LLM-Powered Multi-Agent Framework Automates Polymer Coarse-Graining

Coarse-grained (CG) molecular dynamics is a powerful technique for simulating polymers beyond the scale of all-atom (AA) methods, but the process of creating CG models is typically time-consuming and requires specialized expertise. Each polymer mapping often necessitates a new derivation of potentials, as transferable parameter sets are rarely available. To streamline this, a new multi-agent framework called CGMas has been introduced. CGMas automates the entire workflow, starting from a natural-language description of the polymer and its desired resolution. An LLM reasoning agent infers the all-atom topology, with layered self-correction mechanisms to resolve common physical errors. Subsequent agents handle system equilibration, mapping to CG representation, potential derivation via Boltzmann inversion, and benchmarking against atomistic references. CGMas successfully completed all 27 homopolymer and copolymer tasks, achieving atomistic density within 5% for most cases and drastically reducing simulation time from hours to just one minute. This demonstrates the potential of agentic LLMs for automated polymer coarse-graining.