Machine Learning-Based Screening and Rationalization for Aromatic Guest Inclusion into MOF Crystals

Abstract The exploration of metal–organic framework (MOF) properties arising from guest inclusion, particularly for less-volatile molecules incorporated through solution processes, remains time-consuming and data-intensive compared with framework synthesis, leaving much of the existing MOF library unexplored. We report a machine-learning (ML)-based guest screening approach that predicts favorable guests from a library of ∼19,000 aromatic compounds, trained on 42 inclusion experiments for a single MOF. The ML model, constructed using relevant molecular features extracted from feature sets, predicts guest inclusion amounts across low to high values, with performance superior to models containing irrelevant features. The predictions successfully identify suitable guests that can be accommodated through soaking, enabling selective inclusion from mixed solutions and facilitating crystalline sponge analysis. In the absence of inclusion crystal structures, extensive structural search calculations using a neural network potential provide insights into guest alignment and noncovalent interactions, complementing the ML outcomes. This approach substantially reduces the experimental time and cost required to explore guest-inclusion-based MOF properties.