Uh oh!
There was an error while loading. Please reload this page.
Consistent Jacobian tangent (unwrap-before-differentiate fix) — opt-in, default-off - #258
Merged
Merged
Conversation
Gated behind consistent_jacobian (default False = bit-identical Picard). - symbolic_keep_constants unwrap mode (constants[]-safe, drift-guarded) - consistent_jacobian: False | True | 'continuation' (alpha-blend via constants[]) - model-owned flux_jacobian hook; Nitsche bd_F1 wired Validated: level_1 tier_a 225/225, units 64/64, constant-visc bit-identical. Underworld development team with AI support from Claude Code
Creating the continuation alpha UWexpression in every solver __init__ bumped the global unique-name counter, shifting JIT cache keys enough to flip two known-flaky VEP variable-dt yield-lock tests in full-suite runs. Construct alpha lazily (continuation mode only) so the default Picard path creates no extra expression. Verified: level_2 tier_a forked failure set now IDENTICAL to pristine origin/development (4 pre-existing failures). Underworld development team with AI support from Claude Code
…n-unwrap-to-constants # Conflicts: # src/underworld3/cython/petsc_generic_snes_solvers.pyx
…owned) ViscoElasticPlasticFlowModel.flux_jacobian returns the flux with Min->harmonic (1/(1/eta_ve+1/eta_pl)) and Max smoothed, for the Jacobian source ONLY (exact Min residual preserved). Pure symbolic substitution on the live flux — no state mutation (fixes the earlier err-77 from a yield_mode toggle hack). Runs clean. Guard _newton_flux on consistent_jacobian so the default (Picard) path never evaluates flux_jacobian — keeps default assembly allocation-free / bit-identical (test_1010 6/6). NOTE: smooth-Jacobian + hard-Min residual is an inconsistent tangent and does NOT improve convergence on the BDF-2 VEP loading test (8/15 vs Picard 3/15); only full harmonic (consistent smooth residual+Jacobian) converges (0/15). The hook is correct, clean, and available; VEP convergence remains a separate issue. Underworld development team with AI support from Claude Code
The smooth-Jacobian-with-Min-residual tangent is inconsistent (consistent with the harmonic problem, not Min) and converges worse than Picard on hard-yield VEP — so the VEP-specific harmonic flux_jacobian override is deferred to the yield-law / delta-homotopy follow-up PR. The generic Constitutive_Model. flux_jacobian hook (default None) + the _newton_flux guard remain. Add docs/developer/design/jacobian-consistent-tangent.md: the bug, the opt-in/default-off fix, non-regression evidence, the tangent hierarchy, and the delta-homotopy successor work. Underworld development team with AI support from Claude Code
Contributor
There was a problem hiding this comment.
Pull request overview
This PR introduces an opt-in “consistent Jacobian tangent” path for nonlinear/viscoplastic SNES solves by unwrapping UWexpression atoms before symbolic differentiation (while still preserving truly-constant atoms for the PETSc constants[] mechanism). It also adds a continuation mode to blend Picard→Newton without triggering JIT recompiles, plus a constitutive-model hook to provide a custom Jacobian-only flux.
Changes:
- Add
symbolic_keep_constantsunwrapping mode to expand non-constantUWexpressionatoms while preserving truly-constant atoms as the same symbols forconstants[]. - Add
solver.consistent_jacobianwith modesFalse(default Picard),True(Newton), and"continuation"(alpha-blended Picard→Newton). - Add
Constitutive_Model.flux_jacobianhook (defaultNone) so models can supply a Jacobian-only surrogate flux.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 9 comments.
