Machine Learning for Gas Adsorption in Metal-Organic Frameworks: A Review on Predictive Descriptors

This review addresses a critical gap in the literature by focusing on the features (or descriptors) used in machine learning (ML) studies to predict gaseous adsorption properties in metal-organic frameworks (MOFs). Although ML approaches for predicting adsorption properties in MOFs have been extensively reported in recent years, features employed in ML models have not been thoroughly reviewed. A comprehensive review of these features is crucial since they form the foundation for building effective predictive models. These models are also key to facilitating the inverse design of MOFs, as they can be used to efficiently predict the performance of material candidates and explore the structure-property relationship, guiding the creation of optimal MOF structures. Furthermore, ML models can also be naturally employed in inverse design approaches, such as encoder-decoder architectures. This review starts with a brief overview of the importance and applications of MOFs in various fields, followed by a discussion of the historical milestones of MOFs in computational research, highlighting the critical role of ML. This review then discusses traditional features and introduces newly proposed distinctive features, referred to as "beyond traditional features", that have been reported to date. Furthermore, generalized ML models for predicting the adsorption properties of different gases are also outlined. Finally, we offer future outlooks on ML-assisted computational searches for MOFs in adsorption applications. Overall, this review aims to help researchers grasp current developments and offer insights into future directions in this area.