Enable disconnected neighbor detection across disconnected boundary pairs in libMesh #4283

ChengHauYang · 2025-10-21T15:39:04Z

Motivation

Currently, libMesh assumes physical continuity (base on node ID) between neighboring elements, which limits its ability to handle cases where two regions share coincident boundaries but are geometrically disconnected -- for instance, across interfaces with discontinuous fields or temperature jumps.
This PR introduces a mechanism to explicitly register disconnected boundary pairs, enabling the framework to reconstruct logical neighbor relationships even when elements are not directly connected in the mesh topology (base on node ID).
This functionality is particularly relevant for methods such as the Cohesive Zone Method (CZM), where field continuity is enforced weakly across interfaces.

This work is motivated by long-standing MOOSE discussions about “fake” or “pseudo” neighbors -- element pairs that were once topological neighbors but become disconnected due to node replacement (e.g., in BreakMeshByBlockGenerator).
See the following issues for historical context:

MOOSE #12033: Support for non-geometric neighbor connections -- initial proposal to allow interface coupling between geometrically separated regions.
MOOSE #21161: BreakMeshByBlock connectivity -- highlighted that after node replacement, elements that were formerly neighbors lose topological connectivity, motivating the need for a libMesh-level mechanism to handle such cases.

This PR effectively formalizes that capability at the libMesh level, turning what was previously a MOOSE-side workaround into a supported feature.

Design

Disconnected Boundary Framework: Introduced an API in MeshBase (add_disconnected_boundaries() and get_disconnected_boundaries()) that allows users to register and access disconnected boundary pairs. These pairs are stored as PeriodicBoundaries objects and used during neighbor search to establish logical connections between geometrically separated regions.
Neighbor Reconstruction and Validation: Extended UnstructuredMesh::find_neighbors() to detect and link virtual (fake, or disconnected) neighbors across disconnected interfaces (boundaries). A new regression test (disconnected_neighbor_test.C) verifies correct neighbor mapping and demonstrates a temperature-jump interface problem that matches the analytical solution.

Impact

Facilitates coupling between subdomains separated by non-matching meshes or duplicated boundaries.

roystgnr

I really like this, in principle, and the MOOSE test failures should be trivial to fix. Hopefully the distributed unit test sweep failures won't be much harder.

include/base/periodic_boundaries.h

include/mesh/mesh_base.h

src/base/periodic_boundaries.C

tests/systems/disconnected_neighbor_test.C

src/mesh/unstructured_mesh.C

roystgnr · 2025-10-21T19:40:19Z

Do we have discussion of the problem motivating this anywhere on Github that we can link to, for future reference? I know there's been multiple discussions of the "MOOSE can abuse neighbor pointers" issue somewhere, but I can't recall whether any were here rather than on Slack or over video.

I'm not sure how relevant the discussions would be, because IIRC in most of them my opinion had been "we should come up with a different mechanism for MOOSE to use instead of neighbor_ptr" and I think you've managed to convert me to "we should make it so that misuse of neighbor_ptr is actually a supported use", but it'd still be good to be able to look up the history.

roystgnr · 2025-10-21T19:43:45Z

Ooh. One thing I missed in the review: this has a good chance of failing in interesting ways as soon as we use it in the presence of refined elements on either side of the pseudo-neighbor links, doesn't it? And we don't have any coverage of that, do we? We need another unit test or two; it'd probably be adequate to just refactor your existing test but refine both elements in the second test and refine just one element in the third.

ChengHauYang · 2025-10-21T21:09:50Z

Thank you, @roystgnr ! I tried to add MOOSE #12033 and MOOSE #21161 to the Motivation section of the PR so that others can easily look up the historical context.

roystgnr · 2025-10-23T23:10:59Z

Gah, I forgot to re-activate CI for you. Sorry! I'm going to have you sent an Associates team invite so any future pushes can auto-activate CI.

roystgnr · 2025-10-23T23:12:52Z

Jobs manually activated, hopefully the last time that's necessary; let me know if the invite doesn't get to you or gives you any problems.

roystgnr · 2025-10-23T23:16:03Z

I'm also adding a few optional recipes that I'm not 100% sure will pass; I don't want to merge if this passes here only to find we have to fix something up for our devel->master merges.

roystgnr

The only definite fix we need is for that random whitespace futzing that got left around.

