PolyJarvis: LLM Agent for Autonomous Polymer MD Simulations

A research team from Carnegie Mellon University and Stanford has developed PolyJarvis, a novel AI agent that automates the entire workflow for polymer molecular dynamics (MD) simulations. The system connects a large language model (LLM) to the RadonPy simulation platform through Model Context Protocol (MCP) servers, creating an end-to-end pipeline that translates natural language queries into precise computational experiments. Users simply input a polymer's name or its SMILES string—a standard notation for molecular structure—and the agent autonomously executes the complex, multi-step process that typically requires specialized expertise. This includes monomer construction, charge assignment, polymerization, force field parameterization, GPU-accelerated system equilibration, and final property calculation.

In validation tests on common polymers like polyethylene, atactic polystyrene (aPS), poly(methyl methacrylate) (PMMA), and poly(ethylene glycol), PolyJarvis demonstrated promising accuracy. For aPS and PMMA, it predicted material densities within 0.1% to 4.8% of reference values and calculated bulk moduli within 17% to 24%. The agent's prediction for PMMA's glass transition temperature (Tg) was 395 K, matching experimental data within a range of +10 to +18 K. While some Tg predictions for other polymers showed larger deviations—attributed to intrinsic biases in MD simulation cooling rates rather than agent error—the system successfully met strict acceptance criteria for 5 out of 8 property-polymer combinations with direct experimental comparisons. The research, published on arXiv, demonstrates that LLM-driven agents can now produce simulation results consistent with those run by human experts, significantly lowering the barrier to advanced materials modeling.