Mapping the Porous and Chemical Structure–Function Relationships of Trace CH₃I Capture by Metal–Organic Frameworks using Machine Learning

Large-scale computational screening has become an indispensable tool for functional materials discovery. It, however, remains a challenge to adequately interrogate the large amount of data generated by a screening study. Here, we computationally screened 1087 metal–organic frameworks (MOFs), from the CoRE MOF 2014 database, for capturing trace amounts (300 ppmv) of methyl iodide (CH3I); as a primary representative of organic iodides, CH3129I is one of the most difficult radioactive contaminants to separate. Furthermore, we demonstrate a simple and general approach for mapping and interrogating the high-dimensional structure–function data obtained by high-throughput screening; this involves learning two-dimensional embeddings of the high-dimensional data by applying unsupervised learning to encoded structural and chemical features of MOFs. The resulting various porous and chemical structure–function maps are human-interpretable, revealing not only top-performing MOFs but also complex structure–function correlations that are hidden when inspecting individual MOF features. These maps also alleviate the need of laborious visual inspection of a large number of MOFs by clustering similar MOFs, per the encoding features, into defined regions on the map. We also show that these structure–function maps are amenable to supervised classification of the performances of MOFs for trace CH3I capture. We further show that the machine-learning models trained on the 1087 CoRE MOFs can be used to predict an unseen set of 250 MOFs randomly selected from a different MOF database, achieving high prediction accuracies.