Oh, but we might also need to #ifdef LIBMESH_HAVE_SOLVER around those tests, since they involve a solve(); I'm not sure any of our regular recipes use such a decimated libMesh, but better safe than sorry.

src/base/periodic_boundaries.C

include/mesh/mesh_base.h

src/mesh/unstructured_mesh.C

roystgnr · 2025-10-24T20:40:25Z

Well, the good news is that the invite went through and CI is autorunning now. The bad news is that we're failing multiple recipes. "Test mac" is a red herring, but the distributed-mesh test looks like a real failure and I can't even imagine what is going on in the --disable-amr and --disable-deprecated tests.

ChengHauYang · 2025-10-24T22:19:24Z

Thank you, Roy, for your feedback. I think this issue is related to whether disconnected elements are being added as ghosted elements. I recently implemented a RelationshipManager in MOOSE, which allows me to properly handle disconnected neighbors on distributed meshes. However, this functionality is not yet available on the libMesh side. I plan to add a GhostingFunctor for disconnected neighbors in libMesh to see if that resolves the issue. Thanks again for taking the time to look into this.

moosebuild · 2025-10-25T02:41:34Z

Job Coverage, step Generate coverage on d615ffe wanted to post the following:

Coverage

	af134b	#4283 d615ff
	Total	Total	+/-	New
Rate	65.00%	65.03%	+0.03%	87.50%
Hits	76964	77045	+81	63
Misses	41434	41423	-11	9

Diff coverage report

Full coverage report

Warnings

New new line coverage rate 87.50% is less than the suggested 90.0%

This comment will be updated on new commits.

ChengHauYang · 2025-11-03T15:40:17Z

Dear @roystgnr ,

Thank you for your guidance. I’ve fixed the bugs related to distributed meshes, as well as those occurring when using --disabled-amr or --disabled-deprecated.

Here’s a brief summary of my recent changes:

Implemented a new GhostFunctor that utilizes point_locator and disconnected boundary pairs to add disconnected elements as ghost elements. This resolves the errors associated with distributed meshes.
Fixed a bug where internal boundaries could produce invalid side IDs when calling side_with_boundary_id under --disabled-amr. A corresponding unit test was added for verification.
Added three unit tests for disconnected neighbor handling and resolved test failures triggered by --disabled-deprecated. The previous failures occurred because fe->reinit(...) was called before fe->get_phi().

All tests are now passing. Could you please provide your feedback and review when you have a chance?

Thanks again!

roystgnr · 2025-11-03T17:11:09Z

I think that set of OverlappingCouplingGhostingTest failures in the Test No AMR recipe is a bug of mine; rebasing on devel again should fix it. But this:

3) test: DisconnectedNeighborTest::testTempJumpLocalRefineFail (F) 
expected exception not thrown

is from this PR - you'll want to disable that test when LIBMESH_ENABLE_AMR isn't defined.

Review to follow.

…n also access them. (b) add geometry-based detection of disconnected neighbors via paired boundaries - Introduced `Elem::geometrically_equal()` for tolerance-based geometric comparison of elements. - Replaced `_has_disconnected_neighbor` with `_boundary_id_pairs` for boundary-pair registration. - Updated `UnstructuredMesh::find_neighbors()` to auto-detect disconnected neighbors using geometric matching and paired boundaries.

- Introduced PeriodicBoundaries::disconnected_neighbor() to locate geometrically coincident elements across disconnected boundaries using PointLocator. - Replaced legacy geometry-based neighbor search in UnstructuredMesh::find_neighbors() with unified PeriodicBoundaries-based logic.

