Large-Scale and Machine-Learning-Aided Investigations of Metal-Organic Frameworks for Water Harvesting

Water adsorption in metal-organic frameworks (MOFs) has recently drawn considerable attention for its tremendous potential in mitigating water scarcity. An important key to the development of such water harvesting technology to capture atmospheric water is the selection of optimal adsorbent materials. To date, tens of thousands of MOFs have been reported experimentally, while orders of magnitude more candidates have been theoretically predicted. Given the large materials space of available MOF adsorbents, computational studies, by employing state-of-the-art molecular simulations, play an important role in the selection of potential materials. In this study, a large-scale <i>in silico</i> screening, employing state-of-the-art Monte Carlo techniques, is conducted to explore more than 10,000 MOF candidates that are included in the CoRE (Computational-Ready Experimental) MOF database for their potential in water harvesting. While the widely used grand canonical Monte Carlo (GCMC) simulations to compute water adsorption can converge very slowly may therefore yield unreliable results, the large-scale computations performed herein employ a method, denoted as the flat-histogram-based C-map method reported by some of us [1,2], to determine the water adsorption capability and capacity of MOFs. Through extensively studying a diverse set of MOFs, our results identify promising candidates as well as shed light on the structure-property relationships. Moreover, to facilitate the future development of optimal water adsorbents, machine learning models are also developed. Specifically, tree-based methods such as random forest as well as convolution neural networks (CNNs) are employed. For the former, aside from including commonly used structural features such as largest cavity diameter and surface area, a newly developed metric – the so-called continuously adsorption channel (CAC) recently developed by some of us [3], is also used to help develop more accurate models. For the latter, computer vision techniques are exploited to “see” the structures directly for model training, followed by making quantitative predictions. These machine learning models are also investigated to better understand the structure-property relationship in a quantitative manner. Overall, this work represents a synergistic effort between large-scale molecular simulations and machine learning studies, and we anticipate the outcomes achieved herein can facilitate future computational and experimental efforts on the development of optimal water adsorbents.<br/><br/>References:<br/>[1] Datar, A.; Witman, M. & Lin, L.-C.* Improving Computational Assessment of Porous Materials for Water Adsorption Applications via Flat Histogram Methods, <i>J. Phys. Chem. C</i>, 125, 4253-4266, 2021.<br/>[2] Datar, A.; Witman, M.; Lin, L.-C.* Monte Carlo Simulations for Water Adsorption in Porous Materials: Best Practices and New Insights, <i>AIChE J.</i>, 67, e17447, 2021.<br/>[3] Xu, Z.; Wang, Y.; Lin, L.-C.*, Connectivity Analysis of Adsorption Sites in Metal-organic Frameworks for Facilitated Water Adsorption, <i>ACS Appl. Mater. Interfaces</i>, 15, 47081-47093, 2023.