Metal-organic frameworks (MOFs) are promising candidate materials for various applications, including water harvesting, direct air capture, ion separation, and gas storage. However, their commercial application is unrealized in part due to low stability under mechanical stress. MOFs can undergo amorphization under external pressure, which alters the pore size and shape, thereby reducing performance. Hence, screening MOFs with high mechanical stability is essential to identify suitable candidate materials for experimental application. However, the diversity in metal-centric inorganic nodes, organic nodes, and organic edge chemistry makes the MOF design space in the order of millions, making it impossible to identify MOFs with high mechanical stability by performing experiments or computer simulations. Previous work has shown the potential of leveraging atomistic simulation and machine learning (ML) models for the accelerated discovery of mechanically stable MOFs. Still, the limited diversity in the MOF building blocks in the training dataset and categorical encoding of MOF topology to the ML models impede the generalizability of the models over a broader range of unseen MOF chemistries and topologies as well as the interpretability of structure-stability relationships. Here, we develop ML models for predicting mechanical stability in MOFs trained with novel interpretable topological features, resulting in much greater generalizability across the entire MOF design space and more exhaustive structure-stability relationships. We start with a dataset of hypothetical MOFs with over six times inorganic node diversity and ten times topological diversity than previous work. We then use the bulk moduli of the hypothetical MOFs reported in our earlier study as the descriptor of mechanical stability in our work. Unlike prior work, where only MOF geometric features and categorical topological features were used as inputs to the ML models, we develop our ML models with interpretable MOF chemical features, geometric features, and novel topological features that we develop in our work, making our ML models transferable to new unseen MOF chemistries and topologies. Using our ML models, we next identified functional groups from an extensive set with diverse chemistry to further improve mechanical stability in MOFs. Finally, we employ our ML models to virtually screen ~500k hypothetical and experimental MOFs to identify the most mechanically stable MOFs with potential for industrial applications.
