Adsorption-based water harvesting processes using metal-organic frameworks (MOFs) have drawn considerable attention in recent years for their potential in mitigating water scarcity. A key to the success of such processes is the selection of optimal water adsorbents. To date, tens of thousands of distinct MOFs have been reported experimentally and therefore, employing computational studies, particularly molecular simulations and machine-learning-aided investigations, can play a critical role in exploring such a large materials space. Herein, a large-scale Monte Carlo screening is conducted to study 10,000 experimentally reported MOFs for their adsorption properties. Through such large-scale study, our results identify promising candidates as well as shed light on the structure-property relationships. Hundreds of MOFs are further strategically selected with detailed analysis to shed light on the formation of “S-shaped” isotherms, characterized by sudden condensation within a small pressure window. Aside from the adsorption aspect, large-scale molecular dynamics simulations are also conducted to investigate the diffusion properties of water for MOF structures, identifying those of fast diffusion as well as unravel key underlying mechanisms. Moreover, to facilitate the search for potential candidates, machine learning models are as well developed to predict adsorption and diffusion properties. Specifically, convolution neural networks (CNNs) are employed to directly “see” the structures for model training, followed by making quantitative predictions. Traditional tree-based models are also considered for comparison. Overall, this work represents large-scale computational exploration of MOFs for water harvesting, and we anticipate the outcomes achieved herein can facilitate future computational and experimental efforts on the development of optimal water adsorbents.
