AutoMOOSE: An Agentic AI for Autonomous Phase-Field Simulation

Multiphysics simulation frameworks such as MOOSE provide rigorous engines for phase-field materials modeling, yet adoption is constrained by the expertise required to construct valid input files, coordinate parameter sweeps, diagnose failures, and extract quantitative results. We introduce AutoMOOSE, an open-source agentic framework that orchestrates the full simulation lifecycle from a single natural-language prompt. AutoMOOSE deploys a five-agent pipeline in which the Input Writer coordinates six sub-agents and the Reviewer autonomously corrects runtime failures without user intervention. A modular plugin architecture enables new phase-field formulations without modifying the core framework, and a Model Context Protocol (MCP) server exposes the workflow as ten structured tools for interoperability with any MCP-compatible client. Validated on a four-temperature copper grain growth benchmark, AutoMOOSE generates MOOSE input files with 6 of 12 structural blocks matching a human expert reference exactly and 4 functionally equivalent, executes all runs in parallel with a 1.8x speedup, and performs an end-to-end physical consistency check spanning intent, finite-element execution, and Arrhenius kinetics with no human verification. Grain coarsening kinetics are recovered with R^2 = 0.90-0.95 at T >= 600 K; the recovered activation energy Q_fit = 0.296 eV is consistent with a human-written reference (Q_fit = 0.267 eV) under identical parameters. Three runtime failure classes were diagnosed and resolved autonomously within a single correction cycle, and every run produces a provenance record satisfying FAIR data principles. These results show that the gap between knowing the physics and executing a validated simulation campaign can be bridged by a lightweight multi-agent orchestration layer, providing a pathway toward AI-driven materials discovery and self-driving laboratories.

💡 Research Summary

AutoMOOSE is an open‑source, agentic framework that eliminates the steep expertise barrier traditionally associated with running phase‑field simulations in the Multiphysics Object‑Oriented Simulation Environment (MOOSE). By accepting a single natural‑language prompt, AutoMOOSE orchestrates the entire simulation lifecycle—design, input generation, execution, failure recovery, and results analysis—without any manual coding or intervention.

The system is built around a five‑agent pipeline: (1) Architect parses the user request and produces a structured simulation plan (JSON) that encodes the physics formulation, geometry, boundary conditions, sweep parameters, and solver settings. (2) Input Writer is a compound agent that coordinates six sub‑agents (Mesh, GlobalParams, Variables, Kernels, Materials, Postprocessors, Executioner) to generate a syntactically valid MOOSE .i file. Crucially, the Input Writer translates user‑friendly physical quantities (grain‑boundary energy σ, interface width w_GB, mobility prefactor M₀) into the numerical coefficients L, μ, κ required by the Allen–Cahn equations using the analytical relations in Eq. (4), thereby removing a common source of human error.

(3) Runner launches MOOSE, streams solver output, and detects convergence failures. When a failure occurs, the log is handed to (4) Reviewer, which classifies the error (e.g., time‑step too large, mesh insufficient, out‑of‑range material parameters), proposes corrected parameters, and feeds them back to the Input Writer via a single “Retry” arc. This closed‑loop correction completes automatically, typically within one cycle.

(5) Visualization consumes the CSV post‑processor output, extracts the grain count N(t) and the macroscopic rate constant ˜k(T), fits an Arrhenius relationship, and returns both plots and a textual report. All metadata, including every version of the simulation plan and any parameter adjustments, are stored in a provenance record that satisfies FAIR data principles.

A modular plugin architecture decouples physics‑specific logic from the orchestration core. New phase‑field formulations can be added by implementing a two‑function contract (initialization and block generation) and registering the plugin; no changes to the core agents are required. The framework also exposes a Model Context Protocol (MCP) server with ten structured tool endpoints, enabling headless operation and seamless integration with external AI clients or optimization pipelines.

Validation was performed on a copper polycrystalline grain‑growth benchmark. The user request specified four temperatures (300, 450, 600, 750 K), σ = 0.708 J m⁻², w_GB = 14 nm, M₀ = 2.5 × 10⁻⁶ m⁴ J⁻¹ s⁻¹, activation energy Q = 0.23 eV, 15 initial grains in a 1 µm × 1 µm periodic domain. AutoMOOSE generated a MOOSE input file where six of the twelve structural blocks matched a human‑expert reference exactly, and the remaining four were functionally equivalent. All four simulations ran in parallel, achieving a 1.8× wall‑clock speed‑up relative to serial execution. Grain‑coarsening kinetics were recovered with R² = 0.90–0.95 for T ≥ 600 K, and the fitted activation energy Q_fit = 0.296 eV closely matched the human‑written reference Q_fit = 0.267 eV (difference ≈ 0.03 eV). Three distinct convergence‑failure classes were automatically diagnosed and corrected within a single retry cycle.

Limitations include the current focus on Allen‑Cahn‑based multiphase models, limited support for anisotropic boundary conditions, and the need for further scaling tests on large HPC systems. Future work will explore Retrieval‑Augmented Generation for more robust input synthesis, GPU‑accelerated MOOSE execution, and closed‑loop integration with laboratory equipment to realize truly self‑driving computational‑experimental laboratories.

Overall, AutoMOOSE demonstrates that a lightweight, multi‑agent orchestration layer can bridge the gap between scientific intent and validated multiphysics simulation, paving the way for AI‑driven materials discovery and autonomous computational laboratories.