Skip to content

Bump Symbolics v7, SymbolicUtils v4, ModelingToolkit v11 compat #1025

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
ChrisRackauckas merged 14 commits into from
Jan 31, 2026
Merged

Bump Symbolics v7, SymbolicUtils v4, ModelingToolkit v11 compat #1025

ChrisRackauckas merged 14 commits into from
Jan 31, 2026

Conversation

@ChrisRackauckas-Claude
Copy link
Contributor

@ChrisRackauckas-Claude ChrisRackauckas-Claude commented Jan 30, 2026

Summary

  • Bump compat bounds: ModelingToolkit to v11, Symbolics to v7.8+, SymbolicUtils to v4.12+
  • Handle breaking changes from these version bumps:
    • build_expr was removed from Symbolics v7 — defined locally as a simple Expr constructor helper
    • PDESystem.defaults field renamed to initial_conditions in MTK v11 — updated accessor
    • Removed stale imports (defaults from MTK, build_expr from Symbolics)

Test plan

  • Package precompiles successfully with new deps
  • All core PDE tests pass (NNPDE, adaptive_loss, additional_loss, dgm, direct_function, neural_adapter)
  • CI full test suite (some pre-existing failures in SDE/PINO tests unrelated to this change)

🤖 Generated with Claude Code

@ChrisRackauckas-Claude
Copy link
Contributor Author

ChrisRackauckas-Claude commented Jan 30, 2026

Fix constraints.md segfault

The constraints.md tutorial was segfaulting during the docs build because Zygote attempted to differentiate through IntegralProblem/HCubatureJL (C-based cubature) inside the norm_loss_function. The IntegralsZygoteExt has no adjoint for getindex on IntegralSolution, causing Zygote to fall through to the C code and crash.

Changes:

  • docs/src/tutorials/constraints.md: Replaced IntegralProblem/HCubatureJL with a pure-Julia trapezoidal rule (200 points) that Zygote can differentiate through. Also fixed DomainSets imports (infimum/supremum → leftendpoint/rightendpoint).
  • src/training_strategies.jl: Changed solve(...)[1] to solve(...).u in QuadratureTraining — IntegralsZygoteExt provides an adjoint for .u but not for getindex.

Verified locally:

  • Ran the full constraints tutorial example with 600 BFGS iterations — completes without segfault
  • Loss converges as expected

claude and others added 4 commits January 30, 2026 08:30
@claude @ChrisRackauckas
Bump Symbolics v7, SymbolicUtils v4, ModelingToolkit v11 compat
fba058d 
Breaking changes handled:
- `build_expr` removed from Symbolics v7: defined locally in symbolic_utilities.jl
- `defaults` field removed from PDESystem in MTK v11: replaced with `initial_conditions`
- `defaults` import removed from ModelingToolkit (no longer exported)
- `build_expr` import removed from Symbolics (no longer exported)

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
@claude @ChrisRackauckas
Remove MethodOfLines test dependency
1acd90f 
Replace runtime MOL computation in the Burger's equation DGM test with
precomputed reference data. The reference solution was generated using
MOLFiniteDifference([x => 0.01], t, saveat=0.01) with Tsit5() and
subsampled every 10th point for compact storage.

Also remove unused MethodOfLines imports from Poisson and Black-Scholes
tests which already compare against analytical solutions.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
@claude @ChrisRackauckas
Fix iscall check for SymbolicUtils v4 and apply Runic formatting
5087cbe 
In SymbolicUtils v4, BasicSymbolic is a sum type that includes both Sym
(plain symbols) and Term (callable). The old `isa BasicSymbolic` check
matched Sym which has no arguments, causing ArgumentError. Use iscall()
to correctly filter for callable terms only.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Fix constraints.md segfault: replace HCubatureJL with trapezoidal rule
2cd6566 
The constraints tutorial segfaulted because Zygote tried to differentiate
through IntegralProblem/HCubatureJL (C-based cubature), which lacks Zygote
adjoint rules for getindex on IntegralSolution.

- Replace IntegralProblem/HCubatureJL in norm_loss_function with a pure-Julia
  trapezoidal rule that Zygote can differentiate through
- Fix QuadratureTraining to use sol.u instead of sol[1] (IntegralsZygoteExt
  provides an adjoint for .u but not getindex)
- Fix DomainSets imports: use IntervalSets leftendpoint/rightendpoint

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
claude and others added 7 commits January 30, 2026 09:14
@claude @ChrisRackauckas
Fix defaults→initial_conditions in tests, handle scalar loss in Quadr…
d0768e4 
…atureTraining, fix symbolic tuple in IntegroDiff

- Replace deprecated `defaults` keyword with `initial_conditions` in
  PDESystem constructors (MTK v11 breaking change) in test files
- Handle scalar return from loss function in QuadratureTraining integrand
  to fix MethodError with view() on Float64 (Wave Equation test)
- Handle symbolic tuple from toexpr in SymbolicUtils v4 for multi-variable
  integration domains (IntegroDiff multi-input tests)
- Fix Runic formatting (1e5 → 1.0e5)

