Two-grid scheme for characteristics finite-element solution of nonlinear convection diffusion problems

Two-grid scheme for characteristics finite-element solution of nonlinear convection diffusion problems

Applied Mathematics and Computation 165 (2005) 419–431 www.elsevier.com/locate/amc Two-grid scheme for characteristics finite-element solution of nonl...

274KB Sizes 1 Downloads 35 Views

Applied Mathematics and Computation 165 (2005) 419–431 www.elsevier.com/locate/amc

Two-grid scheme for characteristics finite-element solution of nonlinear convection diffusion problems q Xin-qiang Qin

a,b,*

, Yi-chen Ma

a

a

b

School of Sciences, Xi’an Jiaotong University, Xi’an 710049, China School of Sciences, Xi’an University of Technology, No. 5 Jinhua Nanlu, Xi’an 710048, China

Abstract A two-grid scheme for characteristics finite-element solution of nonlinear convection-dominated diffusion equations was constructed. The L2 and H1 error estimates of the scheme were derived. A numerical example was also presented. The scheme involves solving one small, nonlinear problem on the coarse-grid system, one linear problem on the fine-grid system. The error estimates and the numerical example show that the new scheme is efficient to the nonlinear convection-dominated diffusion equations.  2004 Elsevier Inc. All rights reserved. Keywords: Convection–diffusion equations; Characteristics finite-element; Two-grid scheme

q

This work supported by National Science Foundation of China grants 10371096. Corresponding author. Address: School of Sciences, XiÕan University of Technology, No. 5 Jinhua Nanlu, XiÕan 710048, China. E-mail address: [email protected] (X.-q. Qin). *

0096-3003/$ - see front matter  2004 Elsevier Inc. All rights reserved. doi:10.1016/j.amc.2004.06.021

420

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

1. Introduction Convection–diffusion equations occur in numerous applications, noteworthy among them being mathematical models of environment pollution and neutron transportation. In this paper we discuss the nonlinear form 8   ou ou o ou > < cðxÞ ot þ bðxÞ ox  ox aðxÞ ox ¼ f ðu; x; tÞ; ðx; tÞ 2 I  ð0; T ; ð1:1Þ uðx; tÞ ¼ 0; ðx; tÞ 2 oI  ð0; T ; > : uðx; 0Þ ¼ u0 ðxÞ; x 2 I; we use the modified method of characteristics finite-element developed by Douglas–Russell [1], to discrete in space with equal-order accuracy in u. To linearize the resulting discrete equations, we use a two-grid scheme, which allows us to iterate on a grid much coarser than that used for the final solution. We owe the impetus for using a two-grid approach to Xu [2,3,8], who demonstrates the possibility of mapping a fine-grid Galerkin finite-element problem onto a coarse grid. In that context, one solves the nonlinear problem via Newton-like iterations. After convergence on the coarse grid, one then extrapolates back to the fine grid using a Taylor expansion. Dawson and Wheeler [4] extend the method to nonlinear reaction–diffusion equations. Li Wu and Myron [5] extend it to another nonlinear reaction–diffusion equations. The remainder of the paper is organized as follows: Section 2 describes the algorithm. Section 3 gives an error analysis establishing the methodÕs convergence. Section 4 presents some computational results confirming the methodÕs utility, and Section 5 draws conclusions. We assume that the coefficients a, b and c are bounded and that d bðxÞ (H1) 0 < a0 6 aðxÞ 6 1; 0 < c0 6 cðxÞ 6 1; j bðxÞ j þ j dx ðcðxÞÞj 6 C; x 2 I cðxÞ 2 o f (H2) j of j þ j j 6 C; x 2 I 2 ou ou

Throughout this paper, we shall use the letter C to denote a generic positive constant which may for different values at its different occurrences. Let Wm,p(X) denote the Sobolev space on I. And the norm is defined by 8 1=p > P > p < kDa vkLp ðXÞ ; 1 6 p < 1; a6m kvkW m;p ¼ > > : max esssupjDa vj; p ¼ 1: a6m

m,2

We write W (X) = Hm(X), kvkHm(X) = kvkm, and kvkL2(X) = kvk0 = kvk and also assume that the solution u of (1.1) satisfies: (H3) u 2 L1(0,T;Hr + 1(I)), 2 L2 ð0; T ; H rþs Þ; r ¼ 1 if s = 1 and r P 2 if s = 0, (H4) ou ot 2 (H5) oot2u 2 L2 ð0; T ; L2 ðIÞÞ.

