air–water flow does not give rise to convergence pro- blems, since the number of Newton iterations is virtually
the same for both LIQUID_STATE switching and no LIQUID_STATE switching runs, before the NAPL is
injected. Convergence difficulties arise only when the NAPL is injected. Although the cure for convergence
problems is similar to that used in
26
, the cause of these problems appears to be different.
In the following examples, we will always use LIQUID_STATE variable substitution.
9.2 Jacobian selection for the flux limiter: experiments
In this Section, we will carry out some tests to determine the effect of using various approximations to the Jacobian when
using a flux limited discretization eqns 22–26. Tables 12–14 compare the performance of pure upstream
weighting and flux limiters based on van Leer and MUSCL approaches. For the flux limiters, the nonlinear iteration is
carried out using the full Jacobian, an approximate Jacobian derivatives respect to i
2up
in eqns 22–26 is ignored, and a first-order Jacobian Jacobian constructed using first-
order upstream weighting. These methods are described in more detail in Section 5.1. All the results in Tables 12–14
were obtained using the maximum potential method see Fig. 1 for location of the second upstream point.
Examination of these tables reveals that, at least for this interphase mass transfer data, which is similar to that in
13
, the nonequilibrium models are easier numerical problems
than the equilibrium models, for the same scenario. This will be discussed more detail in Section 9.3.
Another interesting result form Tables 12–14, is that there is very little difference between using the full Jaco-
bian, the approximate Jacobian, and the first-order Jacobian, in terms of total Newton iterations. This holds for MUSCL
and van Leer weighting, equilibrium and nonequilibrium models, and all three scenarios. For any given weighting
method, model formation, and scenario, the number of Newton iterations varied by at most . 15. Of course,
the CPU times for the full Jacobian methods were about double the CPU cost compared to the approximate and
first-order Jacobians, due to the increased size of the Jacobian.
Consequently, on the basis of these results, either the first- order or approximate Jacobian method would appear to be a
good choice for solution of the nonlinear equations. There does not seem to be any advantage to using the full
Jacobian.
Table 12. Comparison of full and approximate Jacobian, stra- tegies, heterogeneous DNAPL examples, 100 days: CPU times
are for a SUN Sparc-10
Method Total nonlinear
iterations Total linear
iterations CPU time s
Equilibrium model Upstream weighting
Full Jacobian 250
1540 241
MUSCL weighting Approximate
Jacobian 443
2413 431
First-order Jacobian
430 2271
413 Full Jacobian
435 2261
831 Van Leer weighting
Approximate Jacobian
440 2248
400 First-order
Jacobian 450
2392 424
Full Jacobian 427
21398 784
Nonequilibrium model Upstream weighting
Full Jacobian 288
1806 390
MUSCL weighting Approximate
Jacobian 280
1782 406
First-order Jacobian
272 1772
393 Full Jacobian
265 1603
727 Van Leer weighting
Approximate Jacobian
289 1835
403 First-order
Jacobian 275
1798 397
Full Jacobian 271
1686 765
Table 13. Comparison of full and approximate Jacobian, stra- tegies, heterogeneous DNAPL examples, 100 days: CPU times
are for a SUN Sparc-10
Method Total nonlinear
iterations Total linear
iterations CPU time s
Equilibrium model Upstream weighting
Full Jacobian 309
1647 282
MUSCL weighting Approximate
Jacobian 512
2094 460
First-order Jacobian
517 2094
460 Full Jacobian
509 2195
941 Van Leer weighting
Approximate Jacobian
622 2377
547 First-order
Jacobian 713
2540 615
Full Jacobian 729
2519 1319
Nonequilibrium model Upstream weighting
Full Jacobian 249
1383 324
MUSCL weighting Approximate
Jacobian 256
1390 356
First-order Jacobian
257 1391
351 Full Jacobian
252 1379
692 Van Leer weighting
Approximate Jacobian
252 1394
355 First-order
Jacobian 255
1401 348
Full Jacobian 250
1386 699
Nonlinear multiphase flow 445
Recall from Figs 7 and 8 that MUSCL and van Leer weighting give very similar computed solutions. Tables
12–14 also indicate that, in terms of numerical perfor- mance, there is little difference between MUSCL and van
Leer methods. This is somewhat surprising, since the MUSCL method in eqn 26 was specifically designed to
use a smooth limiter function. The van Leer limiter function eqn 25 has a discontinuity in slope at r ¼ 0.
In order to verify that the above trends are also observed in convection-dominated problems, the homogeneous
DNAPL nonequilibrium problem was run again, this time with all dispersion parameters in Table 3 reduced by a factor
of 10. This resulted in a grid Peclet number of . 20. Table 15 shows that even for this problem, there does not seem to
be any advantage of using the full Jacobian, compared to either the first-order Jacobian or the approximate Jacobian.
