Tupek/system solver by tupek2 · Pull Request #1556 · llnl/smith

tupek2 · 2026-03-16T17:11:20Z

Consolidated Time Integrators: Removed specific integrators (e.g., solid_mechanics_time_integrator.hpp) and replaced them with a generic MultiphysicsTimeIntegrator capable of handling coupled systems.
Generic System Templates: Replaced specific dynamic/static struct definitions (solid_dynamics_system.hpp, solid_statics_with_internal_vars_system.hpp) with configurable templates (solid_mechanics_system.hpp, solid_mechanics_with_internal_vars_system.hpp) that accept custom time integration rules.
Solver Relaxation: Introduced a relaxation_factor into the CoupledSystemSolver to support under-relaxation for staggered multi-physics solves (e.g., x n e w = ω x s o l v e d + ( 1 − ω ) x o l d ).
Cycle-Zero Acceleration Solves: Built explicit support into the systems and time integrators for "cycle zero" solves to establish initial accelerations from initial displacements and velocities.
System Solver Robustness: Improved SystemSolver convergence logic (using global L2 norm of the un-staggered residual) and replaced brittle string-matching with O ( 1 ) block index mappings.
Deprecation of Old Physics Forms: Deleted legacy solid_weak_form.hpp and heat_transfer_weak_form.hpp from the smith/physics directory, routing users entirely through the new differentiable numerics block solvers.
Extended Field Store: Added dependent fields and state indexing to the FieldStore to handle states and time derivatives ( VAL , DOT , DDOT ) seamlessly for multiple variables.
Test Suite Overhaul: Re-wrote tests (test_thermo_mechanics.cpp, test_solid_dynamics.cpp, test_thermal_static.cpp) to utilize the updated block solvers and newly abstracted simulation builders.

…re_interface

… integrator.

… a state advancer.

…everse pass.

…physics interface.

…nternal vars

…boundary conditions.

… docs and style.

…l the mechanics weak forms.

…ltiblock situations.

tlroy · 2026-04-08T15:28:29Z

Does LinearSolverOptions.sub_block_linear_solver_options allow for nested block solvers? We should test that.

tlroy · 2026-04-08T15:30:43Z

LinearSolverOptions.sub_block_linear_solver_options requires a linear solver even if one wants to apply the preconditioner by itself. We should set up a LinearSolver:None

ebchin

Very elegant, very flexible; a lot of good stuff here.

I know the focus here is coupled PDEs, but it seems like a lot of this can be used for solving constraints too. Excited to try it out for contact!

ebchin · 2026-04-09T02:32:47Z

@@ -1,3 +1,4 @@
+


Remove extra line

ebchin · 2026-04-09T22:53:53Z

+   * @brief Enum for different types of time derivatives.
+   */
+  enum class TimeDerivative
+  {


Does it make sense to add an integral time derivative type? I'm thinking of places where the independent field might already be a time derivative.

ebchin · 2026-04-09T23:06:05Z

+  }
+
+  return {new_states, reactions};
+}


Maybe I haven't gotten there yet, but is there something that works with multi-stage time integration schemes like RK4?

ebchin · 2026-04-10T02:27:55Z

+  /// @param exact_staggered_steps If true, always perform exactly @p max_staggered_iterations
+  ///        sweeps with no early-exit convergence check.  Useful when a fixed number of
+  ///        partitioned-stagger steps is required regardless of residual level.
+  CoupledSystemSolver(int max_staggered_iterations, bool exact_staggered_steps = false);


It seems like you can set this up to do Uzawa iterations to solve for constraints?

ebchin · 2026-04-10T04:29:49Z

-        tangent_weights[field_index_to_diff] = 0.0;
+        if (field_index_to_diff != invalid_block_index) {
+          tangent_weights[field_index_to_diff] = 1.0;
+          auto jac_ij = residual_evals[row_i]->jacobian(time_info, shape_disp_ptr.get(),


Should we refactor WeakForm::jacobian() to only take the derivative w.r.t. a single field? If you had a thermomechanical system or a constrained system, the size of the blocks would be different anyway.

tupek2 and others added 30 commits February 9, 2026 14:22

Working out the new interface.

39cf5ab

Merge branch 'tupek/differentiable_block_solver' into tupek/field_sto…

f8458a6

…re_interface

Some progress on the new design.

ea0ca71

update new gretl.

