marimo-flow

Guidance for composing 3D PINA problems — sampling budgets, activation choices, and gotchas specific to three spatial axes plus optional time.

Residual-based adaptive refinement for PINNs — use RBAPINN to grow attention on high-loss regions without hand-crafting a refined mesh. True h/p-AMR is out of scope for now.

Use an unstructured mesh (STL/OBJ/VTK/GMSH) as the spatial domain for a PINA problem — attach a MeshSpec and reference tagged cell regions via SubdomainSpec.mesh_ref.

Compose inverse / parameter-identification PINA problems — declare UnknownParameterSpec, attach ObservationSpec from data or synthetic sampling, and the composer wires PINA InverseProblem automatically.

Compose coupled multi-field PINA problems (e.g. thermo-elasticity, magnetohydrodynamics) by listing multiple EquationSpecs on one ProblemSpec — no new schema needed.

Pick and build a neural-network architecture for a registered PINA Problem via ModelManager

Compose a PINA Problem from primitives (equations + subdomains + conditions) via compose_problem. No hardcoded kinds — any PDE that sympy + PINA operators can express is reachable.

Wire a PINA Solver (PINN-family or supervised) onto a registered Problem + Model via SolverManager

Train the registered PINA solver via pina.Trainer (collocation sampling + gradient descent)

Interactive reactive Python notebook development with marimo - best practices, UI components, MCP integration, and deployment workflows

MLflow for ML lifecycle management - experiment tracking, LLM/GenAI tracing, model registry, and deployment with GenAI and MCP support

Physics-Informed Neural Networks with PINA - solve PDEs, inverse problems, and operator learning with PyTorch