for the vertex-centered finite volume discretization, the relevant
position for calculating the gradient is *not* the position of the
center of the *sub*-control volume, but the center of the whole
control volume.
Since this patch changes the gradients of all VCFV simulations that
use twopoint gradients, quite a few new reference solutions are
required by this PR. I've looked at all of them manually and made sure
that they look reasonable. (The PR also includes a new reference
solution for the test for the adaptive finger problem -- which uses
ECFV -- but that's because the grid refinement seems to be quite
sensitive to the solution proceedure for that problem, not because
there's any difference in the solutions that can be noticed visually.)
note that all simulations that use element centered finite volume
discretization because there sub-control volumes are identical to full
control volumes. In particular, ebos -- and by extension `flow_ebos`
-- is unaffected because it does not even use that code.
for some reason, there are differences between my laptop with
optimizations and my workstation in debug mode. since the results look
_very_ similar in paraview, let's treat them both as equally valid.
as a side effect, two-point gradients are now used by default for the
vertex-centered finite volume discretization. (because P1-FE gradients
require the FE shape functions and those are provided by
dune-localfunctions.)
as a consequence, this triggered an update of quite a few reference
solutions: the differences are measurable, but as far as I can see,
the results are compareable. also, this commit regresses the
performance of the test for the reservoir problem with the vertex
centered finite volume scheme a bit. While I would not bet a house on
the reason, I'm pretty sure that this is caused by the switch from P1
FE gradients to two-point ones.
Note that even though I'm the author of this patch, it shamelessly
rips off substantial parts of the @dr-robertk's patch:
https://github.com/OPM/ewoms/pull/69
this problem did not work properly anyway: it oscillated like hell
(very likely to the spatial discretization used being inappropriate)
and it did not even converge if more than a single iteration was
required.
dune-alugrid >= 2.4 changed the element ordering changed from
lexicographical ordering to one defined by a space filling curve. the
old reference solutions are still valid (and obtained if older
versions of dune-alugrid are used) and are thus retained.
- start with an initial "do nothing" episode of 100 days to get
hydrostatic conditions.
- after that, produce oil and inject water for 900 days. (thereafter
the reservoir will be empty.)
- make the problem work with element centered FV discretizations. this
requires to make the width of the injection/production areas at
least one cell wide. This is achieved by using the new "WellWidth"
property which specifies the with of wells as a factor of the total
domain width.
- make the problem work with fully compositional models. This implied
to calculate the full composition for the fluid states which specify
the initial condition and the thermodynamic state at the wells.
- add tests and reference solutions for any combination of the {ECFV,
VCFV} discretizations and the {black-oil, NCP} models.
- the residual now does not consider constraints anymore
- instead, the central place for constraints is the linearizer:
- it gets a constraintsMap() method which is analogous to residual()
but it stores (DOF index, constraints vector) pairs because
typically only very few DOFs need to be constraint.
- the newton method consults the linearizer's constraint map to update
the error and the current iterative solution. the primary variables
for constraint degrees of freedom are now directly copied from the
'Constraints' object to correctly handle pseudo primary variables.
- the abilility to specify partial constraints is removed, i.e., it is
no longer possible to constrain some equations/primary variables of
a degree of freedom without having to specify all of them. The
reason is that is AFAICS with partial constraint DOFs it is
impossible to specify the pseudo primary variables for models which
require them (PVS, black-oil).
because of this, the reference solution for the Navier-Stokes test
is updated. the test still oscillates like hell, but fixing this
would require to implement spatial discretizations that are either
better in general (e.g., DG methods) or adapted to Navier-Stokes
problems (e.g., staggered grid FV methods). since both of these are
currently quite low on my list of priorities, let's just accept the
osscillations for now.
they are required because the element ordering of the latest
dune-alugrid has changed. (it now uses a space-filling curve instead
of lexicgraphic ordering.)
