Heat diffusion coefficient study of polymers based on interpretable machine learning

Abstract. Polymers hold significant application value across various fields of modern society, with different application scenarios requiring specific thermal diffusivity coefficients. Finding polymer materials with targeted thermal diffusivities is crucial. However, due to the vast variety and complex structures of polymers, constructing a unified structured dataset for machine learning modeling is challenging. Although machine learning has shown great potential in materials science, it has rarely been applied to predict the heat diffusion coefficient of polymers. This paper constructs a dataset for predicting the thermal diffusion coefficient of polymers using a publicly available dataset by transforming the SMILES code of polymers into eight features with practical physical and chemical meanings. Using the Random Forest algorithm, training with 400 of these data and randomly selecting 200 of them for cross-validation, the accuracy of the test set reached 0.9. Additionally, through interpretability analysis, we found that the molecular weight of the polymer monomers, the TPSA (the polar surface area of the molecule), and the NRB (the number of rotatable bonds) are the main features affecting the polymer thermal diffusion coefficient. An increase in the TPSA and the NRB positively contributes to the thermal diffusivity, while an increase in molecular weight negatively contributes. Our study provides a new method for the prediction of polymer thermal diffusivity and creates a new paradigm for the study of polymer thermal diffusivity, promoting further development in this field.