Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Solution of transient transport equation using an upstream finite element scheme

P. S. Huyakorn
Department of Geoscience and Research and Development Institute of Mining and Technology, Socorro, NM, 87801, Division, USA New Mexico

K. Nilkuha
Department of Mining and Petroleum Engineering, and Technology, Socorro, NM, 87801, USA (Received December 1976; revised October 19 78) New Mexico Institute of Mining

An upstream finite element scheme is presented for overcoming the problem of numerical oscillations associated with the transient transport equation for convective dominated or purely convective flow. This scheme differs from the standard Galerkin scheme in that spatial discretization is performed via a general weighted residual technique which employs asymmetric weighting functions and yields upstream weighting of the convective term in the transport equation. The time derivative term of this equation is weighted using the standard basis functions which lead to a consistent mass matrix. The proposed finite element scheme is compared with the conventional upstream finite difference scheme both in one and two dimensions. Numerical results indicate that the new scheme is superior to its finite difference counterpart and is capable of producing reliable solutions for cases which cannot be handled satisfactorily by the standard Galerkin technique.

Introduction
In recent years, there have been an increasing number of attempts to develop suitable numerical methods for approximating the transient convective-dispersive (transport) equation. The increasing interest in this equation stems from the fact that its solution by standard finite difference or Galerkin finite element approximation can exhibit pronounced oscillatory behaviour and/or excessive numerical dispersion when the convective term is dominant. Various finite difference schemes for overcoming the numerical difficulties have been proposed by several workers.1-9 The most recent examples of such schemes include the improved upstream finite differencing by Lax and Wendroff, Chaudhari* and Todd,3 the box finite difference scheme by Keller,4 the flux corrected transport by Book et aZ.,5 and the third-order accurate time stepping, based on cancellation of truncation errors, by Laumbach.6 While several alternative schemes now exist for finite differences, an equivalent of these procedures still remains to be established for the finite element method. Recent
0307-904x/79/010007-11$02.00 0 1979 IPC Business Press

studieslo-* have shown that for certain convective dominated flow situations, standard Galerkin-based finite element techniques can produce oscillatory solutions whether linear, quadratic or cubic isoparametric elements or Hermitian-type elements are used. Although a limited degree of oscillation can be tolerated when one is dealing with a linear transport-flow problem, experience13-r4 has shown that in a nonlinear situation, such as that of multi-phase flow in geothermal systems or miscible displacement in petroleum reservoirs, slight oscillations near sharp fronts can be greatly amplified and eventually lead to convergence difficulties or instability in the iteration scheme unless the time step size is kept sufficiently small. In many instances, the numerical oscillations can be reduced or eliminated by the use of excessively refined grids. Such an approach is often impractical because of prohibitively expensive computation. An effective numerical procedure has to be established to make the finite element method more competitive with its finite difference counterpart in solving problems involving the transient transport equation. Such a procedure is developed in this paper.

Appl.

Math.

Modelling,

1979,

Vol 3, February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha

The present paper is a sequel of the earlier paper by Huyakorn,15 where a so-called upwind (upstream) finite element scheme was presented for handling the steady form of the transport equation. The scheme was shown to be capable of eliminating numerical oscillations while maintaining good accuracy and preserving all advantages of the finite element approach. This paper extends the original version of the upstream finite element method to deal with the transient transport equation. It will be demonstrated by numerical examples that the new scheme is superior in accuracy to the conventional finite difference schemes and is capable of handling cases for which the standard Galerkin technique produces solutions which are totally unacceptable.

drop the boundary integral terms of equation write it in the form:


A~JCJ +M~JdcJ dt =O

(5) and

(6)

where A~J and M~J are given by:


AZJ=

(7)

MZJ = s R

@NzNJ do

(8)

General formulation
In the proposed upstream weighted residual technique, the spatial (convective and dispersive) terms of the transient transport equation are weighted using asymmetric weighting functions and the time derivative term is weighted using standard basis functions. To illustrate the application of this technique, we consider the following form of the transport equation for an isotropic dispersion tensor:

Performing time integration of equation difference, we finally obtain:

(6) using finite

where 8 is a time weighting factor. To obtain numerical results which are accurate to the second-ordel in time, we use the value 0 = l/2 which corresponds to central time stepping or Crank-Nicolson scheme.

