Accelerating Discovery of Mechanically Stable Metal-Organic Frameworks for Vinylidene Fluoride Storage by Active Learning.

Abstract

Metal-organic frameworks (MOFs) are versatile nanoporous materials for a wide variety of important applications. Recently, a handful of MOFs have been explored for the storage of toxic fluorinated gases (Keasler et al. Science, 2023, 381, 1455), yet the potential of a great number of MOFs for such an environmentally sustainable application has not been thoroughly investigated. In this work, we apply active learning (AL) to accelerate the discovery of hypothetical MOFs (hMOFs) that can efficiently store a specific fluorinated gas, namely, vinylidene fluoride (VDF). First, a force field was developed for VDF and utilized to predict the working capacities (ΔN) of VDF in an initial data set of 4502 MOFs from the computation-ready experimental MOF (CoRE-MOF) database that successfully underwent featurization and grand-canonical Monte Carlo simulations. Next, the initial data set was diversified by Greedy sampling in an unexplored sample space of 119,387 hMOFs from the ab initio REPEAT charge MOF (ARC-MOF) database. A budget of 10,000 samples (i.e., <10% of total ARC-MOFs) was selected to train a random forest model. Then, ΔN in the unlabeled ARC-MOFs were predicted and top-performing ones were validated by simulations. Integrating with the stability requirement, mechanically stable ARC-MOFs were finally identified, along with high ΔN. Furthermore, by Pareto-Frontier analysis, we revealed that long linear linkers can enhance ΔN, while bulkier multiphenyl linkers or interpenetrated frameworks improve mechanical strength. From this work, we efficiently discover top-performing MOFs for VDF storage by AL and also demonstrate the importance of integrating stability to identify stable promising MOFs for a practical application.