AutoMOOSE Brings Agentic AI to Materials Simulations, Automating Complex Physics Models from Natural Language

AutoMOOSE is an agentic software system designed to translate natural-language requests directly into executable, validated MOOSE phase-field simulations—automating workflows that would normally require skilled engineers to set up manually. The framework connects user intent to running computational models, then verifies and interprets the results before delivering them back.

Architecture and Design#

The system is built around a six-agent pipeline that orchestrates the translation from human requests to running physics code. Rather than treating error-checking as an afterthought, AutoMOOSE embeds a screening and recovery loop that can detect physical inconsistencies, flag them as falsifications, and either repair them automatically or halt execution until a human reviewer approves corrections. This design choice ensures that any corrected simulation remains transparent and inspectable—no silent fixes.

The framework also includes:

• A physics plugin layer for extensibility across different material science domains
• A Model Context Protocol interface for interoperability with other tools and agents
• Component separation between physical validation and automatic repair

Scope and Reusability#

Rather than remaining a one-off research tool, the paper frames AutoMOOSE as a reusable framework aimed at supporting a broad range of multiphysics materials design problems. The authors emphasize software design principles that prioritize extensibility and interoperability, enabling other researchers and practitioners to adapt the system to their own simulation needs without rebuilding from scratch.

This positions agentic AI not just as a code-generation novelty, but as a practical layer for scientific computing where correctness and interpretability matter as much as automation.