Physical–chemical coupling machine learning approach to exploring reactive solvents for absorption capture of carbonyl sulfide

Abstract

Most of machine learning (ML) models only explain one (chemical or physical) aspect of the reaction-involved absorption and therefore fail to predict solubilities of reactive species in chemical solvents. Herein, we propose a physical–chemical coupling ML approach to exploring absorption solvents for capturing a reactive organosulfide, COS. COS solubilities for 2,824 molecules were obtained by integrating physical absorption calculated using Henry's law and chemical absorption derived from reaction equilibrium calculation. ML model of ΔG of the reaction was established to examine the contributions of physical and chemical absorption. The coupling ML method was constructed by combining three absorption models. Experimental results of four commercial solvents verify that coupling ML method predicts COS solubilities in reactive and non-reactive solvents well. Furthermore, a descriptor-based molecule generation method was utilized to find 96 COS-preferred compounds. Present research highlights the coupling ML model for predicting absorption of COS and provides a general strategy for designing molecules/materials for intensified physical–chemical synergism.