Add unstructured mesh infrastructure - #5687
Merged
Merged
Conversation
Introduce UnstructuredSubMesh, mesh generators, and Mesh interface coupling so arbitrary 2D/3D domains can be discretised later. Extracted from the unstructured finite-volume stack for isolated review. Co-authored-by: Cursor <cursoragent@cursor.com>
This was referenced Jul 31, 2026
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #5687 +/- ##
==========================================
- Coverage 98.21% 98.09% -0.12%
==========================================
Files 339 340 +1
Lines 31996 32671 +675
==========================================
+ Hits 31424 32049 +625
- Misses 572 622 +50 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
`_combine_unstructured_submeshes` put 110 lines of unstructured-specific node welding in the generic `meshes.py`, which then needed two function-level `from .unstructured_submesh import ...` calls to dodge the module-level cycle (`unstructured_submesh` already imports `SubMesh` and `MeshGenerator` from `meshes`). Move it to `UnstructuredSubMesh.combine`, where `_hex_to_tet` and the class itself are local, and reach `compute_interface_data` through the `pybamm` namespace like the rest of `meshes.py` does. `meshes.py` no longer references `unstructured_submesh` at all. Pure move: the welding and boundary-tag-recovery bodies are unchanged apart from `pybamm.UnstructuredSubMesh(...)` becoming `cls(...)`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…_ordering `interface_data` stores "left_cells" as this mesh's cells and "right_cells" as the neighbour's, but `optimize_ordering` permuted both with this mesh's `inv_perm`. That silently re-pointed the neighbour's indices, and where the neighbour has fewer cells it produced out-of-range indices (an 18-cell mesh next to a 12-cell one yielded `right_cells.max() == 17`). `compute_interface_data` also gives the neighbour a mirrored view of the same pairing, which aliases the very array being replaced, so the mirror went stale. Permute only this mesh's indices and write the result through to the neighbour's mirror. The existing test asserted only on "left_cells" — the one key that was already correct — so add a case that pins the neighbour's indices and the mirror, using meshes of different sizes so a mis-permutation is caught by bounds rather than by luck. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`UnstructuredSubMesh.combine` welds coincident interface nodes so the seam becomes internal faces and TPFA carries flux across it. When the interface nodes don't coincide — mismatched transverse grids, wrong coordinate units, or a non-conforming input mesh — nothing welds, the seam stays boundary faces on both sides, and the domains form separate connected components. The solve then runs to completion with the regions silently decoupled and returns a physically wrong answer. Add a post-weld check that the combined mesh is a single connected component, using the integer face_owner/face_neighbor connectivity so it introduces no distance tolerance. Under the conforming-interface requirement welding is all-or-nothing, so total disconnection is the only failure mode and this check is sufficient. Document the conforming requirement on the two gmsh-reading generators. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
3 tasks
Four review findings on #5687: - Reject non-manifold input: a face shared by more than two cells matched neither the internal (count 2) nor boundary (count 1) branch of _build_face_connectivity and silently vanished from the face list, losing flux paths. Now raises GeometryError. - Make the 3D generator honour element_type: _generate_3d always built hexes regardless of the argument (docstring claimed a tetrahedron default). Hexahedron stays the default, matching existing tests and the 3D DFN model; "tetrahedron" now works and records _hex_gen_params so combine() can rebuild tet domains with a cumulative parity offset, keeping interface triangulations conforming — previously that combine branch was unreachable. Unknown types raise. - Remove dead code: _cell_faces (unused since the vectorised connectivity path) and an n_fpc assignment overwritten two lines later. - Conventions: framework errors now raise pybamm.GeometryError instead of ValueError/KeyError/RuntimeError; meshio is imported via pybamm.import_optional_dependency; the CHANGELOG bullet carries the PR link. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Codecov flagged the meshio-backed generators as entirely untested. Add tests for UserSuppliedUnstructuredMesh (whole-file load with 2D trimming, subdomain filtering, missing-cell-data and unsupported-cell-type errors, lims-to-domain mapping) and TaggedSubMeshGenerator (region extraction with coordinate scaling, missing region, region with no tets), using synthetic meshio objects injected into the class cache so no .msh round-trip is needed. Also cover combine()'s custom boundary-tag recovery, compute_interface_data's no-matching-faces and transverse-mismatch raises, and Mesh leaving interface_data empty when grids don't pair. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
