Machine learning assisted high-throughput computational screening of MOFs for the capture of chemical warfare agents from the air.

[Display omitted] • Structure-property relationships of MOFs were established with the capture of CWAs. • The number of hydrogen bond acceptors was a key factor governing the co-adsorption of CWAs by machine learning. • The similarities of top-performing MOFs were defined for the metals, organic links and topology. To effectively capture the low-concentration chemical warfare agents (CWAs) and their simulants which are extremely harmful to human health and environment, the properties of thousands of Computation-Ready, Experimental Metal-Organic Frameworks (CoRE-MOFs) for the adsorption and separation of four CWAs and simulants (dimethyl methyl phosphonate, soman, mustard gas, and 2-chloroethyl ethyl sulfide) from the air were calculated by high-throughput computational screening. To reasonably identify the top-performing MOFs, the trade-off between selectivity and adsorption capacity (TSN) was introduced to measure the properties of MOFs. Five machine learning algorithms were employed to quantitatively evaluate the structure-performance relationships of MOFs for the adsorption of CWAs and validate that Extreme Gradient Boosting algorithms had the best prediction accuracy. Furthermore, four MOF descriptors (henry coefficient, number of hydrogen bonds, porosity, and volumetric surface area) were found to have significant influence on the properties of MOFs. Finally, it was determined that the number of hydrogen bond acceptors was a key factor governing the co-adsorption of CWAs and their simulants, and the similarities of adsorbents with good adsorption performance included Zn for metal center, trimesic acid for organic linker, and srs for topology. The microscopic insights obtained from our bottom-up approach are very helpful for the development of MOFs and other nanoporous materials for the capture of CWAs from the air. [ABSTRACT FROM AUTHOR]