Explainable ensemble learning reveals site-driven dz2 orbital occupancy tuning for enhanced hydrogen evolution on metal-doped Ni-loaded BN catalysts

Abstract

The activity of single-atom catalysts (SACs) is inherently limited by the intrinsic nature of the supported metal species, and achieving both high activity (the ideal hydrogen adsorption free energy ΔG_H*) and structural stability remains challenging. In this study, we integrated first-principles calculations with ensemble learning (DFT–EL) to conduct high-throughput screening and mechanistic investigations of metal-doped Ni@BN systems (Ni–TM2@BN). Compared with monometallic doping, bimetallic doping can significantly tune the H adsorption strength. Moreover, the H adsorption site is a critical factor governing ΔG_H*. On this basis, AdaBoost.R2 models were developed with ΔG_H1* and ΔG_H2* as the target properties, corresponding to H adsorption on the Ni site and the TM2 site, respectively. The optimal prediction models for ΔG_H1* (AdaBoost-GBR: R² = 0.943) and ΔG_H2* (AdaBoost-GBR: R² = 0.879) were obtained. Based on the DFT–EL framework, the optimal sites (OC1, OC2 and OC3) were screened out. Electronic-structure analysis revealed that bimetallic coupling induces d-orbital electron redistribution and drives the evolution of spin states and magnetic moments, thereby modulating the spin occupancy of the e_g orbitals (d_z² and d_x²−y²) and the TM–d/H–s hybridization strength. To further optimize the performance, axial O/S/P ligands were introduced to tune the e_g orbital electron occupancy at the Co site, thereby enhancing the HER catalytic activity (up to 5-fold). Finally, the SISSO algorithm reveals the dominant roles of the site motif (OC) and the intermetal distance d_Ni–TM2 in governing ΔG_H* (the highest R² is 0.957), providing guidance for the design and screening of bimetallic catalysts.