Note that, as shown in Fig. 15, there is very little differ- ence in the computed solution for the Heterogeneous
DNAPL problem, using either the maximum potential or geometric method for determining the second upstream
point see Fig. 11 for use in the flux limiter expressions. Again, this is somewhat surprising, since the maximum
potential method would appear to be more physically reasonable for heterogeneous problems. This was also
reported in
20
. In order to determine if there was a difference in comput-
ing cost when using maximum potential or geometric methods for determining the second upstream point, the
heterogeneous DNAPL scenario was run again, using the geometric method. The run statistics for the computations
Table 14. Comparison of full and approximate Jacobian, stra- tegies, heterogeneous LNAPL examples, 100 days: CPU times
are for a SUN Sparc-10
Method Total nonlinear
iterations Total linear
iterations CPU time s
Equilibrium model Upstream weighting
Full Jacobian 1023 4788
880 MUSCL weighting
Approximate Jacobian
1345 6606
1225 First-order
Jacobian 1334
6410 1217
Full Jacobian 1319 6365
2246 Van Leer weighting
Approximate Jacobian
1283 6333
1149 First-order
Jacobian 1436
7135 1325
Full Jacobian 1235 5877
2067 Nonequilibrium model
Upstream weighting Full Jacobian
624 2567
736 MUSCL weighting
Approximate Jacobian
612 2481
761 First-order
Jacobian 706
3046 886
Full Jacobian 648
2569 1544
Van Leer weighting Approximate
Jacobian 672
2907 817
First-order Jacobian
714 3078
904 Full Jacobian
599 2427
1439
Table 15. Comparison of full and approximate Jacobian, stra- tegies, homogeneous DNAPL examples, 100 days
Method Total nonlinear
iterations Total linear
iterations CPU time s
Nonequilibrium model Upstream weighting
Full Jacobian 268
1829 376
MUSCL weighting Approximate
Jacobian 290
2061 443
First-order Jacobian
286 1956
424 Full Jacobian
295 2056
857 Van Leer weighting
Approximate Jacobian
310 2265
460 First-order
Jacobian 289
1939 420
Full Jacobian 277
1877 778
This example used dispersion parameters 10 times smaller com- pared to the base case in Table 3, resulting in a grid Peclet number
of . 20. CPU times are for a Sun Sparc-10.
Table 16. Comparison of full and approximate Jacobian, stra- tegies, homogeneous DNAPL examples, 100 days
Method Total nonlinear
iterations Total linear
iterations CPU time s
Equilibrium model MUSCL weighting
Approximate Jacobian
932 3213
812 First-order
Jacobian 368
2036 365
Full Jacobian 935
2861 2461
Van Leer weighting Approximate
Jacobian 1107
3681 890
First-order Jacobian
1359 4272
1110 Full Jacobian 1020
2954 2676
Nonequilibrium model MUSCL weighting
Approximate Jacobian
315 2337
498 First-order
Jacobian 314
2278 477
Full Jacobian 307
1673 1287
Van Leer weighting Approximate
Jacobian 316
2372 499
First-order Jacobian
320 2342
490 Full Jacobian
345 179
1446 This example used the geometric method for determining the
second upstream point. Compare with Table 12. CPU times are for a SUN Sparc-10.
446 P. A. Forsyth et al.
using the geometric method are given in Table 16 and should be compared with Table 12. Note that for the equili-
brium models, there is a large increase in the number of Newton iterations required when using the geometric
method compared to the maximum potential method, with one exception. The one anomaly occurs when using
MUSCL weighting, with the first-order Jacobian equili- brium model. In this case, the geometric method run has
slightly fewer total Newton iterations than the maximum potential computation. However, for the rest of the equili-
brium simulations, the number of Newton iterations for the geometric method is about double that of the maximum
potential methods. For the nonequilibrium assumptions, the geometric-based runs have slightly more Newton itera-
tions compared with the maximum potential runs.
Note that the CPU time requirements for the geometric- based methods, when using the full Jacobian Newton itera-
tion, are much greater than the corresponding runs using the maximum potential method. This is because the number of
nonlinear iterations has increased for the geometric methods and, as well, the geometric Jacobians have more non-zeros
than the maximum potential Jacobians. This results in greater costs for construction and solution of the Jacobian
for the geometric Jacobians. The maximum potential Jacobian has fewer non-zeros than the geometric Jacobian
since the maximum potential method recognizes that many possible non-zeros in the geometric Jacobian are identically
zero in the maximum potential Jacobian. An example of this would be if fluid flows from node j to node i, then i2upi,j
does not exist see eqn 24.
Consequently, even though the solutions for both maxi- mum potential and geometric methods are very similar, the
geometric method is generally slower than the maximum potential method.
9.3 Effect of varying mass transfer parameters