Machine learning-based prediction for stability of hydrogen species on metal-doped CeO₂(111) surfaces

Tuning the stability of hydrogen species on metal-doped CeO_2 surfaces is a crucial aspect in the rational design of more efficient CeO_2-based catalysts. However, traditional theoretical and experimental approaches are limited in their capacity to explore the complex chemical space arising from doping and defect introduction. To address this limitation, we establish a comprehensive dataset using high-precision density functional theory (DFT) calculations for the CeO_2(111) surfaces doped with 43 different metals, encompassing three types of surface structures and five representative hydrogen adsorption configurations. A multi-level feature engineering strategy is then employed, combining elemental properties and multi-center smooth overlap of atomic positions (SOAP) features. The ensemble ExtraTrees regression model can achieve reliable prediction of adsorption energies (R^2 = 0.89, \rm RMSE= 0.55 eV), and its robustness is assessed through repeated cross-validation. Shapley additive explanations (SHAP) analysis shows that model predictions rely mainly on the principal component of local geometric features. Analysis across principal component intervals indicates that the SOAP-derived features differentiate coordination symmetry, anisotropy, and cooperative effects, and reveal their influences on adsorption energies. The results of principal component analysis (PCA) and cross-validation suggest that representing samples within a physicochemical feature space provides a more nuanced understanding of microenvironmental effects on hydrogen stability compared with the traditional configuration-based approaches. This study provides a data-driven framework for high-throughput screening and mechanistic exploration of structure-property relationships in complex multi-component oxide catalysts.