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

421

2. Description of the method 2.1. Characteristics finite-element discretization We begin by briefly reviewing the characteristics finite-element discretization of the problem (1.1). Let qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wðxÞ ¼ c2 ðxÞ þ b2 ðxÞ ð2:1Þ and let the characteristic direction associated with the operator cut + bux be denoted by s = s(x,t), where o cðxÞ o bðxÞ o ¼ þ : ð2:2Þ os wðxÞ ot wðxÞ ox Then, Eqs. (1.1) can be put in form 8   ou o ou > > aðxÞ ¼ f ðu; x; tÞ; x 2 I; t 2 ð0; T ; > wðxÞ  < os ox ox ð2:3Þ uðx; tÞ ¼ 0; x 2 oI t 2 ð0; T ; > > > : uðx; 0Þ ¼ u0 ðxÞ; x 2 I: We introduce the Sobolev space V ¼ H 10 ðIÞ, and the notation (u,v) = Iu(x)v(x)dx,"u,v2V. Let Aðu; vÞ ¼ ða ou ; ovÞ. Then, multiplying the first ox ox equation of (2.1) by any v 2 V and applying Green formula, (2.3) can be written in the equivalent variational form  8 < w ou ; v þ Aðu; vÞ ¼ ðf ðuÞ; vÞ; v 2 V ; os ð2:4Þ : Aðuð0Þ  u0 ; vÞ ¼ 0; v 2 V : We consider a time step Dt and approximate the solution at times tn = nDt, n = 1, 2, . . ., N, Dt ¼ NT . The characteristic derivative is approximated in the following way at t = tn  n ou uðx; tn Þ  uðx; tn1 Þ uðx; tn Þ  uðx; tn1 Þ : ð2:5Þ w  wðxÞ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ cðxÞ os Dt 2 ðx  xÞ þ Dt2 Namely, a backtracking algorithm is used to approximate the characteristic derivative. x ¼ x  bðxÞ Dt is the foot (at level tn1) of the characteristic correcðxÞ sponding to x at the head (at level tn) (see Fig. 1).

Fig. 1. The direction of characteristic.

422

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

Let the mesh size of the grid be h, we denote this grid by Dh and call it the fine grid. Vh be a finite-element subspace of V ˙ W1,1(I). The characteristics finite-element method for (2.4) is defined: For n = 1, 2, . . ., N, find unh 2 V h such that 8  n > un1 < c uh   h ; v þ Aðunh ; vÞ ¼ ðf ðunh Þ; vÞ; v 2 V h ; Dt ð2:6Þ > : 0 Aðuh  u0 ; vÞ ¼ 0; v 2 V h ; where

  bðxÞ Dt; tn1 : uhn1 ðxÞ ¼ uh ðx; tn1 Þ ¼ uh x  unh ¼ uh ðtn Þ;  cðxÞ

ð2:7Þ

The initial approximation u0h in Vh can be defined as any reasonable approximation of u0 such as the interpolation of u0 in Vh. The (2.6) determines funh g uniquely in terms of the data u0. 2.2. Linearization via the two-grid method To solve the system (2.6), we use the Newton-like iterations. The idea is to devote all of the effort of nonlinear iteration to coarse-grid problems. Thus the iterations yield approximate finite-element solutions unH on a coarse subgrid DH having mesh size DH  Dh,H > h and corresponding space VH. To approximate unh on Dh, we then execute iteration of the fine-grid system, using the converged coarse-grid solution unH together with the heuristic  n unh ¼ unH þ enh ; uh  ^ ð2:8Þ f ðunh Þ  f ðunH Þ þ f 0 ðunH Þenh : This tactic yields the following system of equations for the fine-grid solution: 8 n   n  eh uH  un1 > H n 0 n n n > ;v > c ; v þ Aðeh ; vÞ  ðf ðuH Þeh ; vÞ ¼ ðf ðuH Þ; vÞ  c < Dt Dt n > > AðuH ; vÞ; 8v 2 V h ; > : Aðu0h  u0 ; vÞ ¼ 0; v 2 V h : ð2:9Þ Namely, the two-grid method has two stages: Stage 1: On coarse-grid DH, solve the following nonlinear system for unH 2 V H :  8 n un1 < c uH   H ; v þ Aðun ; vÞ ¼ ðf ðun Þ; vÞ; v 2 V H; H H Dt ð2:10Þ : 0 AðuH  u0 ; vÞ ¼ 0; v 2 V H :