Upstream weighting functions


where c is concentration at time t, D is dispersion coefficient , Vi is Darcy velocity vector, and @is porosity of the porous medium. Let W, be a set of weighting functions and let a trial solution be written in the form:
2(xi,

and derivatives

t) = NJ(Xi)CJ(t>,

J= 1,2, . . ..y1

(2)

where NJ(xi) denotes the standard basis functions, cJ(t) denotes nodal values of concentration at time t, and repeated indices imply summation over the total number of nodes. We now discretize equation (1) in the following manner:

For linear one-dimensional and two-dimensional isoparametric finite elements, the expressions of asymmetric weighting functions and the general procedure for obtaining the sign of upstream parameters have been presented elsewhere.13y14,15 For completeness, they are also summarized in Appendix 1 of the present paper. To form the matrix elements AIJ in equation (9), we have to evaluate the derivatives of the weighting functions. In a onedimensional situation, this evaluation is straightforward. The derivatives dW,/ti and dW,/dx for the typical element in Figure 1 are simply obtained as:

(3) where R is the domain over which the integration is performed. Application of Greens theorem to the first integral of equation (3) leads to:

dW, -=-dx

1 h

3a

5-1 ( h*

h)

(11)

W~D~,lidB+ s I

WIi%inidB=O s
B

(4)

where (Yis the upstream weighting factor, and h is the element length. In a two-dimensional case, our experience indicates that to obtain a satisfactory solution, the derivatives aWz/a.$ and a W,lav of the quadrilateral element must be evaluated in such a manner that when differentiation is taken with respect to one particular coordinate, the value of the upstream parameters along the remaining coordinate must be set to zero, i.e.:

where B is the boundary of the solution domain, ni is the outward unit normal vector on the boundary. Substitution of equation (2) into (4) yields:

(12)
(13)

a2 - Hi CD t
axi For convenience in presenting
s B

Wlc^Viili dB

0
I

(5)
scheme, we

the numerical

a_ Figure 7

2
Onedimensional element

i
a

Appl.

Math.

Modelling,

1979,

Vol 3, February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha where (Yiand & are weighting factors associated with various sides of the quadrilateral element. For the typical element shown in Figure 2, the derivatives of the weighting functions may be written in terms of the local iso-parametric coordinates as follows: awr aq. (%O) =;(1-77)(3oiE-l) (14)
Figure .p. .3 I

I, Jtl
I i

- a2 il

P2

.*

11

:;+

i;l,l

~
1. al . i,;-I .

Quadrilateral

element and rectangular

grid

?(oi,o)
aw3 F

={(l-17)(3ar,k-1)

(15)
Vq = (q -

cI_l)/h @At

Cai, 0)

=-b<l

t77)(3~2t-1)

(16)

dc-1 __= (c; Ar dt

aw4 jj-

Ccyi, 0)

=+(l

t17)(3a*t-1)

(17)

awl -$-$OJj) awz +OJZ)

=:(l-v(3PaVl)

(18)

From equation (23), it is apparent that the spatial approximation produced by the upstream finite element method is identical to that obtained from the conventional upstream finite difference technique in which a*c,Qx* and &/ax are approximated by:

=;(1+8(3P,rl-l)

(19)

a*c - = e(,*CJ+A~ t (1-e)(62CZ)r ax* ~=s,~(v~,)"A't(lo)(scr)'+"'l t (l-e)


[cy(vcZ) + (1 -(u)(@)~]

(27)

~(OJj) aw4 +OJj)

=-$0

+5)(3P,n-1)

(20)

(28)

=-$(1-8(3&7)-l)

(21)

The above procedure for evaluating derivatives of the weighting functions has been successfully applied to isothermal and nonisothermal two-phase flow problems involving sharp saturation fronts.is3r4 In the present paper, we will demonstrate with numerical examples that such a procedure is also applicable to the transient transport equation. Furthermore, it will be shown that our upstream finite element scheme produces numerical results which are far more accurate than results produced by the widely used upstream finite difference technique.

Comparison of proposed finite element and upstream finite difference schemes


One-dimensional case To gain further insight into our upstream finite element method, we evaluate the matrix elements in equation (9) for the simplest case of onedimensional transport in a uniform velocity field. For 0 > i, the equation obtained for node Z may be written in the form:

- (1

-a>; (6c&r+Ar]

As can be expected, the upstream parameter Q!plays a significant role in damping out numerical oscillations. Again, we would like to point out that Q!= 1 corresponds to full upstream difference approximation of the convective term, whereas CY = 0 corresponds to central difference (or Galerkin finite element) approximation. We next proceed to examine the approximation of the time derivative term produced by the proposed finite element scheme. It is evident from equation (22) that this approximation indeed corresponds to a Simpson rule averaging of the time derivatives at nodal points Z - 1, Z and Z + 1. Clearly, such an approximation differs from the usual finite difference approximation unless the mass matrix [M] of the finite element equation is lumped by adding the coefficients along row Z and placing the sum on the diagonal. In applying the finite element method to the transient convective transport equation, the mass matrix should not be lumped. With the standard Galerkin finite element scheme, several workers (Gresho et al. l6 and Pinder and Gray) have shown that mass lumping degrades the accuracy of numerical results. Our experience with the upstream finite element scheme also indicates that lumping of its mass matrix also produces the same undesirable effect. We will demonstrate this in a subsequent section by comparing the performance of the recommended consistent-mass upstream finite element scheme (equation (22)) with the conventional upstream finite difference scheme which is represented by:

=(e-l)[h(6%+2(vcr)r -(l-+se,)j

1 (22)
where 62, 8, V and dldt denote difference operators, defined as follows: 6*~~=(~~_~-2c~+c~_~)/h~ 6c, = (crci - cr-i)/2h (23) (24)

e h2(62cr)r+Ar - (Yv; (vc$+At


vh - (1 -a) - (&I) r+Ar D I

- (1 - cr) 5

(&,)j

+2

(29)

Appl.

