arXiv cs.AIWednesday · May 27, 2026FREE

PolyFusionAgent: A Multimodal Foundation Model and Autonomous AI Assistant for Polymer Property Prediction and Inverse Design

Published on arXiv on May 27, 2026, PolyFusionAgent is presented as an interactive framework designed to address the challenges of polymer discovery, which include an expansive chemical design space and fragmented data representations. The system integrates two core components: PolyFusion, a multimodal polymer foundation model, and PolyAgent, a tool-augmented, literature-grounded design agent. PolyFusion aligns various polymer views, such as sequence, topology, 3D geometry, and fingerprints, across millions of polymers. This alignment creates a shared latent space that is transferable across different chemistries and data regimes, enhancing thermophysical property prediction and enabling the generation of chemically valid, structurally novel polymers based on desired properties. PolyAgent then closes the design loop by connecting these prediction and inverse design capabilities with evidence retrieval from the polymer literature. It proposes, evaluates, and contextualizes hypothetical polymer designs, aiming to bridge the gap between AI models and experimental reality to support directly actionable design decisions.