MLIPilot: LLM-Driven Auto-Research for Machine-Learned Interatomic Potentials
Researchers have developed MLIPilot, an automated research framework that uses large language models to optimize machine-learned interatomic potentials. This system leverages LLMs to propose hypotheses, modify training code, and manage high-performance computing jobs, all guided by a physics-based scorecard. When tested with models like GPT-5.5 and Mistral-24B, MLIPilot successfully discovered effective training strategies, including normalization, loss function adjustments, and progressive scheduling, moving beyond manual trial-and-error in scientific machine learning. AI
IMPACT Automates complex scientific workflows, potentially accelerating discovery in materials science and chemistry.