Interpretable Machine Learning Strategies for Accurate Prediction of Thermal Conductivity in Polymeric Systems

Abstract

Polymers, integral to advancements in high-tech fields, necessitate the study of their thermal conductivity (TC) to enhance material attributes and energy efficiency. The TC of polymers obtained by molecular dynamics (MD) calculations and experimental measurements is slow, and it is difficult to screen polymers with specific TC in a wide range. Existing machine learning (ML) techniques for determining polymer TC suffer from the problems of too large feature space and cannot guarantee very high accuracy. In this work, we leverage TCs from accessible datasets to decode the Simplified Molecular Input Line Entry System (SMILES) of polymers into ten features of distinct physical significance. A novel evaluation model for polymer TC is formulated, employing four ML strategies. The Gradient Boosting Decision Tree (GBDT)-based model, a focal point of our design, achieved a prediction accuracy of R$^2$=0.88 on a dataset containing 400 polymers. Furthermore, we used an interpretable ML approach to discover the significant contribution of quantitative estimate of drug-likeness and number of rotatable bonds features to TC, and analyzed the physical mechanisms involved. The ML method we developed provides a new idea for physical modeling of polymers, which is expected to be generalized and applied widely in constructing polymers with specific TCs and predicting all other properties of polymers.