A Machine Learning Study of Polymer-Solvent Interactions

Polymer-solvent interaction is a fundamentally important concept routinely described by the Flory-Huggins interaction (chi), Hildebrand solubility(Delta delta) and the relative energy difference (RED) determined from Hansen solubility in experimental, theoretical and simulation studies. Here we performed a machine learning study based on a comprehensive and representative dataset covering the interaction pairs from 81 polymers and 1221 solvents. The regression models provide the coefficients of determination in the range of 0.86-0.94 and the classification models deliver the area under the receiver operating characteristic curve (AUCs) better than 0.93. These models were integrated into a newly developed software polySML-PSI. Important features including LogP, molar volume and dipole are identified, and their non-linear, nonmonotonic contributions to polymer-solvent interactions are presented. The widely known "like-dissolve-like" rule and two broadly used empirical equations to estimate chi as a function of temperature or Hansen solubility are also evaluated, and the polymer-specified constants are presented. This study provides a quantitative reference and a tool to understand and utilize the concept of polymer-solvent interactions.