OPM
Avoid deleting derivatives with zero pressure difference
- avoid getting zero in matrix in zero pressure/zero flux case
- may make it easier to have lower/upper limits on pressures
This should be the correct, but it will may change many tests so it should be investigated carefully.
Somehow, it affects a lot of cases, maybe most of the regression tests. The results does not look problematic, while some checking need to be done.
https://ci.opm-project.org/job/opm-models-PR-builder/482/
It was reported that this PR and https://github.com/OPM/opm-simulators/pull/5236 helps thermal cases with certain setup. Will trigger a jenkins test to get an overview about how to get the PR and https://github.com/OPM/opm-simulators/pull/5236 in.
The following tests still needs to be fixed manually. Previously, we use smaller time steps to stabilize the time step sizes. But it is still vulnerable to other small changes. I do not have good ideas how to improve it other than relaxing the tolerance. And it is tricky to keep the reference updated automatically.
mpi.python_fluidstate_variables mpi.python_primary_variables
I will go through the test failures and provide evaluation. It looks like all minor and it should not risk the regression failures for other PRs.
The changes here are non-controversial.
Just post the observation of the regression failures, for later information. Not sure whether be able to finish it before the summer.
The following tests require more looking into.
mpi.compareECLFiles_flow+SPE1CASE2_BRINE_GASWATER c (double check) mpi.compareECLFiles_flow+UDT-1D-03 c double check mpi.compareECLFiles_flow+CO2STORE_GASWAT c double check mpi.compareECLFiles_flow+H2STORE_GASWAT c double check mpi.compareECLFiles_flow+MOD4_UDQ_ACTIONX c double check mpi.compareECLFiles_flow+MOD4_GRP_GEFAC c double check mpi.compareParallelSim_flow+AQUFLUX-01 c double check mpi.compareECLFiles_flow+SPE1CASE2_ROCK2DTR ci double check mpi.compareECLFiles_flow+UDQ_M1 c double check mpi.compareECLFiles_flow+KRNUM-03Y c (double check) mpi.compareECLFiles_flow+MAX_WGR_1 c carefully mpi.compareECLFiles_flow+9_3D_GINJ_GAS_MAX_EXPORT_STW c (double check)
The following do not have a SUMMARY output, so did not check in details,
mpi.compareECLFiles_flow+CO2STORE_DRSDTCON c (no plotted output) mpi.compareECLFiles_flow+H2STORE c (no plotted output) mpi.compareECLFiles_flow+CO2STORE_ENERGY o (no plotted output) mpi.compareECLFiles_flow+H2STORE_DIFFUSIVE c (no plotted output) mpi.compareECLFiles_flow+H2STORE_ENERGY c (no plotted output) mpi.compareECLFiles_flow+BC_LAB c no plotting output mpi.compareECLFiles_flow+BO_DIFFUSE_CASE1 c no plotting output
Other regression failures looks fine.
UDT-1D-01B
MSW_MODEL_1 and BASE_MODEL_1 have the same problem,
UDQ_M1
WVFPEXP-02
0_BASE_MODEL6
9_3D_GINJ_GAS_MAX_EXPORT_MSW
H2STORE_GASWAT
CO2STORE_DRSDTCON
2D_OW_CTAQUIFER
The following case might not have summary output, has not checked. GASWATER_VAPWAT_PRECSALT CO2STORE CO2STORE_GW_DIRICHLET CO2STORE_DIFFUSIVE CO2STORE_ENERGY H2STORE H2STORE_DIFFUSIVE H2STORE_ENERGY BO_DIFFUSE_CASE1
For the case UDT-1D-01B
master branch has oscillation of the control mode for the production well and at the end, it chose ORAT control and had BHP lower than bhp limit (100bar, if I understand the UDQ code correct), which is not ideal.
With this PR, the solution result is much better, no control mode oscillation. It is under BHP control at the end. UDT-1D-01B.pdf
For UDQ_M1, using stricter tolerance for the wells either with '--tolerance-wells=1.e-8' or '--relaxed-well-flow-tol=1.e-7' will stablize the simulation some, while it is affected by the time stepping.
Editing the DATA file to use smaller time steps at the beginning, the results look better,
So I will edit the DATA file and add '--tolerance-wells=1.e-8' to the regression test for UDQ_M1.
adding --tolerance-wells=1.e-8 to UDT-1D-01B also help to stablize the simulation.
Will also add that to UDQ-1D-01B regression test.
adding
--tolerance-wells=1.e-8to UDT-1D-01B also help to stablize the simulation.
Will also add that to UDQ-1D-01B regression test.
But it also causes the control mode oscillating for the well PROD and it ends up gets the wrong results with the wrong control mode.
Turning of the local switch control with --local-well-solve-control-switching=false can avoid to get BHP lower than the bhp limit. Maybe it is something interesting to investigate. @steink
Typically we struggle between THP or group control or rate control, this one looks like struggle between BHP and ORAT, which is rather rare. But it can be because the BHP limit is a function of the ORAT, which introduces some cross dependence there.
It might be interesting in improving the local well solve control switching.
For NORNE_ATW2013_B1H_RE-PERF
adding '--tolerance-wells=1.e-8' improves the comparison. Will updating this regression test.
9_3D_GINJ_GAS_MAX_EXPORT_MSW and WVFPEXP-02
looks like due to my PC has different results from the jenkins reference. We can ignore it.
CO2STORE_DRSDTCON
with master branch there are multiple time step cutting, which did not occurs with the Pull request, which I see is a positive thing.
Turning of the local switch control with --local-well-solve-control-switching=false can avoid to get BHP lower than the bhp limit. Maybe it is something interesting to investigate. @steink
There are quite a few tests that violate limits. I've been doing some testing with increasing the default --maximum-number-of-well-switches from 3 to 10, and this seems to improve the overall quality of the solutions (also this test).
In some cases when switching between pressure and rate control, reservoir pressure updates overshoots, and the local well-solve will immediately switch back (causing oscillations). With the old approach (no local switching) one can be lucky and experience less oscillations. There is currently no good remedy for preventing these situations perhaps except trying harder (increasing --maximum-number-of-well-switches) .
I have finalized the investigation of the regression failures. Updated a few regression test. From my side, I am ready to update the reference and get the PR and the downstream PR merged.
Sorry, I forgot the following cases,
GASWATER_VAPWAT_PRECSALT CO2STORE CO2STORE_GW_DIRICHLET CO2STORE_DIFFUSIVE CO2STORE_ENERGY H2STORE H2STORE_DIFFUSIVE H2STORE_ENERGY BO_DIFFUSE_CASE1
I am investigating and will create PR to stablize these tests. For the H2STORE ones, we need to use --tolerance-cnv-relaxed=0.01 to make the results sensible.
Not sure how to check GASWATER_VAPWAT_PRECSALT, but from the jenkins report, it looks fine. Potentially, we can tighten the tolerance in some way to improve it.
BO_DIFFUSE_CASE1 looks fine.
CO2STORE ones look fine.
I will update the H2STORE related (smaller time steps and tightening CNV tolerance ) so that the regression tests are more stable.
There are still regression failures related to H2STORE, the curves match while there is still some different time stepping at the beginning.