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

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