Interpretable Machine-Learning-Assisted Design of Metal Oxide Catalysts by Decoding Key Descriptors for CO₂-Assisted...

The rational design of high-performance catalysts for CO₂-assisted propane dehydrogenation (CO₂-PDH) is hindered by the complex interplay among catalyst properties, preparation parameters, and reaction conditions. Herein, this study develops an interpretable machine learning (ML) framework that accelerates the discovery of optimal catalysts and deciphers the underlying structure-performance relationships. A meticulously curated data set of 606 experimental data points was used to train and optimize seven ML models. The optimized XGBoost model demonstrated superior predictive accuracy (test R² = 0.966) for propane conversion, propylene selectivity, and yield. Shapley Additive exPlanations (SHAP) analysis quantitatively ranked the importance of 17 input features, revealing that catalyst descriptors (45.4%) and reaction conditions (37.2%) dominated the catalytic performance. Specific surface area, the primary support material, and time-on-stream were identified as the most critical descriptors governing the catalytic interface and activity. Partial dependence plots further elucidated the nonlinear influence of these key parameters on the target outputs. The framework was subsequently employed for multiobjective optimization, leading to the identification of Pareto-optimal catalysts (e.g., Cr₂O₃-Co/MSS-2-La₂O₃) achieving a remarkable propylene yield of 74.95%. Independent experimental validation on unseen catalysts confirmed the model's exceptional generalizability and accuracy, with prediction errors ≤ 4.34%. This work provides a robust, data-driven strategy for the rational design of high-performance catalytic materials by translating black-box predictions into actionable design principles.