Math.

Modelling,

1979,

Vol 3, February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha Two-dimensional case Before proceeding to present numerical examples we would like to mention another distinction between our upstream finite element and the conventional upstream finite difference scheme. In multi-dimensional situations, the two numerical schemes will produce different types of spatial as well as temporal approximations. For the typical 2D rectangular grid shown in Figure 2, the finite element equation for node (I, J) is a nine-point equation whereas the finite difference equation for that node is a five-point difference equation. Because of this distinction, one can expect the upstream finite element method to give a higher-order spatial approximation than its counterpart. We will demonstrate this with a two-dimensional numerical example. In this example, we compare the performance of our upstream F.E. scheme with the upstream F.D. scheme presented in Appendix 2. fairly accurate solution and the use of upstream weighting (nonzero o) is unnecessary. For Pe = 100 and 00, the Galerkin scheme produces oscillatory results. To reduce the numerical oscillation to 1 or 2%, we need to use (Y= 0.13 in the upstream finite element scheme. From Figure 4/a-c) it is also evident that the upstream finite element scheme exhibits superior accuracy as compared to the upstream finite difference scheme. The latter requires a high value of (Y(CY > 0.2) to eliminate oscillations. In addition, its concentration fronts are considerably more dispersed than those obtained from the upstream finite element scheme. Next, we examine the effect of the upstream parameter (Yon the accuracy of the numerical solutions for the pure convection case, Pe = m. Figure 5 (a and b) shows concentration profiles obtained using different values of (Y.It can be seen that the numerical oscillations diminish with increasing value of cxand that a high value of cyleads to excessive smearing of the concentration front. To eliminate numerical oscillations but minimize false dispersion or smearing of the concentration front, we need to determine an optimal value of o. to be used in both upstream finite element and finite difference schemes. It is evident from Figure 5 (a and b) that oopt for the F.E. scheme lies
I.2

Numerical results
To test our numerical scheme and compare its performance with the conventional finite difference scheme, we considered three problems for which analytical solutions were available. These problems included (a), ID transport under constant concentration boundary condition at the upstream end; (b), ID pure convective transport of a rectangular waveform;(c), 2D transport of pollutant from an injection well. For each problem, we analysed various cases of convective-dominant transport. Several numerical experiments were performed on the first two 1D examples to gain insight into the behaviour of the numerical solution. We found that the main conclusions drawn from these experiments carried over to two dimensions. Example I The initial and boundary conditions for this 1D example are depicted in Figure 3. As can be seen, the initial condition employed is a linear concentration distribution over the first two grid points. The flow velocity is constant and equal to 0.5. We ran the problem using three meshes whose grid points were equally spaced. To identify different cases solved, we defined the mesh Peclet number and Courant number as Pe = VAX/D and Cr = VAtlAx, respectively. In presenting the results, we begin by comparing the performance of our upstream finite element scheme with the conventional upstream finite difference and Galerkin finite element schemes. Three cases are given here. The first case is a moderately convective-dominated transport with Pe = 10, the second is a highly convectivedominated transport with Pe = 100 and the third is a purely convective transport with Pe = m. Numerical results were obtained using Ax = 0.5 and At = 0.4. Figure 4(a-c) shows concentration profiles at t = 6.4. For Pe = 10 the Galerkin finite element scheme, which corresponds to CY = 0, produces a

OE
u

0.L

1,;

0.f
u

OL

C I ;
Analytic Upstream F.E, a =0,13 F.D, n -020 FE Upstream Galerkln

O-E
c,

o-4

0
Figure 3 Initial and boundary conditions for example 1

\
4

,\On

.2,

6
x

IO

Figure 4 Concentration profiles at time t = 6.4; mesh parameters: Cr = 0.4, AX = 0.5; (a), Pe = 10; (b), Pe = 100; (cl, Pe = -

10

Appl.

Math.