| File | Description |
|---|---|
src/underworld3/function/expressions.py | Adds symbolic_keep_constants unwrapping mode used to expose coefficient dependence during Jacobian differentiation without breaking constants[]. |
src/underworld3/cython/petsc_generic_snes_solvers.pyx | Implements Jacobian source selection (Picard/Newton/continuation), uses unwrap-before-differentiate for Jacobian assembly, and adds continuation solve control flow. |
src/underworld3/constitutive_models.py | Introduces flux_jacobian optional hook for Jacobian-only tangent substitution and documents intended usage. |
docs/developer/design/jacobian-consistent-tangent.md | New design note explaining the unwrap-before-differentiate bug, the opt-in fix, and the tangent/continuation rationale. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+45
to
+47
| if isinstance(expr, sympy.NDimArray): | ||
| return sympy.Array([f(e) for e in expr], expr.shape) | ||
| return f(expr) # scalar expression |
| This is a no-op for constant-viscosity problems (eta has no grad-v | ||
| dependence), so those Jacobians stay bit-identical. | ||
| See ``docs/developer/design/jacobian-unwrap-constants-bug.md``. |
| # the model's own smooth law (constitutive_model.flux_jacobian) when it | ||
| # provides one; otherwise the exact unwrapped flux. | ||
| # | ||
| # See docs/developer/design/jacobian-unwrap-constants-bug.md. |
| # The residual fns above (self._u_F0/_u_F1/_p_F0) are left untouched — | ||
| # getext() unwraps those itself. For constant-viscosity problems this | ||
| # is a no-op (eta has no grad-v dependence) so the Jacobian is | ||
| # bit-identical. See docs/developer/design/jacobian-unwrap-constants-bug.md |
Comment on lines
+1069
to
+1071
| # Restore a clean Picard tangent for any subsequent solve (next step). | ||
| self._set_newton_alpha(0.0) | ||
Comment on lines
+1024
to
1029
| if self.consistent_jacobian == "continuation": | ||
| self._continuation_solve(gvec, verbose=verbose) | ||
| else: | ||
| self.snes.solve(None, gvec) | ||
| if divergence_retries <= 0: | ||
| return |
Comment on lines
+580
to
+582
| Returns ``None`` by default, meaning the solver differentiates the | ||
| exact :attr:`flux` (the Newton fix unwraps it first; a generic Min/Max | ||
| kink-smoothing fallback then rounds any remaining yield kink). |
Comment on lines
+1718
to
+1719
| # (ramp the softmin softness δ→0), not a smooth tangent. See the design doc | ||
| # docs/developer/design/jacobian-unwrap-constants-bug.md. The generic |
Comment on lines
+2486
to
+2489
| # JACOBIAN: unwrap (keep constants) + smooth Min/Max kinks so the | ||
| # derivative sees the field-dependence of any nonlinear coefficient | ||
| # (full Newton) while the residual keeps the exact form. No-op for | ||
| # constant coefficients -> bit-identical. See _jacobian_source. |
…n-unwrap-to-constants
bknight1
commented
Jun 25, 2026
Member
@lmoresi I've had a look at this PR and it's vastly improved the convergence of the VP models I've been working on, both when using continuous and True. The only issue is compilation time with multimaterials, which can be >5 mins, with 'continuous' taking longer than True. Another issue is the compilation appears to occur twice after updating the constitutive model (one to check the hash is different then a rebuild the jacobians). I have made some modifications to fix these issues, happy to discuss committing these changes. |
…n-unwrap-to-constants # Conflicts: # src/underworld3/cython/petsc_generic_snes_solvers.pyx
lmoresi
commented
Jul 2, 2026
MemberAuthor
Reviewed + rebased onto current development (resolved the petsc_generic_snes_solvers.pyx conflict — kept both this PR's Jacobian methods and development's set_custom_mg; builds clean, Stokes 6/6 passes). Safe to merge: the consistent tangent is opt-in and default-off ( Underworld development team with AI support from Claude Code |
Uh oh!
There was an error while loading. Please reload this page.