I cannot really say which one is better, but at least the new one
looks more reasonable: gas appears at the top of the reservoir where
the pressure is lower instead of close to the bottom.
the solution itself did not change, but yesterday's change of the
phase indices of the black-oil fluid system caused the fields to be
outputted in different order.
seems like the changes for vaporized oil caused it to be
different. Since SPE1 and SPE9 still deliver the same result and the
new reference solution is closer to what one would expect, the new
result is probably the correct one...
before this, gradients had the direction of the face between two
finite volumes, now they exhibit the direction of the two
FV-centers.
For axis-aligned grids the result is identical for interior faces, but
it is different for boundary faces or if faces are not
axis-aligned. This patch fixes the SPE9 troubles with anisotropic
permeabilities on tilted grids...
the goal is to make it faster on computers with many cores: The
easiest way to do this is to ensure that the longest running tests are
not taking too much time and that they need about the same time. Thus
this patch contains the following changes which limits the CPU time
taken by each test to about two minutes in debug mode on my machine:
- the water-air problem using the non-isothermal primary variable
switching model now uses an 16x16 instead of a 32x32 grid. as a
compensation it now runs for a year instead of 5000 seconds and the
global grid refinement is now tested.
- the end time of the lens problem ctests is now 3000 instead of
30000 seconds. The binary itself does not change at all.
- sort the tests in the CMakeLists.txt roughly in the order of their
required time. (this will cause ctest not having to wait for long
running test which were started late for too long.)
this is required so that the element-centered finite volume method
does not handle each partition of the domain separately. (i.e. so that
fluxes accross faces on the process boundaries are considered) mea culpa!
The fix for this is to also include these entries in the matrix which
uses domestic indices. This required some rather extensive changes to
the blacklisting mechanism as for this it must be possible to
translate the index of a blacklisted entity (i.e., an entity in a
ghost or an overlap cell) to a domestic index (i.e., the corresponding
index in the algebraic overlap).
Note that the code for algebraic overlaps is *fun* and the person who
wrote it should be tarred and feathered. (*ouch* ;)) Seriously: Better
approaches than "lets-throw-this-away-and-use-grid-overlaps" are
deeply appreciated. (The grid overlap is not really useful in Dune
because only "Mickey Mouse grids" like Dune::YaspGrid support it.)
For some reason, it changed because of the transition to the primary
variable switching approach. Because I trust the new result more than
the old, let's make this the new reference solution.
this allows to retrieve the name of the problem before it is
instantiated. this is required to be able to print the "Initializing
problem" message at the correct point (i.e., before instantiating the
problem).
This should make things a much more robust, partially because now the
linear and the non-linear solvers use the same convergence criterion.
Also, this patch includes some collateral indentation improvements.
instead of directly from the discretization. this should make the
generic quantities re-appear in the output for visualization.
this implies that the reference solutions for the models which
directly derived from the discretization also had to be updated.
This makes eWoms multi-discretization capable. Along the way, this
fixes some bugs and does a medium sized reorganization of the source tree.
This is a squashed patch of the following commits:
--------
1st commit message:
add initial version of the element centered finite volume discretization
currently, it is a misnomer as it is just a copy of the vertex
centered discretization plus some renames...
--------
2nd commit message:
rename [VE]cfvModel -> [VE]cfvDiscretization
--------
3rd commit message:
ecfv: prelimary changes required to make it compile
but not work yet...
--------
4th commit message:
Rename *FvElementGeometry to *Stencil
"Stencil" seems to be the standard expression for this concept...
(also, it is not specific to finite volume methods and is shorter.)
--------
5th commit message:
refactor the stencil class for the element centered finite volume discretization
--------
6th commit message:
ECFV: some work on the stencil class
--------
7th commit message:
ECFV: make the boundary handling code compile
--------
8th commit message:
rename elemContext() to elementContext()
--------
9th commit message:
ECFV: make the VTK output modules compile
--------
10th commit message:
stencil: introduce the concept of primary DOFs
also save an vector of all element pointers in the stencil.
--------
11th commit message:
ECFV: try to fix assembly; add missing timeIdx arguments to the num*() methods
--------
12th commit message:
ECFV: fix stupid mistake in the assembler
--------
13th commit message:
ECFV: remove a few implicit DOF == vertex assumptions
the black-oil example now runs without valgrind complaints until it encounters
a negative oil mole fraction.
