Skip to content

AgentFEM 0.2.2

AgentFEM 0.2.2 establishes an open scientific-learning interface without turning the finite-element core into a machine-learning framework. A laboratory-owned PyTorch, JAX, DeepXDE, or other field solver can now enter the ordinary AgentFEM lifecycle through one framework-neutral boundary:

step = model.step(
    target=neural_field_spec,
    executor=my_solver,
    executor_name="laboratory.my_solver",
    executor_version="1.0",
    executor_options={"epochs": 2000},
    output="results/neural_field",
)
result = step.solve_result()

The executor owns its tensors, architecture, optimizer, devices, and model state. AgentFEM owns the immutable scientific request, executor identity, common SimulationResult, portable manifest, and verification evidence. No AgentFEM model inheritance or official learning package is required.

Scientific-learning contracts

The public agentfem.learning namespace now distinguishes four roles:

  • surrogates approximate a declared parameter-to-response map;
  • neural operators learn a function-to-function map;
  • neural-field solvers optimize one physical field problem;
  • learned constitutive models update local material state.

NeuralFieldSpec records fields, representations, residual or variational objectives, conditions, sampling, trainable physical parameters, and required independent checks. The contract can describe PINN, VPINN, Deep-Ritz/DEM, XDEM, and related methods without claiming that arbitrary UFL forms can be translated automatically or that their external trainers are bundled.

The optional AgentFEM-Learning companion provides maintained method bindings, examples, dependency policy, and benchmark evidence. Its first experimental XDEM subdomain proves the full NeuralFieldSpec -> Step -> SimulationResult path against a Williams Mode-III reference. User and private providers remain free to use the same core contract directly.

Agent-first entry

The README now gives coding agents a direct first instruction before the manual installation route. The repository already ships the corresponding agent guide, reusable Skill, machine-readable capabilities, project templates, health checks, structured results, and verification commands. Human and agent users therefore operate the same public engineering workflow rather than two separate products.

Release evidence

The candidate is accepted only after:

  • version, citation, tag, release contract, wheel, and source archive agree;
  • the complete serial suite and distributed MPI suites pass;
  • checkpoint state remains portable across supported MPI rank counts;
  • the built wheel contains the neural-field contract and execution boundary;
  • a user-owned executor completes the common Step and result lifecycle in serial and MPI tests;
  • optional PyTorch dataset and surrogate bridges pass from the candidate wheel;
  • every installed project template and release-facing FEM workflow executes;
  • the documentation, machine entrypoints, and scientific knowledge catalog pass their strict checks.

This release creates an integration boundary; it does not claim bundled general XDEM, universal PINN training, automatic neural operators for arbitrary meshes, or validity of an external learned model without independent evidence.