Fragment-based representation learning for polymers and heterogeneous formulations

Real polymer systems include not only simple linear homopolymers, but also copolymers, branched architectures, crosslinked networks, and multicomponent formulations. Yet most polymer machine-learning models still rely on idealized repeat-unit representations. This mismatch limits not only the scope of polymer informatics, but also the development of foundation models for polymer systems. Here we investigate chemically meaningful fragments as a shared representation across small molecules, linear homopolymers, and component-defined polymer formulations. We develop CGMNet (Coarse-Grained Molecular Network), a framework that learns fragment-level representations from unlabeled small molecules through self-supervised pretraining and transfers them to downstream polymer property prediction. Using this framework, we show that fragment-based representations support strong predictive performance for small molecules, linear homopolymers, and component-defined polyester, epoxy-resin, and vitrimer formulations, while retaining interpretability through attribution to specific chemical motifs. Together, these results establish fragments as a practical representation interface for bridging small-molecule pretraining, linear-homopolymer benchmarks, and component-defined polymer formulations beyond a single-repeat-unit input.