Modelling,

1979,

Vol 3, February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha

a
-Upstream ---x--Analytic FE,a=O,I Upstream Upstream F.E,a=O FE., a-0 2 5 08

08

Anaiytlc Upstream Upstream Upstream FD,a=O FD, a=0.3 FD, a=05 0.8

I.2

---+--

Analytic Upstream Upstream FE, Ax= FE.,Ax=O

---x--m
--

I
25

-x-Upstream

FE,Ax-O-5

0.8

04

04

4
x

1__~I 8

IO

1 b?<\ x
0

IO

Figure 5 Effect of parameter on 01on numerical results, mesh parameters: Pe = m, Cr = 0.4, Ax = 0.5, t = 6.4. (a), upstream F.E. solution; (b), upstream F.D. solution

x Figure 6 Convergence of upstream F.E. solutions, mesh parameters: Cr = 0.4, r = 6.4, (Y = 0.13. (a), Pe = 100; (b), Pe = -

between 0.1 and 0.2 while aopt for the F.D. scheme lies between 0.3 and 0.5. Thus we conclude that the F.D. scheme require a higher value of cythan the F.E. scheme in order to eliminate numerical oscillations. Furthermore, for the same value of QI,the concentration front produced by the F.D. scheme is considerably more dispersed than that produced by the F.E. scheme. We then proceed to study convergence of the upstream F.E. solution in the situations of convectivedominant transport (high Peclet number) and pure-convective transport (infinite Peclet number). In the latter situation, it is well known that the standard Galerkin F .E. and central F.D. schemes fail to give satisfactory convergence. To test our upstream F.E. scheme, we solved two cases corresponding to Pe = 100 and Pe = 00 using three meshes. The value of At was selected in such a way that the Courant number, Cr = 0.4 for all meshes. Figure 6 (a and b) show a comparison of numerical and analytical solutions. The numerical solutions were obtained using cr = 0.13 which was just sufficient to eliminate oscillations. It can be seen that as Ax decreases the numerical solution converges to the analytical solution. Next, we study the behaviour of the numerical solution produced by a particular mesh when At and hence Courant number (Cr = vat/Ax) increase. Figure 7/a-e) shows the results obtained for the pure convection case when Cr varies from 0.8 to 6.4. The key observation is that with an increasing value of Cr we need to use a higher value of cyto damp out oscillations. The computed concentration fronts become more dispersed at higher value of Cr. At values of Cr greater than 2, unacceptable oscillations still occur even though the value of 01is set equal to the full upstream value, (Y= 1. This general observation leads us to conclude that in solving the convectivedominant transport equation using upstream numerical techniques we should select Ax and At such that the resulting value of Cr is less than 2. By performing several

numerical experiments, we found that the optimal value of upstream parameter, Q,nt was related to the Courant number. This relationship is illustrated in Figure 8. It is an approximate relationship determined from the numerical experiments. When Cr lies in the range 1< Cr Q 2, the numerical solution obtained using aopt given by the curve in Figure 8 exhibits a mild degree of oscillation during the first few time steps. The oscillation diminishes quite rapidly at subsequent time levels. This behaviour is illustrated in Figure 9 for the purely convective case with Cr = 1.8. In the next two examples, it is shown that the value of oopt determined from the empirical curve in Figure 8 is also applicable to other 1D and 2D convective-dominant transport problems.
Example 2

In this example, we deal with pure convective transport of a 2-unit width rectangular slug, the initial position of which is shown in Figure 10. The exact solution for this particular case is the same rectangular slug but translated by a distance x = vt. We ran the problem using a mesh consisting of 41 nodes equally spaced at Ax = 0.5, We selected a constant value of At = 0.4, thus giving Cr = 0.4. The computed results at t = 2 and t = 10 are shown in Figure 11 (a-d). By comparing Figure I la with 11 b, it can be seen once again that the upstream F .E. scheme requires a smaller value of (IIthan the F.D. scheme to reduce the numerical oscillation to an acceptable level. Several key observations may be derived from Figure II (c and d). First, we note that the standard Galerkin F.E. and central F.D. schemes cannot produce satisfactory solutions for this particular case of pure convective transport. On the other hand, the use of full upstream weighting OL = 1 results in excessive numerical dispersion in both the upstream F.D. and F.E. schemes. For the latter scheme, the value of oopt is approximately equal to 0.13.

Appl.

Math.

Modelling,

1979, Vol 3, February

11

Solution

of transient

transport

equation:

P. S. Huyakorn
1

and K. Nilkuha

Analytic Upstream FE, a -0.1 FE, a- 02

08 c,

---X-- Upstream

04

0.8 Cr

12

I6

20 01and

Figure 8 Relationship between optimal value of parameter Courant number for upstream F.E. scheme, Pe = 0 Analytic Upstream Upstream Upstream FE,a=OI FE,a=0,2 F.E.,a-05

4t

IC

0
3(,

Anal@ Upstream
X

4 Y

0.t

x\ i\

--

FE,a=l.O FD,a

Upstream

-1.0

Figure 9 Behaviour of upstream F.E. solution for Cr = 1.8 when aopt = 0.76 is used, mesh parameters: Pe = _, Ax = 0.25, At = 0.9. Time steps are indicated by ringed numbers

