(611j) Accelerating the Characterization and Design of Nanoporous Materials with Data-Driven Models

Metal-organic frameworks (MOFs) are a class of nanoporous materials with highly tunable pore shape and chemistry. They are synthesized via self-assembly of metal nodes connected by organic linkers to form crystalline porous materials which have been explored for diverse applications including gas storage and catalysis. High-throughput molecular simulations can identify promising materials out of thousands of candidate MOFs, but the interrogation of this vast amount of generated data has been only rudimentary in the literature. We want to use more sophisticated techniques to glean more insights from this data. In this work, we applied machine learning methods to accelerate high-throughput computational screening of MOFs and to develop better structure-property relationships. Our approach uses least absolute shrinkage and selection operator (LASSO) regression to extract insights from MOF energy landscapes, which are rapidly calculated using a probe molecule. The model is highly accurate, interpretable, robust, and three orders of magnitude faster than detailed molecular simulations. To demonstrate the usefulness of this model, we screened 55,000 MOFs and identified a candidate MOF with high hydrogen storage capacity, which we also confirmed experimentally. The approach can be easily applied to other classes of porous materials and adsorbate molecules such as methane.