Interpretable Ensemble Learning Reveals the d z 2 -Orbital-Regulated H Adsorption Mechanism of Metal-Doped Ni/BN Catalysts

Abstract

The activity of single-atom catalysts (SACs) is limited by the intrinsic properties of the supported metal elements, and it is difficult to simultaneously achieve high catalytic activity and structural stability. Due to the wide variety of doped metal species and the fact that their performance is influenced by multiple factors, traditional experimental methods and high-throughput calculations face problems such as high cost and long time consumption. In this study, we constructed a training dataset using intrinsic features, site features, and bimetallic difference features as model inputs, with H adsorption on the TM2 metal site as the output. This dataset was successfully applied to small-sample ensemble models (AdaBoost-RFR, AdaBoost-XGB and AdaBoost-GBR) to predict the hydrogen evolution performance of Ni–TM2@BN structures. The results showed that more than 30 candidate structures with excellent performance were identified, including Mn@Ni-BN and Fe@Ni-BN. To further enhance the catalytic performance, O/S/P axial ligands were introduced to regulate the eg orbital electron occupancy of the Co site, thereby improving HER activity by up to approximately fivefold. This study demonstrates that small-sample ensemble learning can still efficiently screen potential catalytic materials under data-scarce conditions, and proposes regulation strategies to further enhance catalytic performance.