Tupek/differentiable block solver by tupek2 · Pull Request #1518 · llnl/smith

tupek2 · 2026-01-08T22:20:58Z

A few ongoing fixes and improvements

Add differentiable block solver. This block solver uses block linear solvers, but otherwise calls our existing nonlinear solver (e.g., Newton) over the entirely block system. This allows monolithic nonlinear solves.
Remove duplication in test_solid_mechanics_state_advancer
Initial version of test_thermo_mechanics.cpp a coupled test that uses the new FieldStore as a prototyp
Update to latest gretl, which fixes an asan issue and also improves performance in some cases
Change the default NewtonLinesearch to use 10 cutbacks. Small interface changes around this to ensure Newton still defaults back to 0 cutbacks.
Taylor series style for residual update in warm-start to better handle physics which start out of equilibrium

…rder issues.

…earch to have no-zero line searches.

…rations when appropriate.

…oint algorithm.

src/smith/differentiable_numerics/tests/test_field_state.cpp

src/smith/differentiable_numerics/tests/test_thermo_mechanics.cpp

src/smith/differentiable_numerics/differentiable_solver.cpp

src/smith/differentiable_numerics/nonlinear_solve.hpp

src/smith/differentiable_numerics/time_integration_rule.hpp

src/smith/physics/common.hpp

src/smith/differentiable_numerics/time_integration_rule.hpp

src/smith/differentiable_numerics/nonlinear_solve.cpp

Co-authored-by: Chris White <white238@llnl.gov>

btalamini

I read through all of this, and it looks good to me! Sorry to make you wait for a no-feedback response.

btalamini · 2026-03-09T18:10:39Z

src/smith/differentiable_numerics/differentiable_solver.cpp

+  std::vector<DifferentiableBlockSolver::FieldPtr> u_duals(static_cast<size_t>(num_rows));
+  for (int row_i = 0; row_i < num_rows; ++row_i) {
+    u_duals[static_cast<size_t>(row_i)] = std::make_shared<DifferentiableBlockSolver::FieldT>(
+        u_bars[static_cast<size_t>(row_i)]->space(), "u_dual_" + std::to_string(row_i));
+  }


In the future, we could make utilities for (block vector -> subvector) and the inverse and stick them in common place.

btalamini · 2026-03-09T18:17:59Z

Pinging @white238 , @ebchin, @lihanyu97 . Any outstanding concerns?

tupek2 added 5 commits December 30, 2025 16:11

Adding a block nonlinear solver.

1862206

Fix style, adding differentiable nonlinear block solver.

8299a98

Fix style.

a94e2fb

Merge branch 'develop' into tupek/differentiable_block_solver

906499b

Remove boundary condition option that causes some template priority o…

6aa3047

…rder issues.

tupek2 added the WIP Work in progress label Jan 8, 2026

tupek2 added 11 commits January 12, 2026 14:10

Still working on block nonlinear solver.

218b75e

SOme minor updates.

b3e8349

Trying to simplify adding new physics even more.

c2ced3f

Working through new interface ideas, change newton linesearch default.

8b30cd8

Some minor adjustments to interface, change default for NewtonLinears…

39819b3

…earch to have no-zero line searches.

Get an apparently working version of automating multiphysics solves.

c2684e8

Fix style.

8e093ae

Merge branch 'develop' into tupek/differentiable_block_solver

3257700

Small adjustment to equation solver defaults to ignore linesearch ite…

9631ff4

…rations when appropriate.

Style.

14fa86e

Address an assert and remove or improve some comments.

e4a2c73

tupek2 requested a review from btalamini January 27, 2026 19:18

tupek2 added 4 commits January 27, 2026 12:32

Working through thermal mechanics details.

2dd9dfd

Remove some duplication.

59af3fe

Fix style.

956d6df

Merge branch 'develop' into tupek/differentiable_block_solver

348ab9f

tupek2 force-pushed the tupek/differentiable_block_solver branch from 5b88e69 to 348ab9f Compare February 2, 2026 15:55

tupek2 and others added 5 commits February 4, 2026 13:29

Merge branch 'develop' into tupek/differentiable_block_solver

553e03b

Spelling.

066976e

Use gretl with memory fix.

c8c40a7

Merge branch 'develop' into tupek/differentiable_block_solver

e376248

Update gretl interfaece, try to fix memory, use more efficient checkp…

c69acfb

…oint algorithm.

tupek2 requested a review from ebchin February 10, 2026 05:22

Point to main gretl after updates.

0acd385