Substructure-Enhanced MPNN for Polymer Discovery and Knowledge: A Study in Predicting Glass Transition Temperature

Understanding the relationship between polymer structures and the glass transition temperature (Tg) is crucial for the design of high-performance polymers. Machine learning models have shown great potential in accelerating the discovery and development of such materials by uncovering structure–property relationships. However, traditional machine learning models often overlook key structural features, such as functional groups, limiting their ability to effectively capture and represent complex structures of polymers. To address this, we developed the Substructure-Enhanced Message Passing Neural Network (SE-MPNN), which incorporates substructure information across multiple scales, from single atoms to functional groups, to evaluate their contributions to Tg. Results demonstrate that, by decomposing polymer structures into substructures, SE-MPNN provides more intuitive interpretations from the substructure level. These insights are consistent with established chemical understanding and offer valuable guidance for the rational design and optimization of polymer materials.