High-Throughput Screening of Hypothetical MOFs for Predicting Xenon Uptake Using Machine Learning Methods

Abstract

Xenon (Xe) gas adsorption in Metal-Organic Frameworks (MOFs) is a critical area for noble gas separation due to Xe's scarcity and high market value. Despite its importance, previous studies have largely overlooked the role of diverse Machine Learning (ML) models in predicting gas adsorption behavior under varying pressures. This study aims to fill this gap by developing a comprehensive database of hypothetical MOFs and applying advanced ML frameworks to predict Xe adsorption. Key structural descriptors—Void Fraction, Gravimetric Surface Area, Volumetric Surface Area, Pore Limiting Diameter, and Large Cavity Diameter—were integrated alongside adsorption pressure to enhance predictive accuracy. We trained and evaluated multiple ML models, including Ensemble Learning, Exponential Gaussian Process Regression, Fine Gaussian Support Vector Machines, and Bilayered Neural Networks, based on metrics such as RMSE (0.937 for EGPR), R² (0.83 for EGPR), and processing speed (up to 58,000 observations per second for FGSVM). Our screening identified four optimal MOFs—hMOF-30258, hMOF-30132, hMOF-5001015, and hMOF-30001—with superior Xe adsorption capabilities, featuring pcu and sql topologies that offer high surface area and porosity. These results highlight the potential of ML-driven approaches to revolutionize MOF design, paving the way for efficient noble gas separation technologies.