Combining large-scale screening and machine learning to predict the metal-organic frameworks for organosulfurs removal from high-sour natural gas

Abstract

High-sour natural gas usually contains organosulfurs besides H2S, the majority of which exist in the form of mercaptans. These impurities of organosulfurs are required to be removed efficiently and economically for commercial application and the environment. In this work, the adsorption performance of organic sulfur gases [methanethiol (MeSH) and ethanethiol (EtSH)] in 137 953 hypothetical metal-organic frameworks (hMOFs) and 4764 computation-ready experimental MOFs (CoRE-MOFs) were evaluated by a high throughput computational screening technique. The highest adsorption capacities are predicted to be approximately 700 and 980 mg/g for MeSH and EtSH, respectively, which is substantially higher than that in zeolites (∼100 mg/g). Quantitative structure-performance relationships are established between adsorption capacities and MOF textural/energetic properties (including the largest cavity diameter, surface area, void fraction, and isosteric heat). Two machine learning techniques, the back propagation neural network (BPNN) and the partial least-square (PLS) methods, are applied to predict 4764 CoRE-MOFs after training all the data of hMOFs from the large-scale screening. Compared with PLS, BPNN shows better prediction accuracy for MeSH and EtSH, and finds that the isosteric heat among seven MOF features possesses the highest weight for the adsorption of organosulfurs. Finally, the best 8 MOFs are identified for the removal of gaseous organosulfurs from the high-sour natural gas in a variety of industrial situations.