The First Generation of Open-Source AI-Native FEM

全球第一代开源 AI 原生有限元平台

AgentFEM

AgentFEM is an open-source finite-element platform that turns an engineering
analysis into a readable Python workflow: define the study, model, materials,
loads, solution procedure, outputs, and verification in one place. The same
workflow can be understood and operated by researchers, scripts, IDEs, future
GUIs, and AI agents.

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

Its immediate goal is practical: to become a dependable and unusually usable
open-source FEM platform. Its longer-term vision is to make finite-element
simulation an accessible scientific workspace connecting engineering,
computation, data, and AI.

All You Need Is an Agent

[!TIP]
Give this prompt to Codex, DS Harness, or another AI agent:

Bring AgentFEM to life.

Use https://github.com/haoming-luo/agentfem as the guide. Follow the tested
route in INSTALL.md; if I say "走镜像通道 / use the mirror channel", use its
single verified mirror route. Read AGENT_GUIDE.md and run `agentfem doctor`.

When it is ready, reply briefly with the environment, AgentFEM
version, and health-check result.

AgentFEM includes the guidance and machine-readable interfaces an agent needs.
Prefer manual setup? Continue to Install.

Why AgentFEM

• AI-Native FEM — finite-element software designed from the start for
  agents to construct, operate, and automate naturally, without replacing
  deterministic mechanics and numerical computation with AI.

• Humans and Agents, Together — people and AI agents work through the same
  readable materials, regions, loads, solution steps, and results. AI work
  remains understandable, editable, and reusable by humans.

• Results You Can Check — convergence, failures, required outputs,
  benchmark comparisons, and applicability limits remain attached to the
  result instead of being separated from the simulation that produced it.

• One Run or Thousands — the same model can support an individual
  analysis, parameter campaigns, parallel execution, restartable studies, and
  reproducible data generation.

• Simulation to Learning — results can flow into scientific datasets,
  user-owned models, PyTorch, surrogate models, and high-fidelity fallback
  without rebuilding the workflow around separate glue scripts.

• Open at Every Layer — users can begin with a clear engineering workflow
  and still reach operators, UFL, DOLFINx, PETSc, and custom constitutive
  models whenever needed.

Our conviction: Open FEM for everyone. Useful simulation within reach
with AI. Engineering AI grounded in physical models, observations, and
verification.

Install

AgentFEM supports Linux, macOS, and Windows through WSL2.

Recommended: conda-forge

Install AgentFEM and its compatible FEniCSx/PETSc/MPI foundation together:

mamba create -n agentfem-env -c conda-forge agentfem
mamba activate agentfem-env
agentfem doctor

Already in a conda-forge environment? Run mamba install agentfem.

On Windows, run these commands inside an Ubuntu WSL2 terminal, then protect
projects and results from distribution removal:

agentfem workspace --protect

If you tell an agent "use the mirror channel", or the canonical package
source is unreachable, use the TUNA conda-forge mirror without changing your
global conda configuration. AgentFEM does not infer this choice from IP or VPN
location:

mamba create -n agentfem-env --no-rc --override-channels \
  -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge \
  python=3.11 fenics-dolfinx=0.11 agentfem
mamba activate agentfem-env
agentfem doctor

The mirror carries byte-identical conda artifacts for synchronized releases
but may briefly lag a new release. See the installation notes
for verification, fallback, and cache-refresh commands.

Optional: Complete Runtime Preview

For offline setup without managing an environment, preview bundles are
available for Apple Silicon macOS and Windows 10/11 through WSL2.

Download the latest Complete Runtime →

On Windows, extract the ZIP and run in PowerShell:

powershell -ExecutionPolicy Bypass -File .\Install-AgentFEM.ps1

See the runtime guide for upgrades, durable
project storage, removal, and Preview-specific troubleshooting.

Existing FEniCSx environment: PyPI

If you already maintain a compatible FEniCSx environment:

python -m pip install agentfem
agentfem doctor

In mainland China, the same Python package is also available from the TUNA
PyPI mirror; use that route only inside an already compatible FEniCSx/PETSc
environment.

See INSTALL.md for platform and MPI details, or the
runtime guide for Preview installation and
integrity checks.

Optional capabilities stay separate from the core Python package:

python -m pip install 'agentfem[mesh-formats]'   # Abaqus/NASTRAN meshes
python -m pip install 'agentfem[gmsh]'           # Gmsh model/.msh import
python -m pip install 'agentfem[visualization]'  # ParaView-ready helpers
python -m pip install 'agentfem[ml]'             # PyTorch adapters

Gmsh is an optional, separately licensed GPL component. It is not contained in
the Apache-2.0 Python package; the recommended offline Complete runtime may
aggregate it with its license and corresponding source for a one-click
CAD-to-mesh workflow.

AgentFEM can share a minimal anonymous reliability signal to improve the free,
open-source software; models, meshes, parameters, code, paths, and results are
never included. Inspect or disable it with agentfem telemetry status|off. See the
feedback and privacy contract.

