Reaction-Condition- and Ensemble-Aware Screening of Dual-Atom ORR Catalysts via Physics-Informed Machine Learning
Abstract Metal–nitrogen–carbon dual-atom catalysts (M1/M2–N–C DACs) have emerged as promising alternatives to Pt-based catalysts for the oxygen reduction reaction (ORR), yet their rational discovery is hindered by an enormous chemical and structural design space. Here, we develop a physics-informed machine learning (ML) framework integrated with high-throughput density functional theory (DFT) to systematically screen 22,599 M1/M2–N–C DAC structures spanning 729 transition-metal pairs and 31 structural configurations. Rather than relying on idealized bare-surface models, we construct voltage-dependent ab initio thermodynamic phase diagrams to identify realistic active-site structures and ORR limiting potentials under electrochemical operating conditions. To accelerate screening, we train equivariant transformer graph neural networks on DFT-generated adsorption energetics, achieving high predictive accuracy with mean absolute errors as low as 0.015 eV for adsorption energies and 0.022 V for derived ORR limiting potentials. We further introduce an experimentally relevant descriptor, the percentage of catalytically active structural configurations for each metal pair, which captures the ensemble nature of experimentally synthesized DACs and provides more reliable catalyst ranking than conventional single-configuration approaches. The framework identifies 3431 DAC structures with predicted ORR limiting potentials exceeding 0.8 V and reveals that only a small subset of metal pairs exhibits consistently high activity across configurations. Several top-performing candidates, including Co/Cr, Co/Ag, Co/Ru, Co/Ir, and Co/Zn, are validated by additional DFT calculations and show strong agreement with available experimental trends. In addition, stability screening uncovers multiple DACs predicted to possess both higher ORR activity and greater thermodynamic stability than benchmark Fe/Co–N–C catalysts. This work establishes a scalable ML-assisted paradigm for realistic electrocatalyst screening and provides design principles for the discovery of next-generation dual-atom ORR catalysts.