- Moved new MeshBase functions out of the header to use forward declarations, reducing unnecessary includes. - Reverted the PeriodicBoundaries::neighbor() logic to allow self-matching, since the existing boundary ID check already prevents unintended self-detections in periodic boundary cases.

(a) Revert periodic BC source file (b) Fix indentation in UnstructuredMesh::find_neighbors and wrap tests (c) involving solve() in DisconnectedNeighborTest with #ifdef LIBMESH_HAVE_SOLVER guards to ensure correct conditional compilation.

…rove)

The DisconnectedNeighborTest::testTempJump fails in certain CI environments (e.g., 'Test No AMR' recipe), potentially due to issues with the default solver/preconditioner selection, numerical errors (0 vs 0.05), or SEGVs observed in logs. To attempt to bypass these issues and diagnose the failure, this commit explicitly sets the linear solver to GMRES and no preconditioner.

Introduce `include_internal_boundary` switch to `side_with_boundary_id()` so we can locate boundaries even when a neighbor exists (e.g. manually disconnected interfaces). This fixes interface-side lookup in the temperature jump CZM-style test without requiring AMR fallback. Updated the test to explicitly use include_internal_boundary=true for interface IDs.

…l interface support This change introduces correct handling of intentionally disconnected interfaces (e.g., CZM or jump conditions) on distributed meshes: - Replace ID-based lookup in `MapBasedDisconnectedGhosting` with direct `Elem*` neighbor mapping to avoid invalid lookups post-renumbering. - Register the disconnected ghosting functor in `find_neighbors()` so ghosting stays consistent with the neighbor graph established by `prepare_for_use()`. - Extend `BoundaryInfo::side_with_boundary_id()` to optionally allow internal/logical boundaries when explicitly requested by the caller. - Fix face FE assembly under `--disable-deprecated` by explicitly requesting shape and map computations before reinit() calls. - Remove older per-test ghosting setup now handled by the mesh.

Restore previous behavior in `side_with_boundary_id()`. When AMR is disabled, internal boundaries incorrectly returned invalid_uint. Add fallback return for non-external sides. Add `testInternalBoundary()` to verify correct side detection on a QUAD4 mesh. Update `disconnected_neighbor_test` to use default API.

This commit removes the `map_based_disconnected_ghosting` implementation and fully migrates the build/runtime logic to `disconnected_neighbor_coupling`. - Remove map_based_disconnected_ghosting.{h,C} and all associated build rules - Add disconnected_neighbor_coupling.{h,C} into all Makefile targets - Drop ghosting implementation in UnstructuredMesh::find_neighbors() - Update tests to use `DisconnectedNeighborCoupling` directly - Enable fine-mesh test (testTempJumpRefine)

- Ensure boundary ID is added to a locally owned elem - Select valid left-bottom elem with neighbor for internal side

…ession test. - Added safety check in MeshRefinement::refine_elements() to prevent refinement when disconnected boundary interfaces are present. Refinement on such meshes can produce invalid neighbor relations (e.g., rev = invalid_uint). - Extended disconnected_neighbor_test with new testTempJumpLocalRefineFail() that verifies the exception is thrown.

ChengHauYang · 2025-11-03T18:24:34Z

Thank you, @roystgnr, for your carefully checking. I just fixed that test and rebased the code again.

roystgnr

Sorry for the "oh, whoops, why didn't I notice that before" change requests, when previously it had seemed we were a hair's breadth away from being done. Clearly my reviews need reviews.