MarcBerliner
reviewed
Aug 3, 2026
Volumes from the 5-tet split are ill-defined on warped (non-planar-faced) hexes: the two valid diagonal splits differ by ~20% on a displaced-corner cube. Enforce planar faces at construction and restrict UserSuppliedUnstructuredMesh to tetra/triangle/quad cells, since gmsh/VTK hex meshes of non-box geometry are warped in general. Pattern-A tets were also inconsistently oriented (signed volumes sum to zero on a unit cube), masked by the per-tet abs. Reorder to a consistent orientation and sum signed so inverted or degenerate cells become detectable. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
numpy 2.0.0 (the declared floor) returns return_inverse with shape (n, 1) instead of (n,) — a regression reverted in 2.0.1. The 2-D inverse makes every face-connectivity mask 2-D and construction raises IndexError on any mesh. reshape(-1) at both np.unique sites is a no-op on fixed numpy versions. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 5-tet hex split must alternate mirror patterns cell-to-cell to keep shared-face diagonals conforming, so multi-domain meshes needed a cumulative parity offset. combine() applied it by regenerating every domain after the first but never wrote the result back to the Mesh: with an odd preceding nx, mesh[domain] and the combined view described geometrically different cells (silent wrong volumes and connectivity for the same state entries). The Kuhn (Freudenthal) 6-tet split is translation-invariant — every hex splits identically and opposite faces carry the same face-local diagonal — so seams conform with no parity bookkeeping. Delete the regeneration block and _hex_gen_params breadcrumb entirely. Strengthen test_combine_tetrahedron_domains_conforming with an odd left nx (the old even-nx choice made the offset branch a no-op) and assert both views agree cell-for-cell. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Vertex-mean centroids are exact only for simplices. On trapezoid quads the error is 12% in the off-axis coordinate; TPFA builds transmissibility from centroid-to-centroid distances, so distorted 2D file meshes silently lost first-order accuracy. Use the polygon (shoelace) centroid for quad cells, signed-volume-weighted tet centroids for hex cells, and area-weighted triangle centroids for quad faces — all exact for the planar-faced cells the mesh accepts, verified against a symmetric trapezoid and a square frustum. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Auto-detection bucketed every exterior face of any mesh by dominant normal axis, so a file mesh of curved geometry (e.g. a triangulated disc) got six 'left'/'right'/... buckets that are not surfaces, and downstream BC/interface code consumed them silently. Construction now assigns no tags by default; the detection becomes an explicit detect_box_boundaries() that the built-in generator calls on its own output, where the assumption is true by construction. combine() propagates all input tags by centroid matching instead of re-detecting on the welded mesh. File meshes get no guessed names — boundary tags must come from the mesh file or be set explicitly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
boundary_mapping was stored and never read, and with guessed box tags gone file meshes had no way to name their boundaries. Tagged facet blocks (gmsh physical surface groups, or any cell-data key) are now matched to submesh boundary faces by sorted vertex indices — through the quantisation weld and node compaction — and become named boundary_faces entries. TaggedSubMeshGenerator gains the same parameter, resolving group names via field_data. Unmatched entries log a warning instead of vanishing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A swallowed GeometryError in Mesh._compute_unstructured_interfaces meant a domain pair could silently lose all flux coupling — the exact failure combine() raises for. Keep the no-raise property of Mesh.__init__ but log which pair was skipped and why. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Consecutive-order pairing stapled whichever domains happened to be
adjacent in the geometry dict: declaring {separator, negative
electrode} out of spatial order paired sep.right (x=2) with neg.left
(x=0), and because interface matching uses transverse coordinates
only, the bogus coupling succeeded silently with cell distances
