Interpretable Latent-Factor Machine Learning for Vapor Solubility in Amorphous Polymers

Machine-learning methods are applied to predict the Henry solubility of vapors in polymers through low-rank matrix-factor models. A systematic evaluation of the predictive capability of varying model ranks shows that rank-2 models have the best predictive power without the risks of instability or overfitting for the dataset in this work. The rank-2 model predicted the Henry solubilities within a multiplicative factor of 1.5 in 95% of cases for 43 vapors in nine polymers and at six temperatures. The vapors and polymers in this study span a range of chemical characteristics, including alkanes, olefins, oxygen-containing organic molecules, and halogenated hydrocarbons. Comparison of the latent factors against the polarizability, hydrogen bonding, and volumetric character of each vapor suggests that the two latent features learned by the model are the nonspecific dispersive character and specific polar/hydrogen-bonding character of the vapor/polymer. This model enables the determination of the solvation character of additional polymers by calibrating their feature vectors on as few as four chemically distinct vapors. The feature vector for the newly calibrated polymer can then be used to predict the solubility of the remaining untested vapors within this polymer, enabling a significant reduction in the number of experiments required to complete the solubility matrix and providing an avenue for estimating the solubilities of difficult-to-measure vapor-polymer interactions. Finally, the UNIFAC 2.0 model with free-volume correction is evaluated for predicting Henry solubilities of vapors in polymers, and its performance is compared to the low-rank model derived in this work.