build-aux/ltmain.sh

src/ghosting/disconnected_neighbor_coupling.C

src/mesh/boundary_info.C

src/mesh/mesh_refinement.C

src/mesh/unstructured_mesh.C

tests/systems/disconnected_neighbor_test.C

include/mesh/mesh_base.h

@roystgnr

Refactored the ghosting workflow following @roystgnr’s review. Removed the separate DisconnectedNeighborCoupling class and integrated its logic directly into GhostPointNeighbors for a cleaner and more consistent approach. This avoids duplicate functor additions in `UnstructuredMesh::find_neighbors`. - Deleted `DisconnectedNeighborCoupling.[C,h]` and updated Makefiles - Moved all related logic into `GhostPointNeighbors::operator()` - Added move assignment and equality checks for `_disconnected_boundary_pairs` in `MeshBase`. This was a great catch from Roy and is needed for mesh cloning/comparisons to work correctly. - Added `libmesh_deprecated()` warning to `UnstructuredMesh::stitching_helper` for disconnected meshes - Replaced a `libmesh_error_msg` in `MeshRefinement` with `libmesh_not_implemented_msg` - Updated unit tests: added `LOG_UNIT_TEST` for `disconnected_neighbor_test`

roystgnr

Getting closer, at least!

With subsequent changes, could you append commits rather than force-pushing them? It is literally asymptotically faster to re-review when I only need to look at new changes rather than all changes.

roystgnr · 2025-11-04T13:52:41Z

src/mesh/mesh_base.C

  _point_locator_close_to_point_tol = other_mesh.get_point_locator_close_to_point_tol();

+#ifdef LIBMESH_ENABLE_PERIODIC
+  _disconnected_boundary_pairs = std::move(other_mesh._disconnected_boundary_pairs);


This should probably work for the move constructor (it's possible for users to derive a PeriodicBoundary subclass that would break here because its clone() kept a reference to the original mesh, but they'd have to work at it), but we need a copy in the copy constructor above too.

The copy constructor copy is going to be a bit trickier - each entry in the new PeriodicBoundaries will need to be a clone() of the corresponding entry from the source.

Thanks, @roystgnr! In both the move assignment operator and the copy constructor, I now perform a deep copy of each entry in _disconnected_boundary_pairs using clone().

Hmm... that's probably not necessary in the move constructor, but it probably is safer; I like it.

src/mesh/mesh_base.C

…nment - Added deep copy logic for `_disconnected_boundary_pairs` in both the copy constructor and move assignment operator of `MeshBase`, ensuring that `clone()` preserves disconnected boundary data. - Updated `locally_equals()` to perform a safer weak comparison on `_disconnected_boundary_pairs` (compare sizes only). - Added `DisconnectedNeighborTest::testCloneEquality()` to verify that cloned meshes correctly replicate disconnected boundaries.

roystgnr

I want to give @jwpeterson a chance to look at this too before merging, but I think this is ready to go so unless John or CI scream at us we'll merge tomorrow.

I'm going to add a bunch of the devel->master recipes to this PR first, too, though, just to make sure I'm not missing anything that we might get hung up on after merging.

roystgnr · 2025-11-04T18:04:44Z

This was kind of a brutal review process; sorry! Thanks for pushing through it with me! Though, honestly, even after all that this is a much less massive change than I'd have expected it to become. We are still at the most basic stages if we're not supporting stitching or refinement yet, but I wouldn't have expected even the most basic support to be possible with only ~400 lines of library code. I've still got a "waiting for the other shoe to drop" feeling, and I won't be too surprised if trying to use this in MOOSE this brings up some unpleasant surprises, but with ~600 lines of tests hopefully we've preempted anything too nasty.

ChengHauYang · 2025-11-04T18:13:28Z

Thank you, @roystgnr, for providing me with much constructive feedback during this process! Really appreciate your guidance!

hugary1995 · 2025-11-04T18:20:20Z

We should probably test idaholab/moose#31636 against this PR branch and make sure it's green.

jwpeterson · 2025-11-04T19:08:45Z

I want to give @jwpeterson a chance to look at this too before merging,

It sounds like you've already reviewed it pretty thoroughly so 👍 from me.