0.L

12 -x Analytrc Upstream Upstream FE.,a = I.0 FD,aIO Figure 10 Initial and boundary conditions for example 2

ix , kx_
-x Analytic Upstream Upstream

E
FE, a = I.0 FE,a = I.0

We now make a direct comparison of the upstream FE. and F.D. schemes by plotting their results together as shown in Figure 12 (a and b). Once again, it is evident that the upstream F.E. scheme, with a consistent mass matrix, is superior to the upstream F.D. scheme. For the same value of CY = 0.25 used in both numerical schemes, the F.D. result not only displays larger oscillations but also significant phase lag. Overall, the F.E. result provides a significantly better approximation to the exact analytical solution, particularly when the optimal value of (Ywas used. It will be demonstrated in the next example that this improvement in accuracy is even more drastic in the two-dimensional flow situation.
Example 3

0.8 u

04

In this example, we consider a common problem of two-dimensional, contaminant transport in groundwater reservoirs. The contaminant source is assumed to be a finite diameter injection well located at the centre of a square-shaped confined aquifer. Figure I3 shows a plan

x\X\
4

x-yx_
6 Y

Figure 7 Effect of Courant number Cr on upstream F.E. solution, mesh parameters: Pe = m, Ax = 0.25, t = 6.4. (a), Cr = 0.8; (b), Cr = 1.6; (c), Cr = 2; (d), Cr = 3.2; (e), Cr = 6.4

12

Appl.

Math.

Modelling,

1979,

Vol 3, February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha


c F.E FE., F.E,

Analytic

--x--*-

Analytrc Golerkrn Upstream a=013 Upstream a=025

0.8
-04

0 CI --x-I.2 -Upstream FD, -Analytic Upstream a=0 Upstream a=025 a= I.0 FD, FD,

08-

Figure 17 Effect of parameter 01 on upstream F.E. and F.D. solutions, mesh parameters: (b), F.D. solution at t = 2; (c), F.E. solution at r = 10; (d), F.D. solution at t = 10

Pe = _, Cr = 0.4, Ax = 0.5. (a), F.E. solution at t = 2;

-o0.8 -----

Upstream Upstream Upstream

FE,a=O

13

FE,a=O25 FD.,a=025

04

L
\

h=c=o

F: ?

K, =Ky = I

ah ax

ax

=o

Dx=Dy = D Pe =0/O 0 =I

h-0

es=,
b
13 25 Y

c=o

-Analytic

-Upstream

FE ,a=0 FE, a-0

, ----Upstream --Upstream O-8 -

FD, a=0 25

C-l
Figure example 18 x Figure 72 Comparison of upstream F.E. and upstream F.D. solutions, mesh parameters: Pe = m, Cr = 0.4, Ax = 0.5. (a), concentration profiles at r = 2; (b), concentration profiles at r = 10 20

,=I6
73 3

j,

ah_ac_ ay ay and physical parameters for conditions

Boundary

written in the form: e--UZDt


0

view of a quadrant

of the aquifer with a quadrant of the injection well located at the origin of the coordinate axes. It is further assumed that the aquifer is homogeneous and isotropic and that the injection well geometry may be represented by a 2 ft x 2 ft square. In the analysis of this problem, we solved the groundwater flow and transport equations separately. The steady flow equation was solved first using the standard Galerkin scheme. Darcy velocities were calculated and fed into the transient transport equation which was then solved using the upstream F.E. or upstream F.D. scheme. The material parameters and boundary conditions employed in the calculation are depicted in Figure 13. To assess the accuracy of the numerical results, we employed the analytical solution given by Carslaw and Jaeger.18 This solution may be

24L&G

+ Y,2(4V)J
(30)

where J, and Y, are Bessel functions of the first and second kinds respectively and v is defined as:
v=--

Q
hD

(31)

Equation (29) was derived for a finite circular well in a semi-infinite confined aquifer. When the time value is such that the effect of the external aquifer boundary is negligible, this equation provides a very close approximation to our problem. We computed numerical and analytical solutions for two cases. To identify these cases, the Peclet number is defined as Pe = Q/D. In the first case, we used D = 0.152

Appl.

Math.

