods Eng. 34, 637{653 (1992). 12] P. J. Roache, Computational Fluid Dynamics, Hermosa, Albuquerque, NM, 1973. 13] S. Scha er, High order multi-grid methods, Math. Comp. 43, 89{115 (1984). 12

14] W. F. Spotz and G. F. Carey, High-order compact scheme for the stream-function vorticity equations, Int. J. Numer. Methods Eng. 38, 3497{3512 (1995). 15] W. F. Spotz, Accuracy and performance of numerical wall boundary conditions for steady, 2D incompressible stream-function vorticity, Int. J. Numer. Methods Fluids, 28(4), 737{757 (1998). 16] C. P. Tzanos, Higher-order di erence method with a multigrid approach for the solution of the incompressible ow equations at high Reynolds numbers, Numer. Heat Transfer: Part B, 22, 179{198 (1992). 17] M. C. Thompson and J. H. Ferziger, An adaptive multigrid technique for the incompressible Navier-Stokes equations, J. Comput. Phys. 82, 94{121 (1989). 18] P. Wesseling, Multigrid solution of the Navier-Stokes equations in the vorticitystreamfunction formulation, in E cient Solutions of Elliptic Systems , Notes on Numerical Fluid Mechanics, Vol. 10, W. Hackbusch, Ed., Vieweg, Braunschweig, 1984, p. 145{154. 19] N. G. Wright and P. H. Gaskell, An e cient multigrid approach to solving highly recirculating ows, Computers & Fluids, 24, 63{79 (1995). 20] P. M. de Zeeuw, Matrix-dependent prolongations and restrictions in a blackbox multigrid solver, J. Comput. Appl. Math. 33, 1{27 (1990). 21] J. Zhang, Accelerated multigrid high accuracy solution of the convection-di usion equation with high Reynolds number, Numer. Methods Partial Di erential Equations, 13, 77-92 (1997). 22] J. Zhang, On convergence of iterative methods for a fourth-order discretization scheme, Appl. Math. Lett. 10 (2), 49{55 (1997). 23] J. Zhang, Multigrid Acceleration Techniques and Applications to the Numerical Solution of Partial Di erential Equations, Ph.D. Thesis, The George Washington University, 1997. 24] J. Zhang, On convergence and performance of iterative methods with fourth-order compact schemes, Numer. Methods Partial Di erential Equations, 14, 263{280 (1998). 25] J. Zhang, Residual smoothing techniques in multigrid, I: equivalence proof, Appl. Math. Comput. 86 (2-3), 283{303 (1997). 26] J. Zhang, Residual scaling techniques in multigrid, II: practical applications, Appl. Math. Comput. 90 (2-3), 229{292 (1998).

13

27] J. Zhang, A note on an accelerated high accuracy multigrid solution of the convectiondi usion equation with high Reynolds number, Numer. Methods Partial Di erential Equations, 16 (1), 1{10 (2000).

14

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

Figure 2: Streamline and equivorticity curves for Re = 5; 000 with h = 1=128.

15

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

Figure 3: Streamline and equivorticity curves for Re = 7; 500 with h = 1=128.

16