Extrapolative prediction of polymer properties using physics-informed hierarchical descriptors

Accurately predicting the properties of polymers is essential for data-driven materials design. However, such predictions are often challenged by the limited availability of polymer-related data and the fact that high-performance polymers of interest typically lie outside the distribution of existing datasets. In this study, we develop a machine learning model that enhances extrapolative prediction accuracy beyond the training data by leveraging hierarchical, physics-informed descriptors. Specifically, we utilize quantum mechanical (QM) descriptors derived from density functional theory calculations, molecular dynamics (MD) descriptors representing structural and dynamical properties, and force field (FF) descriptors characterizing the interaction parameters used in MD simulations. We investigated two types of extrapolation tasks: extrapolation beyond the range of physical properties and extrapolation to structurally dissimilar molecules. By systematically evaluating all non-zero combinations of QM, MD, and FF descriptors, we find that selected subsets often outperform models using the full descriptor set. This highlights the critical role of dimensionality reduction and descriptor relevance, especially under data-scarce conditions. Comparisons with structure-based models employing molecular fingerprints or molecular graphs further demonstrated the superiority of the proposed model based on selected physics-based descriptors.