Toward Agentic Polymer Data Factories for Scalable Foundation Model Training

Polymer foundation models require datasets that are large, diverse, standardized, and physically meaningful, yet current polymer data remain sparse, fragmented, and difficult to curate. Simulation can help close this gap, but polymer systems relax slowly because of chain entanglement, packing frustration, glassy dynamics, and morphology-dependent behavior. Consequently, short high-throughput molecular dynamics simulations may yield structures and properties that are not representative of equilibrated ensembles. We propose agentic polymer data factories: AI-agent-driven workflows that automate polymer construction, force-field assignment, parameter refinement, multiscale simulation, convergence monitoring, property extraction, and database curation. AI agents assess equilibration, flag unreliable trajectories, trigger additional relaxation or coarse-graining, and
preserve provenance-rich metadata. By combining automated decision-making with physically grounded quality control, this framework provides a scalable route to reliable simulation-derived datasets for polymer foundation model training.