Accelerating the discovery of acid gas-selective MOFs for natural gas purification: A combination of machine learning and molecular fingerprint

Abstract

The rational construction of high-performing MOF adsorbents for H2S/CO2/CH4 separation is essential for natural gas (NG) purification. However, trial-and-error experiments usually face significant challenges associated with unclear design routes. Especially the chemical design, which is decisive in geometrical properties and adsorption sites, thus playing substantial roles in molecule separation, has not been comprehensively understood yet. In this work, a combination of machine learning (ML)-assisted computational screening protocol based on ∼ 1,000 simulation data and molecular fingerprint was performed to accelerate the discovery of promising MOFs for H2S/CO2/CH4 separation. The reliability of the adopted ML algorithm on our proposed comprehensive separation metric (TNSR) was further verified by another 50 MOFs with a small absolute mean error (AME) of 2.95. Encouragingly, 463 out of over 10,000 unpredicted MOFs were rapidly identified using the developed ML-assisted screening protocol. Among them, five additional top-performing MOFs (TNSR > 65) were successfully located, with the best one showing a TNSR of 82.39, 5–9 times higher than some well-known adsorbents (i.e. CuBTC, ZIF-69) at the same condition. During establishing the screening protocol, the chemical characteristics particular to the high-performing MOFs and the different effects of aromatic and alkyl ligands on separation were understood in detail. Different from the previous work, it was found that MOFs with transitional metals can also be unfavorable, mainly due to the specific metal coordination formed under various conditions. This work emphasizes the importance of chemical design and aims to provide an accurate screening strategy to discover promising MOFs for natural gas purification.