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

Stage 2: On the fine-grid Dh, solve the following linear system for ^unh 2 V h ! 8 n1 n  > ^ ^  u u < c h h ; v þ Aðun ; vÞ ¼ ðf ðun Þ þ f 0 ðun Þð^un  un Þ; vÞ; h

Dt

> :

Aðu0h  u0 ; vÞ ¼ 0;

H

H

h

H

423

v 2 V h;

v 2 V h: ð2:11Þ

This linearization is efficient if the coarse-grid DH be coarser than the finegrid Dh without sacrificing the accuracy associated with the fine-grid solution unh . The analysis in Section 3. demonstrate that, as with the Galerkin-based scheme proposed by Xu [2], one can choose H = O(h4/7) and still retain the error estimates associated with the characteristics finite-element scheme. Numerical examples in Section 4. confirm the efficiency to solve the nonlinear problems with two-grid method. 3. Convergence analysis The two-grid method affords a remarkably efficient linearization. It is possible to execute all of the Newton-like iterations on very coarse grids, then use the heuristic (2.9) to obtain an accurate fine-grid solution in one additional step. We devote the rest of this section to the analysis of the two-grid scheme. We begin by describing the discretization space and associated projection operator. We then give a convergence analysis for the characteristics finite-element method, using the key parts of this argument in the analysis of the two-grid scheme. In what follows, we consider (1.1) on I. For v 2 Hr + 1 the following approximation property holds: inf ðkv  vh k þ hkv  vh k1 Þ 6 Chrþ1 kvkrþ1 ;

vh 2V h

ð3:1Þ

where r P 1. Let wh: [0,T] ! Vh, satisfy Aðu  wh ; vÞ ¼ 0;

v 2 V h:

ð3:2Þ

Set g = u  wh, n = uh  wh, then u  uh = gn. It is well known that [1], for p = 2,1 and 1 6 s 6 r + 1, kgkLp ð0;T ;L2 ðIÞÞ þ hkgkLp ð0;T ;H 1 ðIÞÞ 6 Chs kukLp ð0;T ;H s ðIÞÞ for r P 2 and 1 6 s 6 r + 1 og og s ou þ h 6 Ch : ot 2 ot 2 ot L2 ð0;T ;L2 ðIÞÞ L ð0;T ;H 1 ðIÞÞ L ð0;T ;H s1 ðIÞÞ

ð3:3Þ

ð3:4Þ

424

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

The inverse property on Vh holds [6], namely, for vh 2 Vh, kvh k1 6 h1=2 kvh k:

ð3:5Þ

And we also have the approximation property [7]: kun  wnh k1 6 kun krþ1;1 hrþ1 :

ð3:6Þ

3.1. Error estimates for the characteristics finite-element We now estimate the errors. From (3.3) and (3.4), to obtain error bounds for uuh, it suffices to estimate n. Theorem 1. Let u and uh be the respective solutions of (2.4) and (2.6). Under assumptions (H1–H5). We have the error estimate " o2 u n n max ku  uh k 6 C Dt ot2 2 16n6N L ð0;T ;L2 Þ !# ð3:7Þ ou rþ1 þh kukL1 ð0;T ;H rþ1 ðIÞÞ þ os L2 ð0;T ;H rþs ðIÞÞ 6 CðDt þ hrþ1 Þ for sufficient small Dt, where s = 1 if r = 1 and s = 0 if r > 1. Proof. At t = tn, subtracting (2.4) from (2.6) and let v = nn to give  c

   nn  nn1 n oun un  un1 n ; n þ Aðnn ; nn Þ ¼ w c ;n Dt os Dt !  n   n1  n1 n1 n1 g g g  g nn1  n n n n þ c ;n þ c ;n  c ;n Dt Dt Dt þ ðf ðun Þ  f ðunh Þ; nn Þ:

ð3:8Þ

At any point x2I, by the Taylor theorem, un ðxÞÞðun ðxÞ  unh ðxÞÞ f ðun ðxÞÞ  f ðunh ðxÞÞ ¼ f 0 ð~

ð3:9Þ

~n ðxÞ. Therefore for some value u un ðxÞÞðun ðxÞ  unh ðxÞÞ; nn Þ ðf ðun Þ  f ðunh Þ; nn Þ ¼ ðf 0 ð~ un ðxÞÞgn ; nn Þ  ðf 0 ð~un ðxÞÞnn ; nn Þ: ¼ ðf 0 ð~

