The removal of CO2 from gases is an important industrial process in the transition to a low-carbon economy. The use of selective physical (co-)solvents is especially perspective in cases when the amount of CO2 is large as it enables one to lower the energy requirements for solvent regeneration. However, only a few physical solvents have found industrial application and the design of new ones can pave the way to more efficient gas treatment techniques. Experimental screening of gas solubility is a labor-intensive process, and solubility modeling is a viable strategy to reduce the number of solvents subject to experimental measurements. In this paper, a chemoinformatics-based modeling workflow was applied to build a predictive model for the solubility of CO2 and four other industrially important gases (CO, CH4, H2, and N2). A dataset containing solubilities of gases in 280 solvents was collected from literature sources and supplemented with the new data for six solvents measured in the present study. A modeling workflow based on the usage of several state-of-the-art machine learning algorithms was applied to establish quantitative structure-solubility relationships. The best models were used to perform virtual screening of the industrially produced chemicals. It enabled the identification of compounds with high predicted CO2 solubility and selectivity toward other gases. The prediction for one of the compounds, 4-methylmorpholine, was confirmed experimentally.
Keywords: carbon dioxide; carbon monoxide; chemoinformatics; gas solubility; hydrogen; industrial gases; machine learning; methane; nitrogen.