Machine learning aided computational exploration of metal–organic frameworks with open Cu sites for the effective separation of hydrogen isotopes

Efficient separation of hydrogen isotopes is of vital importance to develop nuclear energy industry, while it remains a significant challenge to separate D2 from H2 due to their identical physicochemical properties. As one of the efficient alternatives to conventional techniques, the thermodynamic quantum sieving technology using metal–organic frameworks (MOFs) featuring open metal sites (OMSs) has shown a great potential. However, the lack of transferable force fields in conventional molecular simulations and high expense of brute-force screening hinder the quick discovery of MOFs targeted for D2/H2 separation. Herein, based on the established force field with high accuracy and transferability, machine learning and feature engineering are applied to address these challenges. Machine learning comprehensively assesses different descriptors that influence the separation performance of 929 experimentally-reported MOFs bearing Cu(II)-OMS. By employing the same metal nodes, new Cu MOF database (6,748 MOFs) is constructed, in which 45 hypothetical MOFs are firstly identified out through feature engineering that exhibiting high performance. Furthermore, grand canonical Monte Carlo simulations are performed on these MOFs, among which the optimal one exhibits comparable selectivity (36.9) and high adsorbent performance score (315.9) that surpasses the state-of-the-art materials do. This work not only presents a cost-effective approach firstly applying in the separation of hydrogen isotopes, but also provides experimental guidance for the design of high-performance adsorbents.