ð3:10Þ

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

425

Consequently, from Eq. (3.8),  n    n  nn1 n oun un  un1 n ; n þ Aðnn ; nn Þ ¼ w c ;n c Dt os Dt !  n   n1  n1 n1 n1 g g g  g nn1  n n n n þ c ;n þ c ;n  c ;n Dt Dt Dt þ ðf 0 ð~ un ðxÞÞgn ; nn Þ  ðf 0 ð~ un ðxÞÞnn ; nn Þ  T 1 þ T 2 þ T 3 þ T 4 þ T 5 þ T 6 : ð3:11Þ We can estimate the left terms of (3.11) by aða  bÞ P 12 ða2  b2 Þ and (H1)  n  n  nn1 n 1 a0 2 ½ðcnn ; nn Þ  ðcnn1 ; nn1 Þ þ knn k1 ; n þ Aðnn ; nn Þ P c 2Dt Dt 2 ð3:12Þ and use the results provided in [1] to the right terms T1  T4, we have 2 2 2 2 o u o u n 2 n 2 jT 1 j 6 C Dt þ ekn k 6 C os2 2 n1 n 2 os2 2 n1 n 2 Dt þ ekn k1 ; L ðt ;t ;L Þ L ðt ;t ;L Þ ð3:13Þ C og 2 jT 2 j 6 þ eknn k1 ; Dt ot L2 ðtn1 ;tn ;H 1 Þ

ð3:14Þ

2

2

ð3:15Þ

2

2

ð3:16Þ

jT 3 j 6 Ckgn1 k þ eknn k1 ; jT 4 j 6 Cknn1 k þ eknn k1 for any positive constant e. And from (H2), we have 2

2

jT 5 j 6 Cðkgn k þ knn k Þ; 2

jT 6 j 6 Cknn k :

ð3:17Þ ð3:18Þ

Choosing proper e, inequalities (3.12)–(3.18) can be combined with (3.11) to give the recursion relation 1 a0 2 ½ðcnn ; nn Þ  ðcnn1 ; nn1 Þ þ knn k1 2Dt " 2 2 2 o u 1 og n 2 n1 2 6 C kn k þ kn k þ Dt 2 þ os L2 ðtn1 ;tn ;L2 Þ Dt ot L2 ðtn1 ;tn ;H 1 Þ # 2

2

þkgn k þ kgn1 k :

ð3:19Þ

426

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

It follows from (2.6) and (3.2) that n0 = 0. If we multiply (3.19) by 2Dt, sum over n, and apply the discrete Gronwall lemma, it follows that " 2 # o2 u og max kn k 6 C Dt 2 þ þ kgkL1 ð0;T ;L2 Þ 16n6N os L2 ð0;T ;L2 Þ ot L2 ð0;T ;H 1 Þ " !# o2 u ou rþ1 ; þh kukL1 ð0;T ;H rþ1 ðIÞÞ þ 6 C Dt 2 os L2 ð0;T ;L2 Þ ot L2 ð0;T ;H rþs ðIÞÞ n

ð3:20Þ

which, together with un  unh ¼ gn  nn , (3.3) and (3.4) yields the desired result (3.7). h An optimal order estimate for u  uh in H1(I) can be derived in a similar fashion. We can give the theorem below: Theorem 2. Let u and uh be the respective solutions of (2.4) and (2.6). Under assumptions (H1–H5), We have the error estimate n

max ku 

16n6N

unh k1

" 2 o2 u 2 6 C Dt þ hr kukL1 ð0;T ;H rþ1 ðIÞÞ os2 2 2 L ð0;T ;L Þ !# ou þ 6 CðDt þ hr Þ ot L2 ð0;T ;H r ðIÞÞ

ð3:21Þ

for sufficient small Dt and r P 2. In the proof, the embedding theorem was used.

3.2. Error estimates for the two-grid method We now analyze the two-grid method for iteratively solving the nonlinear system (2.6). This scheme is attractive because it requires nonlinear iteration only on the coarse grid, give the fine-grid solution in a single, linear, Newton-like step. The following theorem gives the error: Theorem 3. Let u and ^ uh be the respective solutions of (2.4) and (2.11). Under assumptions (H1–H5), we have unh k 6 CðDt þ hrþ1 þ H 2rþ3=2 Þ max kun  ^

16n6N

for Dt sufficiently small, where r P 1.