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Fix Runic formatting in symbolic_utilities.jl
f7de495 
Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Fix three CI crash errors from SymbolicUtils v4 / MTK v11
8f62625 
- Fix TypeError in NonAdaptiveLoss: wrap initial_conditions values with
  ModelingToolkit.value() since they return BasicSymbolic{SymReal} not Float64
- Fix DimensionMismatch in QuadratureTraining: return [abs2(l)] instead of
  scalar for BatchIntegralFunction compatibility
- Fix KeyError in IntegroDiff: compare against `tuple` function object (===)
  not `:tuple` Symbol, since toexpr produces Expr(:call, tuple, :x, :y)

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Fix QA ExplicitImports and BPINN KeyError from SymbolicUtils v4
2b2d146 
- Use Symbolics.value instead of ModelingToolkit.value in discretize.jl
  to satisfy ExplicitImports ownership check
- Rewrite get_symbols in PDE_BPINN.jl to search equation tree for
  callable depvar terms directly, since SymbolicUtils v4 get_variables
  decomposes u(x,t) into bare u, x, t instead of returning u(x,t)

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Fix parse_equation zero detection and QA issues for SymbolicUtils v4
4fc37d5 
- Use Symbolics._iszero instead of isequal(..., 0) in parse_equation to
  properly detect symbolic zeros from expand_derivatives. In SymbolicUtils v4,
  BasicSymbolic{SymReal}(0) is not isequal to Julia's Int 0, causing
  derivative BCs like Dt(u(x,0)) to be replaced with literal 0 instead of
  the numeric derivative expression. This was the root cause of the Wave
  Equation MethodError crash.
- Revert QuadratureTraining scalar loss workaround (no longer needed)
- Fix Symbolics.unwrap → SymbolicUtils.unwrap in PDE_BPINN.jl (ExplicitImports)
- Fix Runic formatting: if → return if in PDE_BPINN.jl

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
@claude @ChrisRackauckas
Fix Differential.order handling and QA for SymbolicUtils v4
8cd4020 
In SymbolicUtils v4, expand_derivatives produces Differential(x, 2)
(a single higher-order differential with .order field) instead of
nested Differential(x)(Differential(x)(...)). The _transform_expression
loop was only incrementing order by 1 per Differential object, causing
second-order PDEs (Poisson, Wave, etc.) to be treated as first-order.
This resulted in completely incorrect loss gradients and training failure.

Fix: read the .order field from each Differential and push the derivative
variable that many times.

Also fix QA: use SymbolicUtils._iszero instead of Symbolics._iszero
(ExplicitImports ownership check).

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
@claude @ChrisRackauckas
Fix IntegroDiff infinite bounds for SymbolicUtils v4
e292d29 
Two issues:
1. v_semiinf checked `a isa Num` but SymbolicUtils v4 unwraps to
   BasicSymbolic, not Num. Added BasicSymbolic to the check.
2. Boolean mask multiplication with Inf bounds (e.g., false * Inf = NaN
   via IEEE754) caused NaN integration bounds. Replaced mask arithmetic
   with conditional logic.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@ChrisRackauckas-Claude
Copy link
Contributor Author

ChrisRackauckas-Claude commented Jan 31, 2026

Fix IntegroDiff infinite bounds (e292d29)

The IntegroDiff tests were failing because of two issues with SymbolicUtils v4:

  1. v_semiinf type check: The function checked a isa Num to detect symbolic bounds, but SymbolicUtils v4's unwrap produces BasicSymbolic not Num. Added BasicSymbolic to the check.

  2. NaN from boolean mask multiplication with Inf: The original code used _none .* ub where ub contained Inf. Since false * Inf = NaN (IEEE754: 0.0 * Inf), this produced NaN integration bounds. Replaced the boolean mask arithmetic with explicit conditional logic.

Test result: IntegroDiff Example 7 (I(u(x)) ~ 1/x with Integral(x in ClosedInterval(x, Inf))) now produces max abs diff of 0.245 (well within the atol=0.5 tolerance).

claude and others added 3 commits January 31, 2026 09:30
@claude @ChrisRackauckas
Retrigger CI for flaky failures
1b641e6 
Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Fix QA unwrap ownership and increase Wave Equation training iters
4dd58b2 
- Use SymbolicUtils.unwrap instead of Symbolics.unwrap in
  transform_inf_integral.jl (QA ExplicitImports check)
- Increase Wave Equation test training iterations from 500 to 1000
  for both Adam warmup and BFGS to improve convergence robustness

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@claude @ChrisRackauckas
Reduce 2D Poisson QuadratureTraining learning rate and add BFGS
130cb8b 
Adam(0.1) with 500 iters can diverge on some seeds. Changed to
Adam(0.01) with 1000 iters warmup followed by BFGS with BackTracking
line search for robust convergence.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
@ChrisRackauckas ChrisRackauckas merged commit 75c8308 into SciML:master Jan 31, 2026
32 of 34 checks passed
@ChrisRackauckas-Claude ChrisRackauckas-Claude mentioned this pull request Feb 1, 2026
Merged
3 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@ChrisRackauckas-Claude @ChrisRackauckas @claude