MLIPilot: LLM-Driven Auto-Research for Machine-Learned Interatomic Potentials
TLDR
MLIPilot uses LLM agents to autonomously optimize machine-learned interatomic potentials via hypothesis proposal, code editing, and HPC job management.
Reasoning
The paper presents a novel framework integrating LLMs with domain-specific constraints for automated MLIP development, with strong empirical evaluation across multiple LLMs and datasets. However, the abstract lacks details on limitations, such as generalizability beyond MLIPs or potential failure modes.
Read-first score
Read-first score 71.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 78.
Field roles
Rank sensitivity
Stability: volatile; rank range: 17.
Keyword Scores
Deep Analysis
Innovations
- MLIPilot: an auto-research framework where tool-calling LLM agents propose hypotheses, edit MLIP training code, launch HPC jobs, and accept/revert changes using a fixed, physically constrained scorecard.
Methodology
MLIPilot uses tool-calling LLM agents (GPT-5.5, GPT-4.1, Mistral-24B, Qwen3-32B) to iteratively optimize MACE interatomic potentials by editing code, submitting HPC jobs, and evaluating against a physically constrained scorecard. Evaluation is performed on a QM7-derived molecular dataset with B3LYP/6-31G(d) energies/forces and a Cu EMT periodic dataset with ASE Effective Medium Theory labels.
Key Results
The strongest LLM agents transformed initially constraint-violating baselines into accepted models by autonomously discovering training strategies such as output normalization, loss-function changes, progressive training schedules, and model-capacity adjustments.