Elucidating the Physicochemical Basis of the Glass Transition Temperature in Linear Polyurethane Elastomers with Machine Learning

The glass transition temperature (<i>T</i>_g) is a fundamental property of polymers that strongly influences both mechanical and flow characteristics of the material. In many important polymers, configurational entropy of side chains is a dominant factor determining it. In contrast, the thermal transition in polyurethanes is thought to be determined by a combination of steric and electronic factors from the dispersed hard segments within the soft segment medium. Here, we present a machine learning model for the <i>T</i>_g in linear polyurethanes and aim to uncover the underlying physicochemical parameters that determine this. The model was trained on literature data from 43 industrially relevant combinations of polyols and isocyanates using descriptors derived from quantum chemistry, cheminformatics, and solution thermodynamics forming the feature space. Random forest and regularized regression were then compared to build a sparse linear model from six descriptors. Consistent with empirical understanding of polyurethane chemistry, this study indicates the characteristics of isocyanate monomers strongly determine the increase in <i>T</i>_g. Accurate predictions of <i>T</i>_g from the model are demonstrated, and the significance of the features is discussed. The results suggest that the tools of machine learning can provide both physical insights as well as accurate predictions of complex material properties.