Modelling,

1979,

Vol 3, February

13

Solution of transient transport equation:

P. S. Huyakorn and K. Nilkuha

thus giving Pe = 65.8. In the second case, we assumed that


D = 0, thus giving Pe = m. Analytical result for the first case was computed from equations (30 and 3 1). The computed et aZ.,19 who also employed

result coincides with the result presented by Van Genuchten equation (30) to test their
7

numerical solution. For the second case involving only convective transport, the analytical solution takes a cylindrical shape. The concentration front is vertical and corresponds to the boundary surface of the cylinder. At a time t, the position of the front may be easily calculated from :

E -

Analytic
FE,a=P=O,42

--Upstream

i-t-= I 47l

Qt

(32) \
I

Analytic FD,a=P=O FD 42

--Upstream ----Central

I
Analytic F.D, a-P-042 FD

--Upstream ----Central

I Figure

7 x

II

13

I5

74 Comparison of numerical and analytical solutions for Pe = 65.8, mesh parameters: At = 1, Ax, = Ayl, Axi = Ayi = 2, i = 2-10. (a), analytical and F.E. solutions at f = 10 and 25; (b), analytical and F.D. solutions at t = 10; (cl, analytic and F.D. solutions at t = 25

where rf is the radial distance from the centre of the well. We iolved the two cases using two rectangular meshes consisting of 144 and 400 nodes, respectively. For mesh 1 the grid spacing of the first row and column are Ax, = Ayr = 1 and the grid spacings of the remaining rows and columns are AXi = Ayi = 2. For mesh 2, all grid spacings are set equal to 1. Computed solutions for the case corre sponding to Pe = 65.8 are presented in Figure 14(a-c). The depicted values of upstream parameters, (Y= fi = 0.42, are optimal values determined from Figure 8. By comparing Figure 14a with 14b and 14c it is evident that the Galerkin and upstream F.E. schemes produce results which are far more accurate than the central and upstream F.D. schemes. For Pe = 65.8, the Galerkin F.E. result is almost oscillation free while the central F.D. result exhibits considerable oscillation and excessive numerical dispersion at t = 25. Next, we present the solutions obtained for the pure convective case using mesh 1. Figure 15(a-d) shows concentration profiles at t = 10 and t = 25. A number of key observations may be derived from these figures. First, it may be noted that the Galerkin F.E. scheme produces solutions which are totally unacceptable because of large oscillations both in space and time. We can expect the central F.D. scheme to yield the same consequences if not worse. Also noteworthy is the performance of the upstream F.E. and F.D. schemes. The upstream FE. scheme yields solutions which are much more accurate than solutions produced by its counterpart, particularly when the optimal values of upstream parameters, (Y= B = 0.42, are used. At time t = 25, the upstream F.D. result

l2/

,\

\ ------

Analytic Galerkln Upstream Upstream FE FE,a=P=O FE,a=pI.0 42

a.

-.--2 \ N\\ \ .1 \ \\z__ ------

Analytic Upstream Upstream Upstream FE, a+=0,42 F.D., a=P=0,42 ED, a=P=l,O

c,

I----,

Analytic Galerkln Upstream Upstream FE FE) a=P=O-42 FE, a-P -1.0

b
_

-.-----

Analytic Upstream Upstream Upstream FE., a-P-0 FD, a=P-0 FD, a-P=l.O 42 42

----,

I2, , , 0 8 - \ I \ \ \ ,,I 04I 1 ,

I \

--

c,

O,

I 3

7 x

_--

13

15

7 Y

II

13

15

14. (a), analytical and F.E. solutions Figure 15 Comparison of numerical and analytical solutions for Pe = m, mesh parameters are as in Figure at t = 10; (b), analytical and F.E. solutions at r = 25; (c), analytical, F.D. and F.E. solutions at t = 10; (d),analytical, F.D. and F.E. solutions at t=25

14

Appl.

Math.

Modelting,

1979,

Vol 3, February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha


I 2 -

--Upstream
At=l, ----Upstream At=05,

Analyttc

a
08 FE, mesh 2, a=P=0,42 04

FE, mesh I, a-P-0 42

08 0 04

,
0 I.2 -\ u O8 -7 0.4 -.\ \ . \ \ \.
I 1 I

\\.\,_,
Analytic Upstream FE, plot y-or y-0x1s along

Analytic
Upstream FE., mesh I, At=lO,a=P=O42 Upstream FE, mesh 2, Ar-0.5, a-P=042

b
7 08

---

-.-

C Analytic FE ,mesh I, At=l FE, mesh2,At=05 c ---Galerkln --Galerkln 1.;

Figure 17 Comparison of concentration distributions along various directions; (a), upstream F.E. solution at t = 25 for Pe = = obtained using mesh 1, At = 1 and a! = p = 0.42; (b), upstream F.E. solution at r = 10 obtained using mesh 2, At = 0.5 and 01 = p = 0.42

c,

0.E

mesh F.E. solution is quite comparable with the finer mesh F.D. solution.

