AutoMOOSE

Agentic AI for autonomous phase-field simulation

AutoMOOSE is a multi-agent LLM framework that automates the end-to-end workflow of phase-field simulations in MOOSE — from natural-language problem specification through input-file generation, HPC execution, error recovery, and post-processing of results.

The system orchestrates a team of specialized LLM agents — a planner, a MOOSE-input generator, a simulation runner, an error-diagnosis agent, and an analyst — communicating through a shared scratchpad and tool-call API. The goal is twofold: to lower the barrier to multiphysics simulation for non-expert users, and to enable autonomous design-of-experiments loops where the simulator becomes one tool among many in an agentic materials-discovery pipeline.

Status: Active development · Preprint released 2026

Role: Lead developer and corresponding author

Key features

  • Natural-language to MOOSE input translation with self-correction loops
  • HPC-aware job submission and monitoring
  • Agentic error diagnosis grounded in simulation logs and physical reasoning
  • Automated post-processing and figure generation

Citation

@article{manna2026automoose,
  title   = {{AutoMOOSE}: An Agentic AI for Autonomous Phase-Field Simulation},
  author  = {Manna, Sukriti and Chan, Henry and Sankaranarayanan, Subramanian K. R. S.},
  journal = {arXiv preprint arXiv:2603.20986},
  year    = {2026}
}