spanning the whole geometry. Pair domains by comparing the planes of
their right/left boundary faces instead, making declaration order
irrelevant.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Combining a triangle domain with a quad domain leaked numpy's raw concatenate ValueError. Check up front and raise a GeometryError naming both element types. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The weld (1e-9 absolute), tag-match (1e-10 * max(ptp, 1)), and interface-match (1e-8 * max(ptp, 1)) tolerances were unrelated to the mesh's own length scale: the max(..., 1.0) floors made the relative ones absolute for every sub-metre (i.e. every SI battery) mesh, a 3 nm interface jitter on a 100 um domain rejected the whole mesh as disconnected, and tags could vanish because tag matching was tighter than welding. Derive one tolerance from the smallest element edge (1e-3 of it): distinct nodes are at least an edge apart, so no false merges, while file-precision jitter is absorbed. The file-loader quantisation stays separate (a snapping grid, now documented). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The nearest-neighbour interface match is one-directional: a surplus boundary face on the right mesh silently lost its flux, and a doubly claimed face would double-count it. Require equal face counts and unique pairing, and document the assumption that domains stack along x (faces are matched by transverse coordinates). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Everywhere else in PyBaMM, submesh 'nodes' means cell centres (SubMesh1D.nodes) and mesh vertices are 'edges'. UnstructuredSubMesh used 'nodes' for the (n, d) vertex coordinate array, colliding with that vocabulary just as three PRs stack on top of this one. Rename before anything else hard-codes it. BREAKING CHANGE: UnstructuredSubMesh.nodes is now UnstructuredSubMesh.vertices (constructor parameter renamed too). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The per-cell Python loop over 5 tets was ~95% of construction time (~2 s for a 30^3 hex grid). One einsum over the (n_cells, 5, 4, 3) tet array computes identical signed volumes and volume-weighted centroids; 27k-cell construction now takes ~0.1 s end to end. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
contains_points_3d looped over boundary triangles and vectorised over query points, making its cost O(n_triangles) Python iterations even for a single point (~0.3 s for 2 points against 5400 faces). Swap the nesting: few query points iterate, the triangle arithmetic vectorises. Identical boolean output. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The weld built a Python dict per submesh, appended nodes one at a time, and remapped elements with a nested list comprehension — 4M dict lookups for a 1M-cell tet mesh. Replace with a KD-tree query plus cumsum-based index map; identical welded mesh. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replaces the per-face Python loop with array bucketing, and removes its silent tie-break: a zero-area face normalises to a zero normal, argmax returned axis 0, and the face landed in 'right'. Degenerate boundary faces now raise instead. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pure-Python double/triple loops built element lists row by row (~0.05 s at 27k hexes, ~2 s extrapolated at 1M). Vertex ids are affine in (i, j, k), so meshgrid plus index arithmetic produces byte-identical connectivity, verified against the loop versions across grid sizes including the bandwidth-sorted hex ordering. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Review request (MarcBerliner). The str mixin keeps plain strings working everywhere an ElementType is expected, so element_type="quad" and downstream string comparisons are unaffected. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
MarcBerliner
previously approved these changes
Aug 4, 2026
MarcBerliner
left a comment
MarcBerliner
left a comment
Member
There was a problem hiding this comment.
Looks good to go, one non-blocking thing
Comment on lines
+958
to
+968
| if any(block.type == "hexahedron" for block in mesh.cells): | ||
| raise pybamm.GeometryError( | ||
| "Hexahedral cells in mesh files are not supported: warped " | ||
| "(non-planar-faced) hexahedra have ill-defined volumes and " | ||
| "face fluxes. Convert the mesh to tetrahedra (e.g. with " | ||
| "gmsh or meshio) and reload. Hexahedral meshes are still " | ||
| "available through pybamm.UnstructuredMeshGenerator, whose " | ||
| "axis-aligned cells are always well-defined." | ||
| ) | ||
| # Prefer 3D cells when present, otherwise fall back to 2D. | ||
| for cell_type in ("tetra", "triangle", "quad"): |
Contributor
Author
There was a problem hiding this comment.