Conclusions
0.L

Figure 76 Convergence behaviour of upstream and Galerkin F.E. solutions; (a), upstream F.E. solution for Pe = 65.8, t = 10; (b), upstream F.E. solution for Pe = m, f = 10; (cl, Galerkin solution for Pe=m,r=lO

It has been demonstrated that the upstream finite element scheme, developed previously for the steady state transport equation can be extended to handle the transient situation. The extended numerical scheme has been compared with the conventional upstream finite difference and Galerkin finite element schemes. The following conclusions may be drawn from this comparative study: (1) The proposed upstream finite element scheme is capable of overcoming numerical oscillations and producing acceptable solutions for certain cases of convective dominated and purely convective flows which cannot be satisfactorily handled by the standard Galerkin technique. (2) Results from 1D and 2D test examples indicate that the upstream finite element scheme produced solutions which were significantly more accurate than solutions produced by the conventional upstream finite difference scheme. The new scheme not only exhibited superior phase properties and steeper concentration fronts but also required lower values of upstream parameters in order to control oscillations. (3) For the 2D example involving pollutant transport from an injection well, the upstream finite difference solution obtained using a 12 x 12 grid not only displayed excessive numerical dispersion, for the pure convective case, but also exhibited a large discrepancy between concentration profiles along the x- or y-direction and the main diagonal. On the other hand, the proposed upstream finite element solution obtained using the same mesh did not exhibit such undesirable properties. (4) Numerical experiments revealed that a relationship existed between the optimal value of the upstream parameter and the mesh Courant number. This relationship has been obtained for 1D and shown to be applicable to the 2D case as well. We wish to emphasize that the relationship between CY,,,~ and Cr, presented in this paper, is only approximate.

displays excessive numerical dispersion, the concentration front being much more smeared than that at the earlier time t = 10. On the other hand, the concentration front given by the upstream F.E. more or less maintains its original shape. We next consider the convergence of the F .E. solution. Figure 16(a-c) show results obtained using meshes 1 and 2. For mesh 2, which has the nodal spacing Ax = Ay = 1, we select At = 0.5 thus giving the same Courant number as for mesh 1. In Figure I6 (a and b) it can be seen that the upstream F.E. solution approaches the analytical solution as the mesh becomes more refined. On the other hand, the Galerkin F.E. solution for Pe = 00 fails to converge as can be seen from Figure 16~. Finally, we check the symmetry of the numerical solutions. Naturally, one expects the concentration distribution in different radial directions to be the same because of the symmetric feature of the flow problem. The finite element solution satisfies this requirement quite well as can be seen from Figure I7 (a and b), where the plots along the x-axis and the main diagonal are compared. On the other hand, the upstream F.D. solution obtained using mesh 1 displays large discrepancy between concentration profiles along the two directions as can be observed from Figure 18/a-b). To reduce this discrepancy to an acceptable level, we had to use mesh 2 which contains almost three times as many nodes as mesh 1. It is also evident from Figure 19 that the accuracy of the coarse

Appl.

Math.

Modelling,

1979,

Vol 3, February

15

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha

*e-m -_.

. .SI_ 0.8 --%> . \ \

----\ \ \

Analytic Upstream FD,plotalong x-or y-ox/s Upstream diagonal FD,plotalong -\ \

---\ \

Analytic Upstream x-ory-axls FD, plot along

Upstream FD, plot along diagonal

04-

\ -\

b .

Figure 78 Comparison of concentration distributions along various directions; (a), upstream F.D. solution at t = 10 for Pe = 65.8 obtained using mesh 1, Ar = 1 and cz = p = 0.42; (b), upstream F.D. solution at t = 25 for Pe = - obtained using mesh 1, At = 1 and (Y = p = 0.42; (c), upstream F.D. solution at t = 10 for Pe = - obtained using mesh 2, At = 0.5 and LY= p = 0.42; (d), upstream F.D. solution at t = 10 for Pe = = obtained using mesh 2, At = 0.5 and oi = p = 1

I2

-.-

Analytic Upstream FE,mesh At=l, a-6-042 I,

References
Lax, P. D. and Wendroff, B. Comm. Pure Appl. Math. 1964, 17,381 Chaudhari, N. M. Sot. Pet. Eng. J. 1971, 11,277 Todd, M. R. etal. Sot. Pet. Eng. J. 1972, 12,515 Keller, H. B. A new finite difference scheme for parabolic problems (B. Hubbard, Ed.), Academic Press, New York, 1971 Book,D. L.etal. J. Comput. Phys. 1975, 18,248 Laumbach,D.D.Soc.Pet.Eng.J.1975,15,517 Lantz, R. B. Sot. Pet. Eng., AIMEJ. 1971,315 Price, H. S. et al. J. Math. Phys. 1966,45, 301 Price, H. S. et al. Sot. Pet. Eng., AZME J. 1968,8 (3), 293 Lam, D. C. L. Inlroc. 1st Znt. Conf Finite Elements Wat. Res., Princeton University, July 1976 Van Genuchten, M. T. Proc. 1st Znt. Conf. Finite Elements Water Res., Princeton University, July 1976 Huyakorn, P. S. and Taylor, C. Proc. 1st Znt. Conf. Finite Elements Water Res., Princeton University, July 1976 Huyakorn, P. S. and Pinder, G. F. Proc. Znt. Conf. Appl. Numer. Modelling, University of Southampton, 11-15 July 1977 Huyakorn, P. S. and Pinder, G. F. Proc. 2nd ZMACS {AZCA)

