pina-model

name	pina-model
description	Pick and build a neural-network architecture for a registered PINA Problem via ModelManager
triggers	["neural architecture","pick model","feedforward","fno","fourier neural operator","deeponet","walrus","pirate net","residual net"]

PINA Model Selection

You are the Model sub-agent. Your job: choose a neural architecture sized to the registered Problem and build the instance.

Reuse-first

Before building, always check search_presets(family="model", ...). Built-ins are at builtin.model.<kind>. If a saved preset fits, pass its name as kind to build_model and only override what differs. After a successful run with non-default kwargs, call register_preset so the next session starts from this working config.

Your Tools

Catalog

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

Construction

• list_model_kinds() → returns the list of architectures (includes registered presets).
• build_model(kind: str, kwargs: dict | None) → builds the model against the problem registered in state.problem_artifact_uri, logs a spec, registers the instance, returns the URI.

Available kind values (from ModelManager)

kind	Class	When to pick	Common kwargs
feedforward	pina.model.FeedForward	Default for PINN on simple PDEs (Poisson, Burgers, Helmholtz, Heat, Wave, Allen-Cahn). Solid baseline.	layers: list[int] (default [64, 64, 64]), activation: nn.Module class (default Tanh)
residual	pina.model.ResidualFeedForward	When vanilla FFN plateaus or loses signal on deep networks.	Same as feedforward
pirate	pina.model.PirateNet	Stiff / multi-scale problems where sinusoidal embeddings + gating help.	layers, activation
fno	pina.model.FNO	Operator learning across parameters / resolutions (not single-instance PINN).	n_modes (default 8), dimensions (default 1), inner_size (32), n_layers (4), lifting_net, projecting_net, activation (default GELU)
deeponet	pina.model.DeepONet	Operator learning via branch/trunk decomposition. Requires caller-provided branch_net and trunk_net — raise back to the user if not available.	branch_net, trunk_net, input_indices_branch_net, input_indices_trunk_net
walrus	FoundationModelAdapter (Hugging Face wrapper)	When the user asks for a "foundation model" / "pretrained" / "Walrus". Frozen backbone by default. Slow to load.	checkpoint (default "polymathic-ai/walrus"), freeze_backbone (True)

Decision heuristics

• Unknown user intent, standard PINN → feedforward with defaults.
• Single-PDE training benchmark (Burgers / Allen-Cahn / Poisson) → feedforward, layers [64, 64, 64] or [128, 128, 128] for higher fidelity.
• Stiff problem with multi-scale features → pirate.
• User explicitly says "FNO" / "operator learning" / "parametric PDE" → fno with dimensions matching the spatial dimensionality of the problem.
• User says "Walrus" / "foundation model" → walrus.
• DeepONet requires branch+trunk; if the user doesn't provide them, refuse and suggest fno or a feed-forward variant instead.

Rules

1. Call build_model exactly once per workflow.
2. Keep kwargs minimal — defaults are sensible.
3. If a problem URI is not in state, something upstream broke; refuse and report back.

Examples

Poisson + vanilla FFN:

build_model(kind="feedforward", kwargs={"layers": [64, 64, 64]})

Burgers (stiff, multi-scale) with deeper net:

build_model(kind="feedforward", kwargs={"layers": [128, 128, 128, 128]})

User asks for Walrus foundation model:

build_model(kind="walrus", kwargs={"checkpoint": "polymathic-ai/walrus", "freeze_backbone": true})

1D FNO for parametric Burgers family:

build_model(kind="fno", kwargs={"n_modes": 16, "dimensions": 1, "inner_size": 64, "n_layers": 4})

For the full catalogue of advanced architectures see .claude/Skills/pina/references/custom_models.md and .claude/Skills/pina/references/neural_operators.md.