Multi-method material selection for adsorption using Bayesian approaches

Machine learning is reshaping the discovery of new materials, yet a persistent challenge is selecting the most informative training data from large and complex databases. Here we present a framework that integrates inducing points with diverse data acquisition strategies from active learning and Bayesian optimization to guide the selection of material training sets. We compare purely explorative selection, based on Gaussian process regression uncertainty, with exploitation-based approaches such as expected improvement and probability of improvement. Using methane uptake in metal–organic frameworks as a case study, we evaluate structures by properties such as void fraction, pore diameters, and accessible surface area. An intersection analysis across methods identifies a consensus set of 611 frameworks and key pressure points consistently chosen as informative. Training on this reduced subset yields a highly accurate predictive model, demonstrating that principled data selection accelerates adsorption modeling and enables more efficient materials screening. A major challenge in machine learning is selecting the most appropriate training data from large databases. Here, a framework that integrates inducing points with active learning and Bayesian optimization for selecting training sets is reported, and used to identify metal-organic frameworks for methane uptake.