(366b) Active Learning Strategies for Predicting Gas Adsorption and Separations in Metal-Organic Frameworks

Metal-organic frameworks (MOFs) represent an expansive class of porous, crystalline materials, offering extraordinary tunability and potential for gas storage and separation. However, the sheer size and chemical diversity of the MOF design space make systematic exploration and identification of optimal materials a daunting computational challenge. Traditional simulation methods, such as grand canonical Monte Carlo (GCMC) for adsorption prediction, while accurate, are prohibitively computationally expensive when applied exhaustively. Therefore, intelligent data-driven methods are critically needed to accelerate the discovery and optimization of MOFs for gas storage and separation applications.

To address this challenge, we first introduce a comprehensive multi-method framework that integrates inducing points (IPs), active learning (AL), and Bayesian optimization (BO) to systematically identify informative MOFs for adsorption modeling. Specifically, we employ Gaussian process regression (GPR) combined with distinct AL acquisition functions, including Gaussian process standard deviation (GP STD), expected improvement (EI), and probability of improvement (PI), to strategically select MOFs for methane (CH₄) adsorption prediction. These data acquisition strategies target regions of highest model uncertainty or greatest potential performance improvement, selecting a subset of MOFs occupying various regions in the textural space. By intersecting MOF sets identified by these strategies, we isolate a consensus set of 611 representative MOFs. Training a GPR model using this refined dataset yields a robust predictive model.

In the second component of our study, we explore the critical interplay between adsorption and diffusion phenomena within a MOF (CuBTC), integrating both mechanisms into an end-to-end (E2E) active learning framework for enhanced selectivity predictions across wide-ranging operational conditions. Evaluating MOFs for gas separations requires understanding mixture adsorption and diffusion data, traditionally demanding extensive computational resources through combined GCMC and molecular dynamics (MD) simulations. Our approach employs GPR models trained iteratively using AL strategies tailored to different sources of uncertainty: adsorption uncertainty, diffusion uncertainty, or combined selectivity uncertainty through propagation techniques. Our results demonstrate that diffusion-centric AL strategies yield superior model efficiency and predictive accuracy, effectively minimizing redundant simulations and substantially enhancing gas selectivity predictions in CuBTC.