Machine Learning Descriptors for CO2 Capture Materials

The influence of machine learning (ML) on scientific domains continues to grow, and the number of publications at the intersection of ML, CO2 capture, and material science is growing rapidly. Approaches for building ML models vary in both objectives and the methods through which materials are represented (i.e., featurised). Featurisation based on descriptors, being a crucial step in building ML models, is the focus of this review. Metal organic frameworks, ionic liquids, and other materials are discussed in this paper with a focus on the descriptors used in the representation of CO2-capturing materials. It is shown that operating conditions must be included in ML models in which multiple temperatures and/or pressures are used. Material descriptors can be used to differentiate the CO2 capture candidates through descriptors falling under the broad categories of charge and orbital, thermodynamic, structural, and chemical composition-based descriptors. Depending on the application, dataset, and ML model used, these descriptors carry varying degrees of importance in the predictions made. Design strategies can then be derived based on a selection of important features. Overall, this review predicts that ML will play an even greater role in future innovations in CO2 capture.