ð3:22Þ

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

427

n Proof. To estimate the error, set gn ¼ un  wnh ,^ n ¼ ^unh  wnh . At t = tn. Subtractn ing Eqs. (2.11) from itÕs respective counterparts (2.4) and choosing v ¼ ^n , we have !   n n1 n n n1 ^ n n n ^ ^n ; n ^ þ Aðn ^n Þ ¼ w ou  c u  uh ; ^nn ;n c Dt os Dt 0 1 n1  n    ^nn1  ^n n g  gn1 ^n gn1   gn1 ^n ;n þ c ; n  @c ; ^n A þ c Dt Dt Dt n þ ðf ðun Þ  f ðunH Þ  f 0 ðunH Þð^ un  unH Þ; ^ n ÞÞ:

A Taylor expansion of f about

ð3:23Þ

unH

yields 1 2 uÞðun  unH Þ ð3:24Þ f ðun Þ ¼ f ðunH Þ þ f 0 ðunH Þðun  unH Þ þ f 00 ð~ 2 for some function ~ u. From (3.22) and (3.23), we get !   n n1 ^ ^ n n n oun un  uhn1 ^n n n ^ ^ ^ c ; n þ Aðn ; n Þ ¼ w c ;n Dt os Dt 0 1 n1  n   n1  ^nn1  ^n n g  gn1 ^n g  gn1 ^n ;n þ c ; n  @c ; ^n A þ c Dt Dt Dt   n n 1 00 2 ðf ð~uÞðun  unH Þ ; ^n þ ðf 0 ðunH Þgn ; ^ n Þ  ðf 0 ðunH Þnn ; nn Þ þ 2 7 X  T 0i : ð3:25Þ i¼1

Then arguments similar to those used for the characteristics finite-element error estimate (Theorem 1) yield  n n n1 n1 n 2 n1 2 1 a0 n 2 ½ðc^ n ;^ n k1 6 C k^n k þ k^n k n Þ  ðc^ n ;^ n Þ þ k^ 2Dt 2 2 2 o u 1 n 2 n1 2 og þ Dt os2 2 n1 n 2 þ Dt ot 2 n1 n 1 þ kg k þ kg k L ðt ;t ;H Þ L ðt ;t ;L Þ  þ kðun  unH Þ2 k2 : ð3:26Þ Now multiplying both sides of (3.26) 2Dt and summing over n and apply the discrete Gronwall lemma, it follows that: " 2 og o2 u n ^ þ max kn k 6 C Dt 2 16n6N os L2 ð0;T ;L2 Þ ot L2 ð0;T ;H 1 Þ # 2

þkgkL1 ð0;T ;L2 Þ þ max kðun  unH Þ k ; 16n6N

^0 ¼ 0 is used. where the n

ð3:27Þ

428

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

Let h = H in (3.7) and (3.20) and using (3.5) and (3.6), we have 2

kðun  unH Þ k 6 kun  unH k1 kun  unH k 6 ðkun  wnH k1 þ kwnH  unH k1 Þkun  unH k 6 Cðkun krþ1;1 H rþ1 þ H 1=2 kwnH  unH kÞkun  unH k

ð3:28Þ

6 CðH rþ1 þ H 1=2 ðH rþ1 þ DtÞÞðH rþ1 þ DtÞ 2

6 CH 1=2 ðH rþ1 þ DtÞ 6 CðH 2rþ3=2 þ DtÞ: n

Combining un  ^ unh ¼ gn  ^ n together with (3.3) and (3.4) and (3.27) and (3.28) yields the results (3.22). h An optimal order estimate for u  ^ uh in H1(I) can be derived in a similar fashion. We can give the theorem below: Theorem 4. Let u and uh be the respective solutions of (2.4) and (2.11).Under assumptions (H1H5). We have the error estimate unh k1 6 CðDt þ hr þ H 2r Þ max kun  ^

16n6N

ð3:29Þ

for sufficient small Dt and r P 2. Theorems 1 and 3 demonstrate a remarkable fact about two-grid scheme: when we iterate on a very coarse-grid DH and still get good approximations by taking one iteration on the fine-grid Dh. Specifically, for h = H7/4 as r = 1, the two-grid scheme solution ^ uh has the same accuracy compared with that of the nonlinear iteration solution uh. 4. Numerical example To illustrate the effectiveness of the two-grid linearization, we examine the following simple test problem: 8   ou ou o ou > 2 > > < ot þ ox  ox aðxÞ ox ¼ u þ Gðx; tÞ; x 2 ½0; 1; t 2 ð0; T ; ð4:1Þ uð0; tÞ ¼ 1; uð1; tÞ ¼ 0; t 2 ð0; T ; > > > : uðx; 0Þ ¼ 1  x; x 2 ½0; 1; where a(x) = 0.001. The G(x,t) is determined by the exact solution u(x,t) = (1  x)ext. We solve (4.1) by Dt = 0.125 · 105 from t = 0 to 0.20. For H = 24, h = 7/4 H = 27, we use the two-grid scheme presented in this paper. First, we give the coarse-grid partition DH with H = 24 (Fig. 2) and get the nonlinear itera-

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

429

Fig. 2. The coarse-grid DH with H = 24.

