Development of Solubility Prediction Models with Ensemble Learning

The solubility parameter is widely used to select suitable solvents for polymers in the polymer-processing industry. In this study, we established a Hildebrand solubility parameter prediction model using ensemble-learning methods. The database used in the study is from the 2019 edition of the DIPPR 801 database, which includes solubility parameters for 1889 chemicals after removing invalid entries and outliers. Three machine-learning techniques including random forest, gradient boosting, and extreme gradient (XG) boosting were implemented to develop quantitative structure–property relationship analysis (QSPR) models. Subsequently, the ensemble method was applied to achieve higher accuracy. The coefficient of determination (R2) and root-mean-square error (RMSE) were calculated to validate that ensemble-learning models achieved satisfactory predictive capabilities with the overall R2 being 0.9793 and RMSE being 785.3313. Compared with determining the solubility parameter experimentally, the ensemble-learning models can perform a large-scale test within a few seconds. The models can be used to predict promising solvents for newly developed polymers at much lower time costs.