Interpretable Machine Learning Accelerated Discovery of Doped Spinel Co3O4 Catalysts for the Oxygen Evolution Reaction

Abstract The discovery of doped spinel Co3O4 electrocatalysts for the oxygen evolution reaction (OER) is hindered by the large catalyst design space and the high computational cost of density functional theory (DFT) calculations. Here, we establish a hierarchical interpretable machine learning (ML) framework that decouples electronic structure prediction from catalytic activity prediction. By starting with predicting the electronic descriptor (d-band center) from readily accessible structural descriptors and subsequently predicting adsorption energetics, the framework removes the need for expensive electronic structure DFT calculations during catalyst screening while preserving physical interpretability. A random forest model was trained on adsorption free energies of OH*, O*, and OOH* from 144 representative active sites and achieved a test mean absolute error of 0.21 eV. SHapley Additive exPlanations (SHAP) analysis identified surface energy, active site undercoordination, and the d-band center as the dominant descriptors governing adsorption energetics. To avoid additional DFT calculations of the d-band center during screening, a second ML model predicted it directly from structural and elemental descriptors with a test mean absolute error of 0.23 eV. The resulting framework enabled rapid evaluation of OER overpotentials across 252 active sites, while subsequent targeted DFT calculations identified Al-, Ca-, Fe-, and Mn-doped Co3O4 as promising candidates. Targeted DFT calculations and electrochemical measurements further validated the improved OER activity of Al- and Fe-doped Co3O4. This work establishes an interpretable and computationally efficient strategy for connecting structural descriptors, electronic properties, adsorption energetics, and experimentally measurable catalytic activity, thereby accelerating the discovery of high-performance spinel oxide catalysts.