lmoresi added a commit
that referenced
this pull request
Jul 6, 2026
…og, value-first call-site sweep (WE-01..03,05,06,08,09,10) (#338) * docs(WE-01): adopt the one-governing-doc-per-topic authority map Repoint CLAUDE.md's Data Access 'Authoritative Reference' from the stale UW3_Style_and_Patterns_Guide.md to subsystems/data-access.md (the guide it crowned teaches patterns the code deprecates at runtime — DOC-04), and record the Style Charter §10 authority table in docs/developer/index.md as the master authority index. The Charter is added to the Getting Started toctree (removes a baseline 'not included in any toctree' warning). Finding: DOC-04 (docs/reviews/2026-07/DOCS-STANDARDS-COHERENCE.md). Underworld development team with AI support from Claude Code * docs(WE-02): de-drift the Style Guide's four stale normative sections Rewrites the sections DOC-01 verified as contradicting the settled standards: - Docstring format: the 'Markdown Docstrings for pdoc/pdoc3' section is replaced by the NumPy/Sphinx RST standard (worked example with :math: and Parameters/Returns/Examples/Notes; conversion tracked in docs/plans/docstring-conversion-plan.md), per Style Charter section 6. - Doc file format: Quarto .qmd prescription (zero .qmd files exist in the repo) replaced by MyST .md/Sphinx guidance matching CLAUDE.md; migration table row updated. - Data access examples: 'Preferred' coordinate examples now use the real, runnable API — mesh.X.coords (read), mesh.deform() (coordinate changes), and the swarm.coords getter/setter for particle positions. The previous 'Preferred' example swarm.data += displacement raises AttributeError (getter-only property — SWARM-13 evidence); mesh.data warns at runtime. The private-attribute migration advice (swarm._particle_coordinates, mesh._deform_mesh presented as the NEW pattern) is deleted. - Front matter: the 21-line Quarto YAML header is replaced by a minimal MyST title block, and the guide now states that the UW3 Style Charter is the normative contract and wins on conflict. All replacement examples verified against current source: Swarm.coords setter (swarm.py), Mesh.deform (discretisation_mesh.py:3133), uw.synchronised_array_update / NDArray_With_Callback.delay_callbacks_global. Findings: DOC-01, SWARM-13 (style-guide part). Underworld development team with AI support from Claude Code * docs(WE-03): regenerate the docstring review queue; add the sweep to the release checklist The queue (last generated 2026-01-13, cdf5bb2) misrepresented the codebase both ways: it flagged now-complete items (solve, SNES_Scalar) as missing and contained zero entries for the June 2026 API (DOC-02). Regenerated over src/underworld3/**/*.py + **/*.pyx at the current tip. Two bugs in scripts/docstring_sweep.py's regex-based Cython parser made the regenerated queue lie about .pyx docstrings and are fixed as part of making the regeneration meaningful: - the indent group '(\s*)' with re.MULTILINE consumed preceding blank lines, shifting the computed definition line so the docstring search started ON the def/class line and always missed; - the docstring search started at the definition line rather than after the (possibly multi-line) signature, so long signatures hid their docstrings; - raw-string docstrings (r""", the norm in the solver .pyx) were not recognised. DOC-02 cross-validation on the regenerated queue now passes: solve / SNES_Scalar in the solver pyx are no longer flagged 'none'; the queue contains the June API (add_nitsche_bc, add_rotated_freeslip_bc, boundary_flux, set_custom_fmg, consistent_jacobian: 13 mentions) and flags the DOC-05 targets (Swarm.advection x2, read_timestep, write_proxy) as undocumented. Also adds the sweep to the quarterly release checklist (guides/release-process.md) so the queue cannot go stale unnoticed again. Findings: DOC-02 (docs/reviews/2026-07/DOCS-STANDARDS-COHERENCE.md). Underworld development team with AI support from Claude Code * docs(WE-05): backfill the changelog for May - early July 2026; add the changelog sweep to the release checklist The changelog (the quarterly CIG/stakeholder record) ended in April 2026 while ~117 first-parent commits landed May through early July (DOC-03). Backfilled at the existing conceptual granularity — 14 grouped entries, grouped by subsystem rather than by PR, matching the