Refactor periodic utility meshes by connorjward · Pull Request #4836 · firedrakeproject/firedrake

connorjward · 2026-01-23T11:37:17Z

Summary of changes

Use (where possible) DMPlex.createBoxMesh to generate our periodic meshes. This is needed because the current approach that we use results in an invalid DMPlex that cannot subsequently be extruded. This PR is therefore a blocker for pyop3.
Uses a DMPlexTransform to perform "crossed", "left" or "right" refinement instead of rolling our own solution.
Deprecate PartiallyPeriodicRectangleMesh. It's functionality is totally covered by PeriodicRectangleMesh.
Set the 'Face Sets' label of the mesh (the boundary markers) by renumbering the existing one DMPlex provides, instead of performing our own comparison and traversal to rediscover already known structure.

Breaking API changes

For a 3D periodic box mesh we currently label all of the boundaries, meaning that the periodic boundary can be used in interior facet integrals. It substantially simplifies the implementation if we don't do this and only label the non-periodic boundaries. Therefore one can no longer do dS integrals along the periodic boundary for 3D box meshes.

Prerequisite PRs

This is necessary because the previous approach to generating periodic meshes resulted in a DMPlex with a slightly invalid state that prevents subsequent transformations (in particular extrusion). We also avoid manually labelling the boundaries in favour of just renumbering the existing 'Face Sets' label produced by the DMPlex.

Co-authored-by: Connor Ward <c.ward20@imperial.ac.uk>

…dmplex

connorjward · 2026-03-24T13:58:35Z

@pbrubeck or @JHopeCollins are either of you able to review this? It looks daunting but the changes aren't actually that huge.

For the tests I have had to reorder some arrays because the numbering is different. The tests are not ideal in this way but I think making more robust tests is out of scope here.

There is still one PETSc fix that I need to have merged but it doesn't change the Firedrake code so this can therefore be reviewed now.

pbrubeck

The BC label renumbering could be in its own PR, as it is an API breaking change

The firedrake mesh utility refactor (firedrakeproject/firedrake#4836, merged 31 March 2026) changed how UnitSquareMesh generates triangle meshes, now using PETSc refine_tosimplex DMPlexTransform instead of a hand-rolled quad-to-triangle split. The meshes are equally valid but have different cell ordering, shifting numerical results slightly. These tests do not import gadopt. They are pure Firedrake/pyadjoint tests, so the failures on PR #490 were from this upstream change, not from the solution_old reordering proposed there.

Changes are due to node ordering in utility meshes since firedrakeproject/firedrake#4836. In test_hessian_recovery2d and test_vorticity_calculation2d these are just tolerance tweaks. For the bottom friction test with bdm-dg I think this has highlighted that the implicit (vertical) momentum solver using a single ilu(0) sweep does not work very well for bdm-dg (also rt-dg seems to not be very stable). Need to find more general solution. In the test, switching to ILU(1) appears to make it stable again.

connorjward commented Mar 3, 2026

View reviewed changes

Comment thread .github/workflows/core.yml Outdated

connorjward force-pushed the connorjward/periodic-dmplex branch from 2984ac1 to 64d1a00 Compare March 3, 2026 13:59

connorjward commented Mar 4, 2026

View reviewed changes

Comment thread .github/workflows/core.yml Outdated

connorjward force-pushed the connorjward/periodic-dmplex branch from 792a7c5 to fcf088c Compare March 6, 2026 13:55

connorjward added 4 commits March 6, 2026 15:46

linting

6a4fd9a

vom fix

f4af999

submesh fix

8bf3ca2

More fixes, all of them?

077d1d4