Done in 0c92328 — the loop compares enum members now. One subtlety: meshio's vocabulary says tetra where ours says tetrahedron, so ElementType gained a meshio_name property holding that one divergence rather than scattering both spellings.
Review request (MarcBerliner). meshio's vocabulary says 'tetra' where ElementType says 'tetrahedron', so the enum gains a meshio_name property holding the one divergence; all loader comparisons now go through enum members instead of magic strings. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
aabills
enabled auto-merge (squash)
August 4, 2026 16:19
MarcBerliner
approved these changes
Aug 4, 2026
aabills
disabled auto-merge
August 4, 2026 18:00
aabills
enabled auto-merge (squash)
August 4, 2026 18:05
aabills
added a commit
that referenced
this pull request
Aug 6, 2026
Adds FiniteVolumeUnstructured (TPFA Laplacian, fused div_D_grad, Green-Gauss gradient), unstructured processed variables, and the discretisation dispatch for div(D*grad(u)) and graph-topology internal boundary conditions, on top of the unstructured meshing (#5687) and N-component VectorField (#5686) already on main. Includes the fixes from the review of the previous revision: - Neumann values on named axis sides are coordinate-direction derivatives (matching FiniteVolume); custom tags stay outward-normal - unknown BC sides and bc_types raise DiscretisationError instead of being silently dropped; boundary_integral supports "entire" - auxiliary domains are handled in laplacian/gradient BC assembly and in secondary/tertiary broadcasts - divergence of a BC-bearing flux raises rather than silently dropping the boundary flux; the Green-Gauss gradient warns for BC-less buckets - scalar reductions (Max/Min) on unstructured domains post-process as 0D - operator matrices are cached per submesh (invalidated when the cell ordering changes) and BC assembly is vectorised - ParameterSubstitutor.process_boundary_conditions processes every boundary side present, not a fixed whitelist, so tab and named-region Dirichlet conditions are no longer silently dropped Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This was referenced Aug 7, 2026
aabills
added a commit
that referenced
this pull request
Sep 5, 2026
* feat: add unstructured finite volume spatial method Adds FiniteVolumeUnstructured (TPFA Laplacian, fused div_D_grad, Green-Gauss gradient), unstructured processed variables, and the discretisation dispatch for div(D*grad(u)) and graph-topology internal boundary conditions, on top of the unstructured meshing (#5687) and N-component VectorField (#5686) already on main. Includes the fixes from the review of the previous revision: - Neumann values on named axis sides are coordinate-direction derivatives (matching FiniteVolume); custom tags stay outward-normal - unknown BC sides and bc_types raise DiscretisationError instead of being silently dropped; boundary_integral supports "entire" - auxiliary domains are handled in laplacian/gradient BC assembly and in secondary/tertiary broadcasts - divergence of a BC-bearing flux raises rather than silently dropping the boundary flux; the Green-Gauss gradient warns for BC-less buckets - scalar reductions (Max/Min) on unstructured domains post-process as 0D - operator matrices are cached per submesh (invalidated when the cell ordering changes) and BC assembly is vectorised - ParameterSubstitutor.process_boundary_conditions processes every boundary side present, not a fixed whitelist, so tab and named-region Dirichlet conditions are no longer silently dropped Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix: tile per-face BC vectors across auxiliary-domain repeats in div_D_grad A BC value with one entry per boundary face broke with a ShapeError when the variable had auxiliary domains; lift it to n_bnd * repeats entries, matching _bc_contribution's convention. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: raise on y= queries of 2D unstructured processed variables y was silently dropped when building the query grid, returning z-midplane values for a query the user thinks is at y; raise like the existing r/R check. 3D unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * perf: interpolate all time steps of unstructured variables in one pass scipy's interpolators accept (n_points, n_t) value arrays, so the nearest-neighbour KDTree and the domain boundary mask are now built once per query instead of once per time step (~40x on a 100-step animation). