Machine Learning-Guided High-Throughput Screening of Metal–Organic Frameworks for Atmospheric Ozone Capture: A Computational Materials Informatics Approach

Abstract

Computational materials informatics has emerged as a powerful approach for accelerating functional materials discovery through high-throughput screening and machine learning. This study presents the first systematic high-throughput computational screening and machine learning analysis of metal-organic frameworks (MOFs) for atmospheric ozone capture under environmentally relevant conditions (295 K, 250 ppb O₃), establishing a dual-mechanism design framework with broad implications for reactive gas separations. A data set of 2116 representative MOF structures from the CoRE database was evaluated using grand canonical Monte Carlo (GCMC) simulations, with molecular dynamics characterizing transport properties and hierarchical machine learning enabling performance prediction across 5 orders of magnitude (2.58 × 10^-7-1.50 × 10^-2 mol/kg). The key conceptual advance is a dual-mechanism framework governing trace-level capture: chemical affinity mediated by heavy metal centers determines ppb-level uptake efficiency, while geometric packing optimization controls maximum storage capacity. Results revealed that only 14.6% of MOFs exhibit detectable ozone adsorption, with top-performing materials featuring heavy metal centers (Mn, Mo, U, and Cd) achieving capacities exceeding 1.0 × 10^-2 mol/kg. Machine learning analysis achieved exceptional predictive performance (R² = 0.852) and identified pore efficiency as the universal predictor for MOF activity. The validated computational framework enables rational design of next-generation adsorbents for atmospheric remediation and demonstrates the effectiveness of hierarchical machine learning for complex environmental separations. This work establishes fundamental design principles applicable to reactive gas capture and accelerates materials discovery through interpretable computational screening.