PhysicsNeMo-Mesh: adds DomainMesh for combined manipulation of volume + surface meshes.

[peterdsharpe]

PhysicsNeMo Pull Request

Description

This is a proposal for a multi-mesh, which stores volume and surface data together (e.g., what you'd get from a single CFD simulation) and allows for easy manipulation of them together.

DomainMesh is a tensorclass that groups an interior Mesh with zero or more named boundary Mesh objects and optional domain-level metadata (global_data). The interior and boundary meshes can have any combination of manifold dimensions (volumetric, surface, graph, point cloud); the only structural invariant enforced at construction is that all meshes share the same spatial embedding dimension. Dimensional constraints on individual meshes can be checked via Mesh's own parametric types (e.g., isinstance(dm.interior, Mesh[3, 3])).

All Mesh operations that make sense at the domain level are exposed as passthrough methods via a private _map_meshes helper that applies a Mesh -> Mesh function to the interior and every boundary using TensorDict.apply(fn, call_on_nested=True), preserving global_data unchanged. This includes geometric transforms (translate, rotate, scale, transform), cleanup and refinement (clean, strip_caches, subdivide), data conversion (cell_data_to_point_data, point_data_to_cell_data), and calculus (compute_point_derivatives, compute_cell_derivatives). A validate method aggregates per-mesh validation reports into a single domain-level result. All methods return DomainMesh and support fluent chaining (e.g., dm.translate([1, 0, 0]).scale(2.0).clean()).

Note: this is a stacked PR on top of:

• [stacked] Adds caching of neighbor information to PhysicsNeMo-Mesh #1503
• Adds parametric type system for PhysicsNeMo-Mesh #1500
  Do not review or merge until this is in (otherwise the diff will be incorrect)

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.
• If I am implementing a new model or modifying any existing model, I have followed the Models Implementation Coding Standards.

Dependencies

Review Process

All PRs are reviewed by the PhysicsNeMo team before merging.

Depending on which files are changed, GitHub may automatically assign a maintainer for review.

We are also testing AI-based code review tools (e.g., Greptile), which may add automated comments with a confidence score.
This score reflects the AI’s assessment of merge readiness and is not a qualitative judgment of your work, nor is
it an indication that the PR will be accepted / rejected.

AI-generated feedback should be reviewed critically for usefulness.
You are not required to respond to every AI comment, but they are intended to help both authors and reviewers.
Please react to Greptile comments with 👍 or 👎 to provide feedback on their accuracy.

[greptile-apps]

Greptile Summary

This PR introduces DomainMesh, a tensorclass that groups an interior Mesh with zero or more named boundary Mesh objects and optional domain-level metadata (global_data), along with a suite of passthrough transform, cleanup, data-conversion, and validation methods. The design is clean, well-documented, and integrates naturally with the existing Mesh/TensorDict infrastructure.

Key observations:

• Shared global_data reference in _map_meshes: Every transform method calls _map_meshes, which forwards global_data=self.global_data without cloning. Because __post_init__ only mutates batch_size (a no-op on an already-normalised TensorDict), both the original and the transformed DomainMesh end up pointing at the same TensorDict. An in-place write on the result (e.g. dm2.global_data["Re"].fill_(0.0)) silently propagates back to the source. A self.global_data.clone() in _map_meshes would decouple the two instances.
• n_boundaries materialises an unnecessary list: len(list(self.boundaries.keys())) can be simplified to len(self.boundaries).
• Missing test coverage: compute_point_derivatives and compute_cell_derivatives are the only passthrough methods not exercised by the new test file.
• mesh_repr.py improvement: Skipping empty data fields makes the repr cleaner; the change is correct.

Important Files Changed

Filename	Overview
physicsnemo/mesh/domain_mesh.py	New DomainMesh tensorclass grouping an interior Mesh with named boundary Mesh objects and optional global_data. Well-structured with thorough docstrings. Minor issue: global_data is passed by reference in _map_meshes, so transformed instances share the same TensorDict object, enabling unintended cross-instance in-place mutations. Also, n_boundaries materialises an unnecessary list.
test/mesh/test_domain_mesh_transforms.py	Comprehensive test suite covering geometric transforms, cleanup/refinement, data operations, validation, and fluent chaining. Missing test coverage for compute_point_derivatives and compute_cell_derivatives.
physicsnemo/mesh/utilities/mesh_repr.py	Small improvement to format_mesh_repr: empty data fields (point_data, cell_data, global_data) are now skipped, making the representation cleaner for sparse meshes. Change is correct and well-implemented.
physicsnemo/mesh/init.py	Adds DomainMesh to the public physicsnemo.mesh namespace. Straightforward, no issues.

_{Last reviewed commit: db2e28d}

[coreyjadams]

Looking great! Overall, I think this is awesome functionality and we should move forward. It's really valuable for the CAE effort but it's going to be useful beyond that as well. I left some comments; I really have just one that I consider an "issue" and that is this: I think we discussed the possibility to encode things like global vectors (inlet velocity, for example) with a DomainMesh, that would transform with the DomainMesh. As implemented I think those would fall through the cracks - am I missing something?

I also made some comments about meta data but we can keep that as a conversation going on the design, rather than something to resolve before merge.

[coreyjadams]

Thanks for making the updates to the global data tensors and aligning the apis with Mesh. I think this is good to go!

[peterdsharpe]

/blossom-ci

[peterdsharpe]

/blossom-ci