Harmonizing Adsorption and Diffusion in Active Learning Campaigns of Gas Separations in a MOF

This study establishes an active learning (AL) framework designed to enhance the prediction of selectivity in the metal–organic framework (MOF) CuBTC by addressing the adsorption and diffusion mechanisms essential for gas separation applications. Traditional methods of predicting gas selectivity across broad pressure, temperature, and composition (PTX) conditions face considerable computational demands, particularly when simulating both adsorption and diffusion in porous materials. We tackled this challenge by integrating Gaussian Process (GP) models with grand canonical Monte Carlo (GCMC) and molecular dynamics (MD) simulations, applying AL to iteratively select data points exclusively based on the highest model uncertainty (emphasizing exploration), and others schemes, to systematically improve prediction accuracy across diverse conditions. We evaluated several AL strategies tailored to two distinct molecular-level phenomena involved in gas separations: adsorption and diffusion. Specifically, adsorption-based selectivity prediction methods employed AL to preferentially select data points characterized by high uncertainty in adsorption predictions. In contrast, diffusion-based selectivity prediction methods directed AL toward regions exhibiting high uncertainty in diffusion predictions. These approaches included direct prediction of selectivity as well as propagation-based methods, where selectivity uncertainty was calculated from individual component uncertainties in adsorption or diffusion predictions. The culmination of this exploration is an end-to-end (E2E) framework that integrates adsorption and diffusion modeling within a single AL-driven data-labeling process. In this framework, uncertainty in either adsorption or diffusion predictions guides data selection, enabling more precise model training across adsorption and diffusion. Results indicate that the diffusion-focused E2E scheme yields the highest predictive accuracy and a more efficient process. This approach minimizes redundant sampling and improves the efficiency of data acquisition across adsorption and diffusion models.