Accelerated screening of water-stable MOF structures using the digital reticular chemistry method

Abstract

Metal–organic frameworks (MOFs) are a class of porous materials renowned for their potential applications in areas such as gas separation and catalysis, while their practical utilization is often hampered by structural stability under varying humidity conditions. The emergence of digital reticular chemistry has revolutionized traditional MOF development methodologies, which are typically labor-intensive and costly. This study addresses the urgent need for innovative strategies to overcome these limitations by constructing a comprehensive dataset encompassing over 300 entries that detail the water stability characteristics of MOFs. To quantify material affinity for water, we introduced a computational descriptor, the Henry constant. Utilizing only three descriptors—RACs, SOAP, and the Henry constant—we trained machine learning (ML) models employing two classification strategies. These models exhibited robust predictive performance, achieving macro ROC–AUC values of 0.90 and 0.82 for the two- and three-class models, respectively. The generalizability of these models was further validated with newly collected data, and a detailed analysis of the three-class model yielded valuable chemical insights for the design of water-stable MOFs. Furthermore, the trained model was employed to predict the water stability of 100 MOFs, randomly selected from the CoRE database, and the reliability of the predictions was substantiated through the synthesis and stability testing of one predicted structure. The developed ML model facilitates the rapid screening of MOFs with targeted water stability, thereby streamlining the material selection and validation process and accelerating their application in specific domains.