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: raise DomainError for unresolvable unstructured spatial variables spatial_variable silently fell back to the x column for unknown names and directions, matched loose prefixes (zeta -> z), and accepted y/fb on 2D x-z meshes. Match the leading name token exactly and raise otherwise. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * refactor: define set_internal_bcs_for_concat on the base SpatialMethod Replace the hasattr duck-typing dispatch with a documented base-class hook returning None (use the default 1D-stack routine); the unstructured method's implementation becomes a plain override. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: match legacy tab BC sides exactly instead of by substring With arbitrary boundary tags now allowed, a region name containing 'tab' (e.g. 'tab_weld') was misrouted into check_tab_conditions and raised ModelError; share one exact-name set across both check sites. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * style: pre-commit fixes * fix: fail loudly when unstructured meshes lack tags or interface data build() now warns for a submesh with exterior faces but no boundary tags (BCs cannot apply, interface discovery cannot pair it), and _internal_neumann_unstructured raises instead of returning zeros when no interface pairs two meshes, which silently decoupled the domains. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Add implicit non-orthogonal correction to the unstructured TPFA operators The two-point flux only carried alpha (u_j - u_i)/d with alpha = cos(theta) and dropped the k . grad(u) remainder of n = alpha e + k, so the Laplacian of a linear field was nonzero on any non-orthogonal mesh and the scheme did not converge on triangles or tetrahedra (L2 rate ~0 under refinement). - Split the face normal as n = alpha e_ij + k and add the k . grad(u)_f cross flux implicitly, for internal faces (laplacian, div_D_grad) and Dirichlet boundary faces (perpendicular distance delta . n). - "non-orthogonal correction" option: "over-relaxed" (default, alpha = 1/cos theta, floored) or "minimum" (alpha = cos theta). - Replace the Green-Gauss gradient with a batched weighted least-squares reconstruction that is exact on linear fields on skewed meshes; internal, Dirichlet and Neumann faces each contribute one directional-derivative row. Green-Gauss remains the divergence assembly. - Raise GeometryError for inverted cells, warn at build above 70 degrees of non-orthogonality, skip the correction entirely on orthogonal meshes. - _divergence_matrices now aliases _green_gauss_matrices. Linear patch test is at 1e-13 on tri, perturbed tri and Kuhn tet meshes; Poisson convergence rates are 1.95-2.04 (tri) and 1.74-1.88 (tet) for both options, unchanged (2.0) on quad/hex. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Test the build-time non-orthogonality warning; make zip strict Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Return a pybamm.Matrix from the unstructured definite integral definite_integral_matrix returned a raw csr_matrix, so DefiniteIntegralVector failed with AttributeError in process_symbol's shape check; it also ignored vector_type (a "column" request came back as a row) and integral() dropped integration_dimension, silently integrating the primary volume for a secondary-dimension Integral. Support row/column, raise NotImplementedError for non-primary dimensions, and lift by the auxiliary-domain repeats. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Move point-in-domain onto UnstructuredSubMesh; skip nearest fill outside it UnstructuredSubMesh.contains_points now owns the 2D even-odd loop test (loops cached on the mesh) and delegates to contains_points_3d, so the processed variable only asks the mesh. In _interpolate_spatial the outside mask is computed first and excluded from the nearest-neighbour fill, which previously filled ~97% of out-of-hull points only to overwrite them with fill_value. Also correct the get_quiver_data docstring for the 3D return. