CGAgentX: Agentic AI Framework to Develop Transferable Coarse-Grained Models

We present CGAgentX, a general autonomous multi-agent framework in which specialized LLM-based agents coordinate the optimization of coarse-grained (CG) model parameters to reproduce target properties. Using polar solvents dimethyl sulfoxide (DMSO) and N,N-dimethylacetamide (DMA) as representative case studies, we demonstrate the framework’s capability to develop CG models that accurately reproduce key properties from atomistic simulations and experimental literature. Six specialized agents – Mapping, Topology, Boundary, Hypothesis, Diagnostic, and Optimization – operate under a Master Agent that orchestrates closed-loop, iterative parameter refinement by autonomously invoking external tools, including molecular dynamics (MD) simulations and analysis workflows, and evaluating outputs through a fitness function. Central to the framework is a Hypothesis Agent that generates and verifies physically motivated parameter hypotheses by coordinating parallel multi-fork simulations, wherein multiple candidate parameter sets are evaluated simultaneously. This multi-fork strategy expands parameter space exploration, yielding richer datasets that enable more accurate hypothesis refinement across iterations. Agents adaptively propose parameter updates based on intermediate simulation outcomes, enabling efficient navigation of complex trade-offs among structural, thermodynamic, and transport properties. The framework reproduces key experimental properties within 5% accuracy while maintaining consistency with atomistic reference behavior, achieving convergence without manual intervention. The modular architecture is readily extensible to other molecular systems and can accommodate additional targets, constraints, or simulation engines, providing a general agentic-AI platform for transferable CG model development.