--------
14th commit message:
VCFV: make everything compile again
all vertex centered FV examples should now work again...
--------
15th commit message:
rename [ev]cfvmodel.hh to [ev]cfvdiscretization.hh
the classes have already been renamed.
--------
16th commit message:
ECFV: make it work to the point where it can write out the initial solution.
--------
17th commit message:
ECFV: make it work
the local residual/jacobian needed some work in distinguishing primary
and secondary DOFs and there was an minor issue with the serialization
code.
for some reason, it seems still not correct. (-> convergence is too slow.)
--------
18th commit message:
VCFV: make it compile for the black oil model again
--------
19th commit message:
VCFV: make it compile with the remaining models again
--------
20th commit message:
flash model: make it work with ECFV
although this breaks its compatibility with VCFV. (-> next commit)
--------
21st commit message:
adapt the VCFV to make it compatible with the flash model again
--------
22nd commit message:
make all models compile with VCFV again
--------
23rd commit message:
VCFV: more cleanups of the stencil
VcfvStencil now does not have any public attributes anymore. TODO: do
not export attributes in the SubControlVolume and SubControlVolumeFace
classes.
--------
24th commit message:
VCFV: actually update the element pointer
--------
25th commit message:
change the blackoil model back to ECFV
--------
26th commit message:
immiscible model: make it compatible with the ECFV discretization
--------
27th commit message:
PVS model: make it work with ECFV
--------
28th commit message:
NCP model: make it work with ECFV
--------
29th commit message:
rename Vcfv*VelocityModule to *VelocityModule
--------
30th commit message:
richards model: make it work with ECFV
--------
31st commit message:
unify the ECFV and the VCFV VTK output modules
and other cleanups
--------
32nd commit message:
unify the common code of the VCFV and the ECFV disctretizations
--------
33rd commit message:
unify the element contexts between element and vertex centered finite volumes
--------
34th commit message:
unify the local jacobian class of the finite volume discretizations
--------
35th commit message:
replace [VE]vcf(LocalResidual|ElementContext|BoundaryContext|ConstraintsContext) by generic code
--------
36th commit message:
replace the [EV]cfvLocalResidual by generic code
--------
37th commit message:
unify the MultiPhaseProblem and Problem classes, introduce NullBorderListCreator
--------
38th commit message:
remove the discretization specific boundary context
--------
39th commit message:
unify the [EV]cfvDiscretization classes
--------
40th commit message:
Unify [EV]cfvMultiPhaseFluxVariables
--------
41st commit message:
Unify the [EC]cfvNewton* classes
--------
42nd commit message:
Unify [EV]cfvVolumeVariables
--------
43rd commit message:
unify [EV]cfvAssembler
--------
44th commit message:
unified flux variables: fix stupid mistake when calculating pressure gradients
--------
45th commit message:
unify what's to unify for the [EV]CFV properties
--------
46th commit message:
make the method to calculate gradients and values at flux approximation points changeable
Currently, this is used by the vertex centered finite volume method to
be able to use P1-finite element gradients instead of two-point
ones...
--------
47th commit message:
make the restart code work correctly, use the correct DofMapper for VCFV
--------
48th commit message:
actually use the gradient calculator in a model
the immiscible model in this case
--------
49th commit message:
move some files around to where they belong, use the new gradient calculation code in all models
TODO: proper handling of boundary gradients
--------
50th commit message:
fix the stokes model
currently it only works with the vertex centered finite volume
discretization, but the plan is to soon move it to a staggered grid
scheme anyway...
--------
51st commit message:
move all models back to using the vertex centered finite volume discretization by default
--------
52nd commit message:
models: some variable renames and documentation fixes
- scv -> dof
- vert -> dof
- vertex -> dof
- replace 'VCFV'
- fix some typos
--------
53rd commit message:
don't expect UG anymore
since it is quite non-free and hard to get. we now use ALUGrid instead!
--------
54th commit message:
temporarily disable jacobian recycling
--------
55th commit message:
fix writing/reading restart files using the generic code
--------
56th commit message:
fix bug where fluxes were only counted once in the stencil
this only affected the vertex centered finite volumes discretization...