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Apply the non-orthogonal correction across unstructured domain interfaces The interface Neumann value between concatenated domains was the plain two-point difference (u_r - u_l)/d, the same alpha = 1 truncation that made the Laplacian inconsistent inside a domain, so a linear field was not a steady state of a two-domain tet model (residual 4.6 on the Kuhn mesh). - internal_neumann_condition now returns alpha (u_r - u_l)/d + k . grad(u)_f with the same n = alpha e + k split as the interior; the face gradient is the distance-weighted mean of both sides' least-squares gradients, which take the interface faces as cross-mesh rows towards the paired cell and each side's external boundary conditions (passed by set_internal_bcs_for_concat). Orthogonal interfaces are unchanged. - interface_data records left_faces/right_faces on both pairing paths. - Least-squares rows for boundary faces without a condition now fit a zero normal derivative, matching the operators' zero-flux treatment, so cells on untagged boundaries stay fully determined. - UnstructuredSubMesh.contains_points gets a direct test. Two-domain linear field on tets: residual 4.6 -> 5e-14; interface flux is conservative to 1e-10 for a random field. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Test the corrected interface gradient on a linear field The old expectation hard-coded the two-point (u_r - u_l)/d formula, which is only the full interface gradient on orthogonal pairs; check u = x gives exactly 1 on the skewed triangle interface and keep the two-point check for a quad interface where it still holds. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Rename the axis token variable that Bandit mistakes for a password Codacy's Bandit B105 flags `token == "y"` as a hardcoded password purely because of the variable name. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs: drop duplicate Component/Norm entries after merging main main already documents both in unary_operator.rst; the second copy made Sphinx fail with "duplicate object description" under -W. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Cover the inconsistent-face-count error and the no-loops containment path Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Treat every scalar-shaped boundary value as a broadcast scalar Time- or input-dependent scalars (e.g. a current-density Neumann value) evaluate for shape to () rather than (1, 1), so _bc_contribution built Matrix @ scalar and failed once div_D_grad started assembling the least-squares gradient with boundary rows. Centralise the check. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Use the harmonic mean of D at unstructured faces in div(D * grad(u)) The face coefficient was a linear interpolation of the two cell values. For a flux the physically consistent face value is the resistances-in- series (distance-weighted harmonic) mean, which FiniteVolume already uses for coefficients of a gradient: it reproduces the exact two-cell flux across a material interface and gives zero flux into an impermeable cell, whereas the arithmetic mean lets half the conductive side leak through. On combined concatenation meshes this is the current-collector/electrode face, where sigma jumps by ~1e5. Linear weights are kept for the face gradient of the cross term; D must be strictly positive. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Treat faces orthogonal to within 1e-8 as orthogonal, not 1e-12 Centroid rounding on high-aspect-ratio cells (10 um thick, cm wide, as in a pouch cell) puts ~1e-11 into the face direction, so with a 1e-12 cut-off the cross term and boundary cross terms were assembled on every hex mesh with coefficients of 1e-11. That doubled the Jacobian stencil (57k -> 101k nonzeros on the pouch demo) and the IDAKLU time for no change in the answer. k is now zeroed below a dimensionless 1e-8 (an angle far below any real skew) and explicit zeros are dropped, so orthogonal faces never widen the stencil, also on mixed meshes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * Keep plotting out of the unstructured processed variables The visualisation grid (N_VIS/N_VIS_3D and the first/second/third_dim_pts it populated), mid-plane 3D slices, quiver sampling and slice positions only exist to feed QuickPlot, whose unstructured support lives in the plotting PR. Processed variables here now only interpolate at requested points; QuickPlot refuses unstructured variables with a clear message until that support lands, instead of failing deep inside its 2D branch. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
UnstructuredSubMesh, generators (UnstructuredMeshGenerator,UserSuppliedUnstructuredMesh,TaggedSubMeshGenerator), andMeshinterface coupling.Stack
Full pre-split branch preserved on #5397 /
backup/unstructured-finite-volume-full.Test plan
tests/unit/test_meshes/test_unstructured_submesh.py(41 tests)Also in this stack: #5691 deprecates
pybamm.Magnitude.