roystgnr requested changes Oct 21, 2025

View reviewed changes

ChengHauYang force-pushed the break_mesh_libmesh_pr branch from e5f64ed to 120b7c1 Compare October 21, 2025 19:37

roystgnr requested changes Oct 23, 2025

View reviewed changes

src/base/periodic_boundaries.C Outdated Show resolved Hide resolved

include/mesh/mesh_base.h Outdated Show resolved Hide resolved

roystgnr reviewed Oct 24, 2025

View reviewed changes

src/mesh/unstructured_mesh.C Outdated Show resolved Hide resolved

ChengHauYang force-pushed the break_mesh_libmesh_pr branch from 99faccb to 64e4597 Compare October 24, 2025 15:48

ChengHauYang force-pushed the break_mesh_libmesh_pr branch 6 times, most recently from a7b9cfb to 71680c0 Compare October 28, 2025 22:06

ChengHauYang added 8 commits November 3, 2025 12:21

add new unit test for disconnected neighbor in libmesh

c90120a

Add translations customization for disconnected neighbor

114972c

Fix code base on Roy's feedback

0b5260d

(a) Revert periodic BC source file (b) Fix indentation in UnstructuredMesh::find_neighbors and wrap tests (c) involving solve() in DisconnectedNeighborTest with #ifdef LIBMESH_HAVE_SOLVER guards to ensure correct conditional compilation.

fix unit tests

ac5c088

[WIP] simple ghosting functor for disconnected neighbor (rooms to imp…

c034b84

…rove)

ChengHauYang added 9 commits November 3, 2025 12:21

add the ghost functor directly inside prepare_for_use()/find_neighbors()

53cd73a

Fix distributed test failures in internal-boundary

a86e595

- Ensure boundary ID is added to a locally owned elem - Select valid left-bottom elem with neighbor for internal side

add amr flag for a unit test

38787d8

ChengHauYang force-pushed the break_mesh_libmesh_pr branch from f956256 to 38787d8 Compare November 3, 2025 18:22

roystgnr requested changes Nov 3, 2025

View reviewed changes

ChengHauYang force-pushed the break_mesh_libmesh_pr branch 2 times, most recently from 8271f2d to 9857539 Compare November 3, 2025 22:20

ChengHauYang force-pushed the break_mesh_libmesh_pr branch from 9857539 to d100e50 Compare November 4, 2025 00:55

roystgnr requested changes Nov 4, 2025

View reviewed changes

roystgnr approved these changes Nov 4, 2025

View reviewed changes

Enable disconnected neighbor detection across disconnected boundary pairs in libMesh #4283

Are you sure you want to change the base?

Enable disconnected neighbor detection across disconnected boundary pairs in libMesh #4283

Uh oh!

Conversation

ChengHauYang commented Oct 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Design

Impact

Uh oh!

roystgnr left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

roystgnr commented Oct 21, 2025

Uh oh!

roystgnr commented Oct 21, 2025

Uh oh!

ChengHauYang commented Oct 21, 2025

Uh oh!

roystgnr commented Oct 23, 2025

Uh oh!

roystgnr commented Oct 23, 2025

Uh oh!

roystgnr commented Oct 23, 2025

Uh oh!

roystgnr left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

roystgnr commented Oct 24, 2025

Uh oh!

ChengHauYang commented Oct 24, 2025

Uh oh!

moosebuild commented Oct 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Coverage

Warnings

Uh oh!

ChengHauYang commented Nov 3, 2025

Uh oh!

roystgnr commented Nov 3, 2025

Uh oh!

ChengHauYang commented Nov 3, 2025

Uh oh!

roystgnr left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

roystgnr left a comment

Choose a reason for hiding this comment

Uh oh!

roystgnr Nov 4, 2025

Choose a reason for hiding this comment

Uh oh!

roystgnr Nov 4, 2025

Choose a reason for hiding this comment

Uh oh!

ChengHauYang Nov 4, 2025

Choose a reason for hiding this comment

Uh oh!

roystgnr Nov 4, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

roystgnr left a comment

ChengHauYang commented Oct 21, 2025 •

edited

Loading

moosebuild commented Oct 25, 2025 •

edited

Loading