CGMas: Multi-agent LLM framework automates polymer MD, 38-88x faster

Coarse-grained (CG) molecular dynamics is essential for simulating polymers beyond the reach of all-atom (AA) methods, but building accurate CG models is a labor-intensive manual process. Each polymer requires a custom mapping and potential parameter set, making automation a long-sought goal. Now, researchers have introduced CGMas, a multi-agent framework that leverages large-language-model (LLM) reasoning agents to automate the entire CG pipeline—from topology inference and equilibration to mapping, potential derivation via Boltzmann inversion, and validation. Users simply specify the polymer and target resolution in natural language; the system's layered self-correction logic then resolves physical errors common in unsaturated, heteroatom-containing, and polar polymers, which have historically been the hardest cases for automated workflows.

In benchmarking on 27 homopolymer and copolymer tasks, CGMas demonstrated striking performance: it completed all tasks, matched the AA reference density within 5% in 22 of them, and reduced simulation time from 38–88 minutes to just 1 minute per task. This speedup, roughly 40–90x, arises from replacing manual trial-and-error with efficient LLM-driven decision-making. The framework establishes agentic LLMs as a viable route to automated polymer coarse-graining, potentially accelerating materials discovery and enabling non-experts to generate high-quality CG models without deep domain knowledge. While density matching isn't perfect for every polymer, the results suggest that CGMas could become a standard tool in computational polymer science.