Prediction of organic sulfur solubility in mixed solvent using feature-based transfer learning and a hybrid Henry's law constant calculation method

Machine learning (ML) algorithms are playing increasingly important roles in exploring solvents for wide industrial applications. However, most ML strategies for solvent screening neglect the contributions of intermolecular interactions among solvent components, resulting in reduced prediction accuracy for the solubilities of solvent mixtures. In this study, we propose an efficient method combining feature-based transfer learning and a hybrid Henry's law constant (HLC) calculation method to assist the exploration of promising solvent mixtures to remove organic sulfides. The incorporation of predicted HLC values from established models as features significantly enhances the prediction accuracy for various organic sulfides. In the case of 2-propanethiol, the prediction shows a R_test² of 0.91, RMSE of 0.0166, and MAE of 0.0118. The hybrid HLC calculation method, which incorporates non-ideal interactions between two solvent components, outperforms both the conductor-like screening models for real solvents (COSMO-RS) and ideal solution methods in predicting experimental HLC values. The present method successfully predicts a hybrid solvent for methanethiol (MeSH) removal. Both static and dynamic absorption experiments confirm that this designed solvent mixture has the lowest HLC of 370.48 kPa and the highest removal rate of 80.38% for MeSH.