Multitask Machine Learning to Predict Polymer–Solvent Miscibility Using Flory–Huggins Interaction Parameters

Predicting and understanding the phase equilibria or separation in polymer–solvent solutions represent unresolved fundamental problems polymer science. The behavior thermodynamics of miscibility depend on inter- intramolecular interactions a with certain molecular weight distribution mixed solvent. Here, we develop machine-learning framework to achieve highly generalized robust prediction Flory–Huggins χ parameters for solutions. model was trained using experimentally observed temperature-dependent 1190 samples, comprising 46 unique polymers 140 solvent species. However, difficulty that data set quantitatively limited qualitatively biased owing technical issues determining parameters. To overcome these limitations, produced an in-house obtained from quantum chemical calculations thousands pairs large list soluble insoluble pairs. Using three sets, conducted multitask machine learning simultaneously performed “soluble/insoluble” classification quantitative evaluation both experimental calculated Consequently, applicable wide range solution spaces. In this paper, predictive power physicochemical implications are demonstrated, along comparisons existing methods.