08---t, 04.. \

Upstream FD, mesh 2, At-0 5, a-P-1

Ol

3 Comparison

7 of upstream

II I

13 6

15

Figure 79
Pe=-,t=lO

F.E. and F.D. solutions for

5 6 I 8 9 10

Finally, we would like to point out that the proposed upstream finite element scheme differs from the upstream finite difference scheme both in one and two dimensions. In one dimension, the temporal approximation in the finite element scheme leads to a consistent mass matrix while the temporal approximation in the finite difference scheme yields a diagonal or lumped mass matrix. In two dimensions, both the temporal and spatial approximations in the two numerical schemes are different. For a rectangular grid, the upstream finite element method yields a nine-point nodal equation while the upstream finite difference technique yields a five-point nodal equation. Experience indicates that the latter type of spatial approximation can lead to undesirable asymmetric properties such as large discrepancy between concentration distributions in different directions and sensitivity to grid orientation.*O

11 12 13

14

Int. Symp. Comput. Meth. Partial Differential Equations,


15 16 Lehigh University, 22-24 June 1977 Huyakorn, P. S. Appl. Math. Modelling, 1977, 1, 187 Gresho, P. M. et al. Proc. 2nd Znt. Symp. Finite Elemeni Meth. Flow Problems, S. Margherita Ligure, Italy, 14-18 June 1976 Pinder, G. F. and Gray, W. G. Finite Element Simulation in Surface and Subsurface Hydrology, Academic Press, New York, 1977 Carslaw, H. S. and Jaeger, J. C. Conduction of Heat in Solids (2nd edn), Oxford University Press, New York, 1959 Van Genuchten, M. T. et al. Water Resour. Res. 1977,13,451 Settari, A. and Price, H. S. Proc. 4th SPE Symp. Numer. Simulation Reservoir Performance, Los Angeles, California, 19-20 February 1976, No. SPE 5721

17

18 19 20

Acknowledgements
The work reported in this paper was supported in part by funding from the Geophysical Research Center of the Research and Development Division of New Mexico Institute of Mining and Technology, and from the New Mexico Energy Research Board, Project No. EJ-177-084. The authors would like to thank Professor Lynn W. Gelhar of the Geoscience Department, New Mexico Tech, for reviewing the manuscript.

Appendix

Asymmetric weighting functions and procedure for determining signs of upstream parameters
Using the notation illustrated in Figures 1 and 2, the expressions of the weighting functions for 1D and quadri-

16

Appl.

Math.

Modelling,

1979,

Vol

3,

February

Solution of transient transport equation: P. S. Huyakorn and K. Nilkuha

lateral elements are given as follows: One-dimensional elements

The sign of (Yis determined a>0 01<0 if v>O if v<O

in accordance with:

W, =$[(lt[)(3a([-3a-2)t4] W, =$[(l t[)(-3a~+3ff+2)]

Appendix 2
where ,$ is a local iso-parametric co-ordinate and 0~is upstream parameter associated with the element. Quadrilateral elements

2D finite difference scheme used in numerical comparison


Referring to the grid point (I,J) in Figure 2, the upstream finite difference equation used in our numerical comparison may be written in the form: e [D(6;c,+At t @?+At) - vU,(1-_(y)6,C~+A~ - ?+(l -P)6yCr+At t 6$cf)

-vxCxvxc~+A~ - z$ljV$t+At
@ __Ct+At=_ ti

At z*J

At

Ci,J+ (e - 1) [I@$

-vx(l-cu)6,cf-v.&Yv.Jf - v,pvycwhere (g, 77)is a local iso-parametric coordinate system and (or, /3r, (Yeand pa are upstream parameters associated with sides l-2,2-3,4-3 and l-4 of an element, respectively. To determine the sign of (Yor fi let lIJ denote the direction vector of a particular side IJ and let vz and vJ denote velocity vectors at node I and J respectively. (For each type of element, the positive sense of direction of lz~ is depicted in Figure 2.) The component of an average velocity along side IJ is obtained by taking the scalar product: v=;(q+vJ).I~J VJl -@s,cf] (33)

where 0 is the time weighting factor, (Yand fl are upstream parameters in the x andy directions, respectively, Sz,6, and 0 are difference operators defined as follows:
~~C=(CZ-~,J-~~Z,J--Z+,,J>/(A~)~

(34) (35) (36)

6,~ = (cz+ I, J - cz- I, ~)/(2Ax) ?xC = (Cz, J


Cz1, JYAx

The remaining difference operators may be derived in a similar manner. Equation (33) represents a penta-diagonal matrix equation which may be efficiently solved by using the well-known strongly implicit procedure.

Appl.

Math.

Modelling,

1979,

Vol 3, February

17

You might also like