A physically motivated Machine Learning model for accurate gas adsorption predictions in nanoporous materials

In this study we introduce a new set of descriptors to be employed by machine learning (ML) algorithms in order to accurately predict the gas adsorption capacities of Metal-Organic Frameworks (MOFs). The development of these descriptors is based on chemical intuition, specifically the realization that the ability of a porous material to adsorb gas depends not only on the total voids in the material’s framework but also on the distribution of these voids. This distribution is efficiently calculated by computing the statistical moments of the helium void fraction. The new approach requires almost no additional computational cost beyond that needed for calculating the standard structural features of nanomaterials. For the development and evaluation of this approach, as well as for comparison with existing methods, previously published computational data are used for the uptake capacities of rare gases and small linear and branched hydrocarbons in Topologically Based Crystal Constructor (ToBaCCo) MOFs under various thermodynamic conditions. Extensive analysis of all results reveals that, despite its simplicity, the new approach provides reliable predictions with the same or even higher accuracy compared to previous methods. ML models developed using the same approaches and additional theoretical data for the adsorption of ethane gas by computation-ready, experimental (CoRE) MOFs demonstrate lower accuracy predictions. Our analysis, aimed at clarifying this point, leads to useful conclusions about the factors that determine the accuracy of each approach and the features needed in the training data to develop predictive models for a wide range of nanoporous materials. • Physically inspired machine learning descriptors for gas uptake in MOFs • Void fraction moments capture structure with low computational cost • Accurate predictions across gases and thermodynamic conditions vs. GCMC data • Benchmarked on ToBaCCo and CoRE MOFs; generalizability differs