Abstract
Understanding and predicting catalytic activity in transition-metal oxides remains challenging due to the multiscale interplay between geometric and electronic structure. Here, we propose a Hierarchical Embedded Sphere Model (HESM) integrating density functional theory (DFT) with interpretable machine learning (IML) to establish a generalizable descriptor framework for oxide electrocatalysis. By analyzing transition metal-doped anatase MO2 (101) surfaces, HESM disentangles catalytic activity into three hierarchical contributions: Global electronic structure (G-class), atomic-site intrinsic properties (A-class), and local coordination effect (L-class). The combined effect of these three classes leads to the emergence of two catalytic activation paradigms: Rh@MO(2) optimizes activity via dopant-induced electronic modulation (η(TD) = 0.29-0.36 V), while Fe@ZrO(2) enhances performance (η(TD) = 0.49 V) by host-site coordination field tuning. The hierarchical framework explains the competitive adsorption of OH(*) across doped/host sites, reconciles the well-known V-shaped activity trends with site-dependent deviations, and SHapley Additive exPlanation (SHAP) analysis identifies G-class as a critical feature that governs the model's prediction. This work demonstrates that HESM bridges adsorption energetics with multiscale geometric-electronic couplings, offering a generalizable methodology for descriptor discovery and catalyst design in complex systems.