Comparison of Energy-Based Machine Learning Descriptors for Gas Adsorption

Metal–organic frameworks (MOFs) are a class of nanoporous materials that hold great promise for applications involving adsorption. The structural and chemical tunability of MOFs has led to the realization of an enormous number of materials through experimental or in silico techniques. Although a large candidate pool is desirable, the great number of possible solutions renders its exploration nontrivial. With the advent of machine learning (ML), the identification of promising materials for specific applications can be performed in a matter of seconds with the proviso that ML models are accurate and their required input is (computationally) cheap. With regard to gas adsorption in MOFs, energy-based descriptors can significantly improve the performance of ML models compared to structural descriptors, especially when non-negligible host–guest interactions are present. In this work, we investigate the impact of Dprobes and Henry coefficients, two energy-based descriptors, on the performance of ML models, examining CO2, H2S, and H2 under different adsorption conditions. The ML models built with Dprobes perform in general better, and their approximately 2 orders of magnitude faster calculation (compared to Henry coefficients) enables efficient large-scale screening.