Skip to content

Simulation to learning

AgentFEM does not replace PyTorch or a user's neural-network stack. It makes the path from a parameterized finite-element model to accepted, traceable training data much shorter and more systematic.

Workflow

parameter space → reproducible FEM campaign → result policy
                → accepted ScientificDataset → NumPy/PyTorch adapter
                → built-in or user model → validation/applicability guard
                → learned prediction or FEM fallback

Available building blocks

  • parameter campaigns with persistent case identity and parallel execution;
  • dataset acceptance rules rather than silent collection of failed cases;
  • NumPy and optional PyTorch adapters;
  • ridge and POD-ridge baselines plus an optional PyTorch MLP;
  • user-owned PyTorch estimators without inheritance from a proprietary model class;
  • training/validation separation, applicability checks, and high-fidelity FEM fallback.

Two integration paths

A user-owned neural-field solver can enter the same Step lifecycle directly:

step = model.step(target=spec, executor=my_solver)
result = step.solve_result()

This route is framework-neutral and requires neither inheritance from an AgentFEM model class nor an official learning package. When a maintained method binding, standard artifacts, examples, and benchmark evidence are useful, install a provider from the optional AgentFEM-Learning companion. XDEM is its first experimental neural-field subdomain; it is not a separate official project or a general fracture claim.

This observation and dataset contract is the basis for future neural operators and digital-twin state updating. Those directions require stable field encodings, sensor identity, time alignment, and uncertainty—not just another neural-network class.

Go deeper