Skip to content

AgentFEM 0.2.0a1

Released: 3 August 2026 Status: public alpha preview

AgentFEM was initiated by Haoming Luo and open-sourced on GitHub in July 2026.

Version 0.2.0a1 is the first public preview of AgentFEM as an AI-native finite-element platform rather than a collection of isolated helpers. It is a deliberately early release for researchers, engineers, and AI/CAE developers who want to test the workflow, inspect its design, and help shape the 0.2 series.

Highlights

  • A readable Study -> Model -> Step -> Result FEM workflow for humans and agents.
  • Linear, nonlinear, thermal, implicit-dynamic, and explicit-dynamic procedures on the current FEniCSx/PETSc/MPI kernel.
  • Neo-Hookean finite strain and a global small-strain J2 path with quadrature state, consistent tangent, physical cutback controls, cyclic loading, energy histories, and serial restart.
  • Abaqus C3D10 and equation import, external mesh conversion, and distributed periodic workflows.
  • Unified result, progress, checkpoint, Golden-benchmark, and quality-policy evidence.
  • Reproducible campaigns, scientific datasets, PyTorch adapters, surrogate validation, applicability guards, and high-fidelity fallback.
  • Versioned scientific knowledge cards, benchmark contracts, and explicit capability maturity boundaries.

Install

Create the FEniCSx numerical environment first, then opt into the preview:

mamba create -n agentfem-env -c conda-forge \
  python=3.11 fenics-dolfinx=0.11 mpich mpi4py petsc4py h5py
mamba activate agentfem-env
python -m pip install --pre agentfem

Release Evidence

The tagged source is accepted only after:

  • the source version, package version, and Git tag agree;
  • the wheel and source distribution pass metadata and payload checks;
  • the installed wheel passes serial and distributed FEniCSx tests;
  • the two-rank MPI regression passes;
  • the static, heat, creep, wave, and nonlinear release contracts pass; and
  • a separate environment installs the official CPU-only PyTorch wheel and verifies the optional simulation-to-learning interfaces.

Honest Boundaries

This alpha is not a universal CAE replacement. It does not claim global adaptive creep/damage, portable MPI restart for quadrature state, general UMAT/UHYPER binary compatibility, industrial code compliance, automatic arbitrary-mesh neural-operator training, or a complete native-Windows solver stack. WSL2 is the recommended Windows route.

The 0.2 series will deepen the implemented workflows and their external benchmarks before expanding the public vocabulary indiscriminately.