Machine-Learning-Based Prediction of Methane Adsorption Isotherms at Varied Temperatures for Experimental Adsorbents

ABSTRACT

Metal−organic frameworks (MOFs) are crystalline materials andone of the optimal materials for large-scale grand canonical Monte Carlo(GCMC) simulations. Recently, there have been trials for applying machinelearning (ML) to the results of large-scale GCMC simulations to predict gasadsorption on MOFs. However, the functions of the developed algorithms arenot different from those of GCMC simulations, in that they provide a predictionof adsorption properties based on the coordination structures. In this study, wepropose a novel Monte Carlo-Machine Learning (MC-ML) strategy, whichcombines ML with GCMC to provide the function that is distinct from that ofGCMC. To verify the concept of the strategy, we designed an algorithm topredict methane isotherms at a range of temperatures from a methane isothermat a temperature of 298 K. GCMC simulations functioned as a data-producingtool for ML, which yielded adsorption properties of 4951 structures in the CoRE-MOF database. The ML was applied to the GCMCresults using experimentally measurable properties as features. Finally, the algorithm developed from ML was evaluated usingexperimental methane adsorption data for defective MOFs, MOFs with open metal sites, and non-MOF materials, which revealedthe merits of the MC-ML strategy in comparison with typical GCMC.

논문 정보

DOI : https://doi.org/10.1021/acs.jpcc.0c01757