0a44b2b

Merge branch 'tupek/differentiable_block_solver' into tupek/field_sto…

1b78255

…re_interface

Refactoring thermo mechanics example to use generic multiphysics time…

a03c722

… integrator.

Cleanup interface by having a build method to construct a system with…

cf22c89

… a state advancer.

--aned

8e413ee

Working through new interface issues. there is still a crash in the r…

1d8a7a5

…everse pass.

Address interface changes in contact examples.

1e5219f

Still working on debugging reverse pass.

a977f34

Update new continutation solvers.

121ff6a

Progress on debugging issue.

36eeebb

Fix bug in sending multiple params to residuals in some cases.

d5544ff

some docs.

edfe764

Line changes.

9cc8d29

Use current continutation solvers.

a04b7cb

Fix some style.

f44f16b

Fix docs and style.

de69f85

Add addition runtime checks for consistent use of spaces and fields.

fc8abbe

Refactor existing solid mechanics state advancer to use the new multi…

2441fcb

…physics interface.

Refactoring to improve testing of backpropagation.

04370ae

Clearer names for dims and such.

ec0ab77

Fix style.

211bc93

style change to reduce lines.

fe0c79e

refactor solid state advancer to reuse the multiphysics advancer.

1148e7f

A working state, but there is a lot to cleanup and review.

3ca1727

Working tests, a bit cleaner interface.

ae37692

Simplifying tests and implementations.

3e0bd29

Some rename clean remove heat transfer weak form that will be unused.

2206310

mrtupek2 and others added 23 commits March 26, 2026 13:40

fix compile issues in cycle-zero weak form for solid mechanics with i…

401fabd

…nternal vars

Style and trying to get things more consistent.

64ba138

Merge branch 'develop' into tupek/system_solver_reactions

97c2fc7

Merge branch 'develop' into tupek/system_solver_reactions

45074bd

Unique pointer, and slightly better handling of zero cycle solve.

3d9c014

Merge branch 'develop' into tupek/system_solver

a064fa7

Merge branch 'tupek/system_solver_reactions' into tupek/system_solver

9a4a61e

Trying to simplify cycle 0 block solver strategy and get the correct …

f25f681

…boundary conditions.

style.

d5697e7

Remove some thermoelastic material models and tests that came in, fix…

66b05ab

… docs and style.

Remove unneeded paraview shim.

fa4a959

Remove random png.

cdba59e

Update to new paraview usage.

881894c

Do not do initial acceleration solve unless needed. Add density to al…

35434d7

…l the mechanics weak forms.

Put back accidentally deleted materials.

5fb4ba4

Some test and docmentation cleanup.

cd39875

Remove unused time integration rule.

769545a

Fix naming and cmakelist issue.

dec64a9

Add header to cmake.

0da8e6e

Merge branch 'develop' into tupek/system_solver

06f7966

Merge branch 'develop' into tupek/system_solver

c4a7e0c

More nullptr safety checks, subspace solver a bit safer in case of mu…

e37bab2

…ltiblock situations.

style

b0bfce0

ebchin approved these changes Apr 10, 2026

View reviewed changes

tupek2 and others added 4 commits April 16, 2026 10:05

Merge branch 'develop' into tupek/system_solver

8a99672

Working on merge.

bc3bfed

Fix merge issues.

ec1f452

sub options for block preconditioners

d78c213

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Tupek/system solver#1556

Tupek/system solver#1556
tupek2 wants to merge 146 commits intodevelopfrom
tupek/system_solver

tupek2 commented Mar 16, 2026 •

edited

Loading

Uh oh!

tlroy commented Apr 8, 2026

Uh oh!

tlroy commented Apr 8, 2026

Uh oh!

ebchin left a comment

Uh oh!

ebchin Apr 9, 2026

Uh oh!

ebchin Apr 9, 2026

Uh oh!

ebchin Apr 9, 2026

Uh oh!

ebchin Apr 10, 2026

Uh oh!

ebchin Apr 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

tupek2 commented Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tlroy commented Apr 8, 2026

Uh oh!

tlroy commented Apr 8, 2026

Uh oh!

ebchin left a comment

Choose a reason for hiding this comment

Uh oh!

ebchin Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

ebchin Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

ebchin Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

ebchin Apr 10, 2026

Choose a reason for hiding this comment

Uh oh!

ebchin Apr 10, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

tupek2 commented Mar 16, 2026 •

edited

Loading