HDG Navier Stokes #3732
HDG Navier Stokes #3732
Conversation
lindsayad
force-pushed
the
hdg-ns
branch
2 times, most recently
from
4522b49 to
acd9220
Compare
This allows verification of the new APIs. Running with this version of the example we obtain the same results as the previous (git) version
And add some boundary ID info
Set tau = 1 for all faces
Still don't work. There's still an artifact at the flow outlet
This reverts commit 39a2438.
And some clang formatting
This reverts commit 5f843be.
I was missing `JxW_face` for the pressure face kernels
I'm through the `scalar_face_` routines
I think the Jacobian might be wrong though
I'm just not sure how to do this with the "two-step" vector increment
This reverts commit 8a255a7.
This reverts commit 483d9aa.
This ensures that the aux system variables are up-to-date when we perform the residual evaluation which considers both the "aux" and nonlinear solution values. This also ensures that the Jacobian matches the residual for all nonlinear iterations >= 1
There is a fundamental problem with this. The problem is that the computed increment for the mixed system is nonzero even if the increment is zero for the LM system. So basically I think trying to increment the mixed solution outside of a global incrementing is just not gonna work
This reverts commit 19579d0.
Precheck only implemented for PETSc
|
Marking this as ready for review. These current failures seem unrelated? |
|
Job Coverage on da16a3e wanted to post the following: Coverage
Warnings
This comment will be updated on new commits. |
||||||||||||||||||||||||||
|
Definitely unrelated. Logan's looking atone; I have a hypothesis about and an attempted fix for the other. |
|
I've opened #3742 as an acknowledgment that I should try using |
|
I was curious what a profile for this would look like so I made one for the 64x64 lid driven cavity case. I'm pleased that the dominant cost is the global solve as opposed to local inverses during assembly. This means there will be a lot to gain from using field split preconditioning
|
|
What was your ksp/pc type for that profile? But regardless, 64x64 is pretty small; if the assembly was already negligible there then I wouldn't worry about it in general. The only exception I can think of is really high polynomial degree. We already do an O(Np^2d) assembly (p^d bases, p^d quadrature point values for each on each of N elements), even on tensor product elements where there are asymptotically cheaper tricks ... but these local inverses look like O(Np^3d), don't they? You've got m=p^d entries in those matrices and then you're doing O(m^3) inverses on each element? |
Based on work by Nguyen and Cockburn. I've implemented MMS tests for both lid and channel flow type boundary value problems.
Cavity MMS results
grid_size=2
L2 error for u is: 1.543638e-01
L2 error for v is: 1.543638e-01
L2 error for pressure is: 2.879229e-01
grid_size=4
L2 error for u is: 4.007067e-02
L2 error for v is: 4.007067e-02
L2 error for pressure is: 5.268862e-02
grid_size=8
L2 error for u is: 1.012986e-02
L2 error for v is: 1.012986e-02
L2 error for pressure is: 1.111670e-02
grid_size=16
L2 error for u is: 2.543457e-03
L2 error for v is: 2.543457e-03
L2 error for pressure is: 2.581651e-03
Channel MMS results
grid_size=2
L2 error for u is: 2.318716e-01
L2 error for v is: 3.545995e-01
L2 error for pressure is: 6.271178e-01
grid_size=4
L2 error for u is: 6.192154e-02
L2 error for v is: 9.922736e-02
L2 error for pressure is: 1.775775e-01
grid_size=8
L2 error for u is: 1.577982e-02
L2 error for v is: 2.597410e-02
L2 error for pressure is: 4.410009e-02
grid_size=16
L2 error for u is: 3.977834e-03
L2 error for v is: 6.505690e-03
L2 error for pressure is: 1.091284e-02