established format (### Title (Month Year), bold lead sentence, hyphen bullets, inline PR references): - New '2026 Q3 (July - September)' section: the July 2026 quality campaign (#309-#313, #317, #322-#326, #329, #334 as grouped entries), rotated strong free-slip / boundary traction / dynamic topography (#293, #294, #298, #306), generalized geometric multigrid via custom prolongation (#290, #297), consistent Jacobian tangent (#258), swarm correctness (#216, #313, #323, #329), numpy 2 support (#301, #305). - Extended '2026 Q2' section with the May-June entries: mesh adaptation movers (#190, #209, #213, #228, #259, #264, #266), moving-mesh field transfer / deform() (#246, #249, #251), semi-Lagrangian accuracy controls (#164, #183, #185-#189, #208, #220), snapshot/checkpoint toolkit (#146, #195, #196, #198), Stokes_Constrained (#224, #229, #240, #265), local-h Nitsche + boundary-slip surfaces (#225, #241, #275), units interoperability (#277, #278, #283, #284), memory/evaluation/solver infrastructure (#161, #177-#179, #181, #182, #222, #237, #250, ...). Every entry is backed by a merged commit on development (verified against git log --first-parent aed517f..3184a40). Also adds a quarterly-changelog sweep step beside the docstring sweep in the release checklist (guides/release-process.md) per DOC-03's proposed fix. Findings: DOC-03 (docs/reviews/2026-07/DOCS-STANDARDS-COHERENCE.md). Underworld development team with AI support from Claude Code * docs(WE-06): status headers on the unmarked design docs (per-doc git verification) Adds one-to-three-line Status markers to the 13 design docs that lacked one, following the directory's existing conventions (**Status**: line under the title; status: key inside existing YAML frontmatter for the three frontmatter-only docs), and corrects the stale 'Design Phase' marker on MATHEMATICAL_MIXIN_DESIGN.md (the mixin ships in utilities/mathematical_mixin.py). Every stamp was verified against git history (git log --follow dates) and the current source tree before writing: - Implemented: jacobian-consistent-tangent (PR #258, c63cd70), fmg-checkpoint-hierarchy (3cd73cd), petsc-dmplex-checkpoint-reload-plan (PR #146, write_timestep(petsc_reload=True) in tree), fault-refinement-simplification (smooth_mesh_interior / metric_density_from_gradient / fault_comb_metric all in tree), MATHEMATICAL_MIXIN_DESIGN. - Current reference/contract: mesh-adaptation-formulation, ND_UNITS_BOUNDARY_CONTRACT (PR #278, e0ece9a). - Investigation records (preserved via PR #245, 34a9dd4; production geometric-MG is custom prolongation, PR #290): snesfas-feasibility, snesfas-vanka-feasibility-study. - Design notes / prototypes with honest gaps: in_memory_checkpoint_design (not implemented, per its own trailing Status section), submesh-solver-architecture (extract_region/extract_surface exist; coarsened_companion does not). - Historical: ARCHITECTURE_ANALYSIS (persistence.py layout superseded), COORDINATE_MIGRATION_GUIDE (transition shipped), WHY_UNITS_NOT_DIMENSIONALITY (decision record). The audit's ~16 estimate over-counted: re-derived at this tip, 13 docs were unmarked plus one marked-but-stale (DOC-07). Findings: DOC-07 (docs/reviews/2026-07/DOCS-STANDARDS-COHERENCE.md). Underworld development team with AI support from Claude Code * docs(WE-08): convert units.py public docstrings Google -> NumPy style Docstring-only conversion of the 18 public module-level functions that carried Google-style Args:/Returns:/Raises:/Examples: labels (check_units_consistency, get_dimensionality, get_units, non_dimensionalise, show_nondimensional_form, simplify_units, create_quantity, convert_units, to_base_units, to_reduced_units, to_compact, get_scaling_coefficients, set_scaling_coefficients, validate_expression_units, assert_dimensionality, validate_coordinates_dimensionality, enforce_units_consistency, require_units_if_active, convert_angle_to_degrees) to the NumPy/Sphinx standard (Style Charter section 6). dimensionalise was already NumPy style; one-line docstrings and private helpers are untouched. No code, signature, or behaviour changes (verified: every diff hunk is inside a docstring; ast.parse clean). Finding: API-12 (docs/reviews/2026-07/API-CONSISTENCY-REVIEW.md). Underworld development team with AI support from Claude Code * docs(WE-09): sweep call sites of the