Machine learning guided high-throughput screening of dual-atom catalysts for oxygen reduction reaction

Dual-atom catalysts (DACs) can mitigate adsorption-energy scaling relations in the oxygen reduction reaction (ORR) through electronic coupling between adjacent metal sites. However, the complex correlation between geometric and electronic structures hinders rational design. Here, we integrate density functional theory (DFT), machine learning (ML), and experimental validation to screen 378 M 1 M 2 /NC candidates. A hybrid-sampled XGBoost classifier achieved an F1 score of 0.92 and an area under the receiver operating characteristic curve (AUC) of 0.97, reducing DFT evaluations for the remaining candidate space by approximately 80%. SHapley additive explanations (SHAP) identified the M 2 -N 6 bond length and the spin-up d-band centre of M 1 as key geometric and electronic descriptors, respectively. Guided by this principle, CoCu/NC exhibited excellent ORR activity (E 1/2 = 0.914 V) and stability, outperforming commercial Pt/C. Electronic-structure analysis further revealed spin-resolved O 2 -to-metal donation coupled with metal-to-O 2 π back-donation, which increased the occupation of O 2 antibonding states and weakened the O-O bond, providing a physically interpretable basis for catalyst discovery.