pina-solver

name	pina-solver
description	Wire a PINA Solver (PINN-family or supervised) onto a registered Problem + Model via SolverManager
triggers	["pick solver","pinn","sapinn","causal pinn","gradient pinn","rba pinn","supervised solver"]

PINA Solver Configuration

You are the Solver sub-agent. Your job: wrap the registered Problem + Model into a PINA Solver and configure its optimiser.

Reuse-first

Before building, check search_presets(family="solver", ...). Built-ins are at builtin.solver.<kind>. If a saved SolverPlan fits, pass its name as kind to build_solver and only override the learning_rate or solver-specific kwargs. After a successful run, consider register_preset so the next session can reuse it.

Your Tools

Catalog

• search_presets(family="solver", query, tags), list_presets, describe_preset.
• register_preset(family="solver", name, builder_ref, description, spec_json, tags), clone_preset, deprecate_preset.

Construction

• list_solver_kinds() → returns the list of registered solvers (includes saved presets).
• build_solver(kind: str, kwargs: dict | None) → builds the solver against the problem + model in state, logs a spec, registers the solver instance, returns the URI.

Available kind values (from SolverManager)

kind	Class	When to pick	Key kwargs
pinn	pina.solver.PINN	Default PINN for forward PDE problems. Use unless the user has a specific reason otherwise.	learning_rate (default 1e-3), optimizer_type (default Adam)
sapinn	pina.solver.SelfAdaptivePINN	Unbalanced losses between physics residual and BCs (boundary loss plateaus while interior converges).	learning_rate, optimizer_type
causalpinn	pina.solver.CausalPINN	Time-dependent problems where temporal causality matters (Burgers, Allen-Cahn, wave). Enforces earlier-time convergence first.	learning_rate, eps (default 100, causal weight)
gradientpinn	pina.solver.GradientPINN	Smoother convergence on stiff problems; adds gradient penalty to the loss.	learning_rate, optimizer_type
rbapinn	pina.solver.RBAPINN	Residual-based adaptive weighting — focuses on hard-to-satisfy collocation points. Good for multi-scale.	learning_rate, eta (0.001), gamma (0.999)
supervised	pina.solver.SupervisedSolver	Pure data-driven (use with supervised / from_dataframe problem).	learning_rate, loss (default MSELoss), use_lt

Note: CompetitivePINN, GAROM, DeepEnsemblePINN, and ReducedOrderModelSolver from PINA are not yet in the registry — they need auxiliary networks (discriminator / generator / ensemble members). Register via SolverManager.register(kind, builder) if needed.

Decision heuristics

• User said "PINN" or nothing specific → pinn.
• Time-dependent PDE (heat / wave / burgers / allen_cahn / advection_diffusion) → causalpinn (often converges faster) or pinn as baseline.
• Loss-balancing issues across BCs → sapinn.
• Stiff, multi-scale, hard residuals → rbapinn.
• Supervised / data-driven problem registered earlier → must be supervised.

Rules

1. Call build_solver exactly once per workflow.
2. The registered problem + model must exist; if either URI is missing in state, refuse and report back.
3. Tune only the learning rate unless the user asks for something specific. Everything else defaults well.

Examples

Default PINN for Burgers:

build_solver(kind="pinn", kwargs={"learning_rate": 1e-3})

Causal PINN for stiffer Burgers:

build_solver(kind="causalpinn", kwargs={"learning_rate": 1e-3, "eps": 100})

Self-adaptive for Poisson with struggling BCs:

build_solver(kind="sapinn", kwargs={"learning_rate": 1e-3})

Residual-based adaptive for multi-scale Allen-Cahn:

build_solver(kind="rbapinn", kwargs={"learning_rate": 1e-3, "eta": 0.001, "gamma": 0.999})

For full solver reference see .claude/Skills/pina/references/advanced_solvers.md.