Fig. 3. The fine-grid Dh with h = H7/4 = 27.

tion solution uH on DH according to (2.10). Then we give the fine-grid partition Dh with h = H7/4 (Fig. 3) and obtain the fine-grid solution ^uh in one additional step on Dh to use the converged coarse-grid solution uH according to (2.11). To check up the numerical accuracy of the two-grid method, we obtain the nonlinear iteration solution uh on the fine-grid Dh according to (2.6). The numerical results of the two-grid method and the nonlinear iteration method and the exact solution are presented in Fig. 4. From Fig. 4 we can conclude that the two-grid method solution ^uh has the same accuracy as that of the nonlinear iteration solution uh. To confirm the efficiency of the two-grid method, we obtain the CPU time and errors of the numerical solution of the two-grid scheme (2.10) and (2.11) and the nonlinear iteration scheme (2.6). Table 1 shows the L2 errors of ^ uh and the CPU time for two-grid scheme of characteristics finite-element

1 exact solution twogrid solution nonlinear iteration solution

0.9 0.8 0.7

u(x,t)

0.6 0.5 0.4 0.3 0.2 0.1 0

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

x Fig. 4. Three kind of solutions at t = 0.20.

0.9

1

430

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

Table 1 L2 errors for ^ uh at t = 0.20 h 7

2

k^uh uk kuk

H 4

CPU time 4

1.212 · 10

2

00

93

Table 2 L2 errors for uh at t = 0.20 h 4

2

kuh uk kuk

CPU time 4

1.228 · 10

184

00

method when h = H7/4. Table 2 shows the L2 errors of uh and the CPU time for characteristics finite-element nonlinear iteration method over the range of fine grids Dh. Tables 1 and 2 also show that the two-grid solution ^uh and the nonlinear iteration solution uh have the same order of accuracy. And the computational time of two-grid scheme is less than nonlinear iteration method. The convergence speed increases by about two times. Although the CPU time reported here provides only a rough measure, the efficiency of the two-grid approach is obvious.

5. Conclusion Two-grid linearization offers an attractive way to solve the nonlinear problems involving convection–diffusion equations. The key feature of the two-grid method is that it allows one to execute all of the nonlinear iterations on a system associated with a coarse spatial grid, without sacrificing the order of accuracy of the fine-grid solution. The two-grid scheme combined with the characteristics finite-element method, cannot only decrease the numerical oscillation caused by dominated convections, but also save much more computational time for solving the nonlinear convection-dominated diffusion problems.

References [1] J. Douglas Jr., T.F. Russell, Numerical method for convection-dominated diffusion problem based on combining the method of characteristics with finite element or finite difference procedures[J], SIAM J. Numer. Anal. 19 (5) (1982) 871–885. [2] J. Xu, A novel two-grid method for semilinear elliptic equations, SIAM J. Sci. Comput. 15 (1) (1994) 231–237.

X.-q. Qin, Y.-c. Ma / Appl. Math. Comput. 165 (2005) 419–431

431

[3] J. Xu, Two grid finite element discretization techniques for linear and nonlinear PDEs[J], SIAM J. Numer. Anal. 33 (5) (1996) 1759–1777. [4] C.N. Dawson, M.F. Wheeler, Two-grid methods for mixed finite element approximations of nonlinear parabolic equations[J], Contemp. Math. 180 (1994) 191–203. [5] W. Li, M.B. Allen III, Two-grid methods for mixed finite-element solutions of reaction– diffusion equations, Numer. Meth. Partial Differential Eqs. [J] 15 (5) (1999) 589–604. [6] S.C. Brenner, L.R. Scott, The Mathematical Theory of Finite Element Methods, SpringerVerlag, Berlin, New York, 1996. [7] Ph.G. Ciarlet, The Finite Element Methods for Elliptic Problems, North-Holland, Amsterdam, 1978. [8] J. Xu, Local and parallel finite element algorithms based on two-grid discretizations[J], Math. Comput. 69 (231) (1999) 881–909.