Machine Learning the Hydrogen Adsorption Capacity of Metal Organic Frameworks

High-throughput virtual screening and machine learning (ML) are powerful tools for accelerating the discovery of nanoporous adsorbents for gas storage applications, including metal-organic frameworks (MOFs). Besides the nature of the data and the models, ML performance is strongly dependent on the MOF representation. In this work, we extended the molecular atom-atom, bond-bond, bond-atom (AABBA) autocorrelations to generate vector representations of MOFs for ML models. AABBA vectors encoded both the atomic and bond properties of the MOF metal nodes and linkers, enhancing accuracy in the prediction of gas storage properties. In particular, we computed the gravimetric and volumetric adsorption capacities for the adsorption of hydrogen into 11,500 MOFs. The data, which is made openly available together with the ML models code, was computed with Grand Canonical Monte Carlo simulations to train neural networks predicting the isothermal hydrogen deliverable capacity at room temperature. Further, feature engineering based on a gradient boosting algorithm highlighted the benefits of reducing the dimensionality of the AABBA representation, while keeping the autocorrelation of the bond properties.