--------
57th commit message:
boundary gradients: use the center of the sub-control volume adjacent to a boundary segment
--------
58th commit message:
make it compile on GCC
--------
59th commit message:
get rid of most hacks
for this, partial reassemble and jacobian recycling was brought
back. For the this and the remaining stuff the main trick is the
introduction of the GridCommHandleFactory concept which constructs
communication handles suited for the respective spatial
discretization...
--------
60th commit message:
fix a few annoying bugs
first, default the convergence criterion for the linear solver did not
honor the initial residual which lead to linear solver breakdowns,
then some debugging code was left in the discrete fracture model and
then there was a bug in the TP gradient approximation class...
this has the consequence that we need a new reference solution for the
discrete fracture problem...
--------
61st commit message:
iterative linear solver: remove the code for the non-default convergence criteria
--------
62nd commit message:
provide the FE cache instead of the local FE
this fixes a segfault in the stokes model caused by the fact that the
local FE was not initialized at this point.
--------
63rd commit message:
(Navier-)Stokes: fix bug due to the transition to unit normals
now, all tests pass for this branch. The only things which need to be
fixed are some annoying performance regressions compared to master and
some bug in the splices feature of the property system...
--------
64th commit message:
some fix for the local residual of the immiscible model
--------
65th commit message:
Navier-Stokes: implement SCV center gradients
There seems to be a bug in the previous implementation (the jacobian
inverse transposed is evaluated using the local, not the global
geometry), so the reference solution for the stokes2c test problem has
also been updated...
--------
66th commit message:
remove the ALUGrid specialization of the LensGridCreator and the YaspGrid one for the fingerproblem
using different grid seems to sometimes cause a different vertex
order, which in turn causes the respective test to fail if the
reference solution was computed using the other grid...
--------
67th commit message:
VCFV: use the correct BorderListCreator
this makes MPI parallel computations work again. apart from
performance regressions, this branch does not exhibit any known
regressions compared to master anymore...
--------
68th commit message:
make verything compile with the element centered finite volume discretization
except the Navier-Stokes and the two-phase DFM models, of course...
--------
69th commit message:
minor fixes
- make the navier-stokes model slighly more generic by using the
proper (in,ex)teriorIndex() methods on sub-control volumes
- make the signature of the calculateValue() template method of the
common two-point gradient approximator match the one of the vertex
centered finite volume one
--------
70th commit message:
fix fallout from the Big Rebase
--------
71st commit message:
ECFV: some bugs in the boundary
--------
72nd commit message:
make computeFlux() compute area-specific quantities
--------
73rd commit message:
fix more bugs in the element centered FV discretization
now eWoms should match Dumux pretty closely...
--------
74th commit message:
coalesce the common code of the multi phase porous medium models into "MultiPhaseBaseModel"
--------
75th commit message:
update reference solutions
these were changed because of the screw-up with the area of boundary
segments...
--------
76th commit message:
rename "ImplicitBase" to "FvBase"
because in eWoms, everything is implicit and these are currently the
base classes for all finite volume discretizations.
--------
77th commit message:
make the spatial discretization selectable using a splice
This requires an opm-core with a the patches from
https://github.com/OPM/opm-core/pull/446 merged...
--------
78th commit message:
rename the properties used for splices to *Splice
--------
79th commit message:
move the files in 'tests/models' to 'tests'
since 'tests' was empty except for the 'models' subdirectory...
--------
80th commit message:
improve and fix the tutorial
--------
81st commit message:
remove the -fno-strict-aliasing flag from the provided option files
seems like recent versions of Dune have been adapted...
--------
82nd commit message:
also compile all CO2 injection simulations using the element centered finite volume discretization
--------
83rd commit message:
PVS model: make it work properly with the element-centered finite volume discretiation
because DOF != number of vertices
this renames the 'test' directory to 'tests' and 'test/implicit' to
'tests/models'. the latter change reflects the fact that in eWoms all
models are implicit since the IMPET models have been removed.
Andreas Lauser
Robert Kloefkorn