Run Your First Model

Create and run a complete static-solid project in any directory:

mkdir first-agentfem-model && cd first-agentfem-model
agentfem init --template static-solid .
agentfem check
agentfem run --name baseline
agentfem show latest

The generated case.py is ordinary, editable Python. Its public workflow reads
like an engineering analysis:

study = studies.static_solid(dimension=2, assumption="plane_strain")
model = models.create(study=study, mesh=domain, name="cantilever")
u = model.field(fields.displacement(domain, degree=1))

model.material(elasticity.isotropic_elastic(young=210e9, poisson=0.30))
model.clamp(u, on=left)
model.traction((0.0, -1.0e6), on=right)

result = model.step(target=u, name="static_load").solve_result()
result.verify("engineering").require()

The CLI gives the same model a readable run folder such as
outputs/001-baseline/, an immutable evidence identity, a concise result
summary, an MPI launch path, and a machine-readable interface. Use
agentfem runs to find earlier runs and --json when an agent or GUI needs the
complete record.
You can also run case.py directly with Python.

What Works Today

AgentFEM records capability maturity explicitly. A working material-point law,
an integrated global solver, and an externally verified analysis are different
levels of evidence; the software does not silently treat them as equivalent.
See the capability and verification guide
for the detailed scope.

Release Examples

• Static elasticity — the readable beginner
  workflow.
• NAFEMS LE10 3D elasticity — a public curved-
  solid pressure benchmark with stress, displacement, force, energy, and MPI
  evidence.
• Transient heat transfer — implicit time
  integration, progress, and field output.
• Modal cantilever — linear natural
  frequencies, mass-normalized mode fields, residuals, orthogonality, and MPI
  evidence.
• NAFEMS R0016 Test 5H direct harmonic response
  — a public forced-vibration comparison for the provider-neutral
  K/M/C/K_loss/F procedure.
• Wave propagation with an inclusion —
  dynamic fields, source amplitude, and boundary models.
• Abaqus C3D10H periodic cell — direct
  mesh/equation import, quasi-incompressible hyperelasticity, and homogenized
  response.
• J2 plasticity and
  global creep — stateful nonlinear
  material workflows with cutback and restart.
• Simulation-to-surrogate campaign
  — accepted FEM data, surrogate validation, applicability guard, and FEM
  fallback.

These are executable release assets with numerical contracts, not only syntax
demonstrations. More examples are indexed in examples/ and on
the documentation site.

Open and Extensible

AgentFEM has three visible layers:

Engineering workflow
    -> reusable FEM operators, constitutive laws, constraints, and outputs
        -> FEniCSx / DOLFINx / PETSc / MPI numerical kernel

Users can stay in the concise engineering workflow or descend to operators,
UFL, DOLFINx, PETSc, and custom constitutive implementations when a research
problem needs a lower layer. This is also the extension path for user
materials, new elements, private domain modules, GUIs, and agent tools.

Learning follows the same rule. A laboratory-owned model can connect directly
through the framework-neutral model.step(target=spec, executor=...)
boundary. The optional
AgentFEM-Learning
companion provides maintained providers, examples, and benchmark evidence for
selected scientific-learning methods; it is not required to use a user's own
model.

Documentation

• Getting started
• Standard modeling workflow
• Engineering concepts
• Scientific functions and theory
• Results, campaigns, and learning
• AI-native learning contracts
• AI-agent guide
• MCP connection for compatible agents
• Roadmap and release gates

The complete user and scientific reference is available at
haoming-luo.github.io/agentfem.

Support the Project

If AgentFEM helps, one explicit command can Star the project through an
existing GitHub CLI login:

agentfem support --star

It never runs automatically. You can also visit
GitHub directly. Have a question or
a result to show? Visit Q&A
or Show and tell.

Citation

Please cite the AgentFEM version used. GitHub's Cite this repository uses
CITATION.cff.

@software{luo2026agentfem036,
  author  = {Luo, Haoming},
  title   = {AgentFEM},
  year    = {2026},
  version = {0.3.6},
  doi     = {10.5281/zenodo.22847104},
  url     = {https://doi.org/10.5281/zenodo.22847104}
}

For architecture and methodology, also cite the
AgentFEM Technical Report
(PDF).

@techreport{luo2026agentfem,
  author      = {Luo, Haoming},
  title       = {AgentFEM: An AI-Native Open-Source Platform for
                 Finite-Element Computing},
  institution = {AgentFEM},
  year        = {2026},
  month       = {August},
  doi         = {10.5281/zenodo.22847132},
  url         = {https://doi.org/10.5281/zenodo.22847132}
}

Author

Haoming Luo is the initiator and maintainer of AgentFEM. His interests include computational mechanics, materials engineering, finite-element simulation, and AI-assisted scientific computing, with education and research experience associated with NWPU, INSA Lyon and Ecole Polytechnique.

The project is also motivated by engineering needs in materials evaluation,
defect inspection, and simulation analysis for power-generation equipment.

License

AgentFEM is available under the Apache License 2.0. It can be used,
modified, and extended in research, education, and commercial products under
the terms of that license.