3D Adaptivity Keeps Refining Boundary Elements

[cwangaw]

Hi all,

I am solving some system with a singularity at the middle of the cube, and I have to refine the mesh adaptively. If I use a square instead in 2D, everything works well. However, if I move to 3D, when I mark the elements and pass the DG0 marking function to msh = msh.refine_marked_elements(mark), it will ALWAYS refine the boundary elements (see attached, the second picture shows a slice passing through the center), regardless of my forcing the marking function to be 0 on boundary elements:

 W = FunctionSpace(msh,'DG',0)
 w = TestFunction(W)
 boundary_id = assemble(w*ds)

 if u.function_space().mesh().geometric_dimension() == 3:
      for id in range(len(mark.dat.data)):
           if boundary_id.dat.data[id] != 0:
                mark.dat.data[id] = 0

I checked my marking functions, and they were really zero on the boundary.

I migrated from NGSolve to Firedrake. In NGSolve, they had a similar problem, but can be fixed simply by stating that the boundary elements are not going to be refined with the following code

   if gfu.space.mesh.dim == 3:
      for el in gfu.space.mesh.Elements(BND):
         gfu.space.mesh.SetRefinementFlag(el, False)

Is there any similar fix in Firedrake?

[ksagiyam]

I think this is the same issue as discussed here: https://firedrakeproject.slack.com/archives/C1Q0Y6H8A/p1740128015584029.

@UZerbinati seems to have fixed it in NGSolve/ngsPETSc#74, so you probably just need an update.

[cwangaw]

Here is some information about the version of the related packages.

Requirement already satisfied: ngsPETSc in /home/wan01937/Workspace/venv-firedrake/lib/python3.12/site-packages (0.0.8)
Requirement already satisfied: netgen-mesher<7.0,>=6.2 in /home/wan01937/Workspace/venv-firedrake/lib/python3.12/site-packages (from ngsPETSc) (6.2.2501)
Requirement already satisfied: netgen-occt<8.0,>=7.8 in /home/wan01937/Workspace/venv-firedrake/lib/python3.12/site-packages (from ngsPETSc) (7.8.1)
Requirement already satisfied: numpy<3,>=2 in /home/wan01937/Workspace/venv-firedrake/lib/python3.12/site-packages (from ngsPETSc) (2.2.4)
Requirement already satisfied: petsc4py<4.0.0,>=3.22.1 in /home/wan01937/Workspace/venv-firedrake/lib/python3.12/site-packages (from ngsPETSc) (3.22.4)
Requirement already satisfied: scipy<2,>=1 in /home/wan01937/Workspace/venv-firedrake/lib/python3.12/site-packages (from ngsPETSc) (1.15.2)

[cwangaw]

Update: They have submitted a new commit solving the segmentation fault issue (NGSolve/netgen@7b13db7). However, the boundary elements are still being refined with the latest ngsPETSc.

[UZerbinati]

Amazing ! Thanks @ChrLackner for already solving the problem. Are you sure you are using the latest ngsPETSc ? You would not get it installing it using pip because I've not yet generated a pip release. You should clone the repo and install using pip install -e . :)

[cwangaw]

I see! Thanks! Yes, I was installing the latest release. Just to confirm: should I update my ngsPETSc in the same pip install -e . way after the segmentation fault problem fixed?

[cwangaw]

Or should I actually manually install the latest commit of NGSolve and update my ngsPETSc with git clone https://github.com/UZerbinati/ngsPETSc.git followed by pip install ./ngsPETSc? Sorry for my unfamiliarity with the architecture.