Machine learning-guided design of multi-metal catalysts for methane dry reforming: Structure–activity insights and experimental validation

Machine learning (ML) offers a cost-effective and efficient strategy for designing dry reforming of methane (DRM) catalysts with high activity and stability. However, most reported models consider only a limited set of influencing factors—such as a single reaction condition or structural parameter—thereby restricting their predictive power and practical applicability. In this study, we developed an optimized and interpretable CatBoost model, systematically tuned via grid search with five-fold cross-validation, to predict CH4 conversion with high accuracy (R2 = 0.918, RMSE = 0.0718). The model incorporates a comprehensive set of descriptors, including catalyst composition, support type, preparation parameters, and reaction conditions. Pearson correlation and Sankey diagram analyses were used to reveal interdependencies among variables. SHAP analysis further identified the global importance of key parameters such as reaction temperature, gas hourly space velocity (GHSV), calcination temperature, and reduction temperature. In addition, bidirectional partial dependence visualization (PDV) clarified optimal operating ranges and uncovered synergistic effects between variables affecting catalytic performance. Based on these insights, we systematically explored the performance trends and structure–activity relationships of bimetallic catalysts in the DRM process and designed an optimized catalyst under targeted conditions. The reliability of the model was supported by both literature data and experimental validation. The resulting Ni–Ru/MgAl2O4 catalyst exhibited excellent CH4 conversion (90 %) and outstanding anti-coking performance at 850 °C.