The Polymer Society of Korea

Abstract

Computational polymer design with molecular dynamics is challenging due to complications in model selection and generation and the difficulty of ensuring equilibration in systems with intrinsic slow relaxation. We introduce AutoPolySim (APS), a human–agent co-scientist framework that turns natural-language research questions into reproducible molecular dynamics result summaries. APS collaborates with the user to design proper polymer simulation methods and construct robust models; subsequently, it automates the execution and analysis of the simulations to report predicted properties and conformational metrics. APS interoperates with standard force fields (e.g., AMBER, OPLS-AA) and fills missing terms via targeted parameter generation, including quantum-mechanical fitting of charges and torsions. Capable of executing on both local and HPC resources, APS supports both open-weight and state-of-the-art LLMs. We conclude with an APS-driven case study and outline future extensions.