newer BC methods to value-first (conds, boundary, ...) order Wave C (#334) made the ORIGINAL value-first order canonical for add_nitsche_bc / add_rotated_freeslip_bc / add_constraint_bc (maintainer decisions D2/D3; Style Charter section 6) with deprecation shims for the legacy boundary-first and g= spellings. This sweep updates every call site of those THREE methods to the canonical order so nothing in the repository exercises the shims — 74 sites total: - tests/: 63 call sites across 12 files (test_1017, test_1018, test_1060, test_1061, test_1062, test_1064, test_1065 x2 serial; parallel test_1017, test_1062, test_1063, test_1064). tests/test_0641_wave_c_api_shims.py is deliberately untouched — its legacy-order calls ARE the deprecation contract. - docs/: 7 sites (curved-boundary-conditions.md x4, CONSTRAINED_FREESLIP_MULTIPLIER.md call + signature line, examples/submesh_investigation/test_region_ds_nitsche.py). - .claude/skills/: 3 sites (adapt-on-top-faults x2, free-surface-convection x1). - CLAUDE.md: 1 signature reference (free-slip BC preference section). The ~1,370 legacy-trio (add_dirichlet_bc/add_natural_bc/add_essential_bc) sites already conform and are untouched per the D2 decision. The audit review documents under docs/reviews/2026-07/ record the pre-decision state as evidence and are not swept. Discovered while verifying the swept tests run warning-free: the Wave C zero-datum guard in add_rotated_freeslip_bc rejects FLOAT zero (sympy.sympify(0.0) != 0 is structurally True), so the canonical add_rotated_freeslip_bc(0.0, boundary) raises NotImplementedError while conds=0 works. Filed as issue #336 with a TODO(BUG) marker at the guard (comment-only src touch); the swept call sites use the working integer form add_rotated_freeslip_bc(0, boundary). No fix applied here (Charter section 9 scope discipline). Findings: API-01/API-02 sweep (WE-09, REMEDIATION-WORKLIST.md). Underworld development team with AI support from Claude Code
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 freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Consistent Jacobian tangent for nonlinear (viscoplastic) solves — opt-in, default-off
The bug
SNES Jacobian assembly differentiated the residual flux
F1while the effectiveviscosity was still a wrapped
UWexpressionatom, so∂η/∂(grad v)was silently droppedfrom every Jacobian. UW3 viscoplastic Stokes was therefore running an accidental Picard /
defect-correction tangent, not full Newton — the unwrap happened after the derivative
instead of before it. Constant-viscosity problems were unaffected, which is why it stayed
hidden behind the "≈20 Picard iterations is intrinsic" folklore.
The fix (default-off, bit-identical by default)
symbolic_keep_constantsunwrap mode — expandsUWexpressionatoms down to (but notincluding) truly-constant atoms (η₀, τ_y stay symbolic for the
constants[]mechanism).The keep-symbolic predicate is the same
_is_truly_constantused byconstants[]extraction, so they cannot drift (drift-guard test).
solver.consistent_jacobian:False(default, frozen/Picard, bit-identical) /True(full Newton) /"continuation"(Picard→Newton via anα-blend routed throughconstants[], so switching costs no JIT recompile andα=0is bit-identical).Constitutive_Model.flux_jacobianhook (defaultNone) for a model to supply a smoothtangent law;
_newton_fluxis guarded so the default path never evaluates it.The residual never goes through the new path, so a converged solution always satisfies the
exact constitutive law.
Non-regression evidence
test_1010, SolCxtest_1015,test_0610, asymmetric-Jacobian guard, units (64) pass;level_1 tier_a225/225.--forked)level_2 tier_afailure set is identical to pristinedevelopment— the pre-existing reds (test_1012gmsh crash + 3test_1052VEP) are notintroduced here.
Scope / why a smooth Jacobian is not the VEP fix
For a hard-
Minresidual, a smooth Jacobian is the consistent tangent of a different(harmonic) problem — it diverges more than Picard. The robust route for hard-yield VEP is
problem-space homotopy (ramp the softmin softness δ→0; δ=0 is identically
Min), which isa separate follow-up PR that reuses this PR's
constants[]-ramp machinery. Seedocs/developer/design/jacobian-consistent-tangent.mdfor the full tangent hierarchy and thesuccessor plan.
Underworld development team with AI support from Claude Code