Skip to content

AgentFEM: An AI-Native Open-Source Platform for Finite-Element Computing

Haoming Luo
Xi'an Thermal Power Research Institute, Xi'an 710032, China
Preprint / Technical Report ยท August 2026

Download the paper (PDF) View AgentFEM on GitHub

Abstract

AI agents can increasingly construct solver programs, but dependable finite-element simulation still requires explicit models, numerical procedures, and evidence. This paper presents AgentFEM, an open-source platform for finite-element workflows shared by people and agents. Its readable Python language expresses the engineering model, a reusable capability layer implements finite-element methods, and FEniCSx, UFL, PETSc, and MPI provide the numerical foundation. Results record solver status, checks, and origin, and the same workflow extends to parameter studies and surrogate modelling.

Three cases demonstrate the present scope: a linear-elastic beam, wave propagation through an inclusion, and finite-strain homogenization of a periodic porous cell using C3D10H elements. The current release also provides a direct path from parameterized simulation to learning-ready data and guarded surrogate models. AgentFEM is the AI-native CAE platform; agents and development environments provide ways to work with it. Its broader purpose is to make dependable finite-element software freely available and to prepare CAE for a future in which people and AI advance science together.

What the paper covers

  • a shared finite-element workflow for humans and AI agents;
  • the engineering, finite-element capability, and numerical-kernel layers;
  • explicit distinctions among computed, converged, verified, and validated results;
  • representative static, wave-propagation, and finite-strain applications;
  • the path from parameterized simulation to scientific datasets and guarded surrogate models.

Citation

@techreport{luo2026agentfem,
  author      = {Luo, Haoming},
  title       = {AgentFEM: An AI-Native Open-Source Platform for
                 Finite-Element Computing},
  institution = {Xi'an Thermal Power Research Institute},
  year        = {2026},
  month       = {August},
  type        = {Preprint / Technical Report},
  url         = {https://haoming-luo.github.io/agentfem/publications/agentfem-platform/}
}

The manuscript describes AgentFEM version 0.2.1. The project continues to evolve; consult the release notes for subsequent capabilities.

Keywords: finite-element method; scientific software; AI agents; computational mechanics; simulation workflows; surrogate modelling; FEniCSx