Alumina-Supported Catalyst Development for Propane Dehydrogenation via Interpretable Machine Learning and Experimental Validation

The direct propane dehydrogenation (PDH) reaction constitutes one of the key routes for the production of propylene and relies on the development of high-performance catalysts, which is generally achieved following a time-consuming trial-and-error strategy. In this study, a workflow of machine learning running five stages, i.e., data preparation and the development of a reliable machine learning model and its evaluation, interpretation, and application, was established to explore the data-driven research paradigm in the screening and design of catalysts for PDH with propylene yield as the target. Data from the literature on the PDH reaction catalyzed by alumina-supported catalysts were compiled. Twelve algorithms were evaluated, and the CatBoost model exhibits a high accuracy and generalization capability, with a coefficient of determination (R2) value of 0.992 for the training set and 0.973 for the test set. By employing this model, we screened two highly promising ternary catalysts. Experimental validation demonstrates that the predicted values for these two catalysts are in close agreement with the measured instantaneous propylene yields. Among the screened catalysts, PtSnZr/γ-Al2O3 exhibits a high propylene yield and maintains over 50% yield for 13.5 h. The instantaneous propylene yields on these catalysts are predicted to be further improved upon H2S pretreatment conditions. Explainable machine learning tools (Shapley additive explanations and partial dependence plot analysis) were employed to interpret the model. This study offers valuable insights into the application of machine learning in the heterogeneously catalyzed conversion of light alkanes and aids in the development of catalysts by uncovering a hidden structure–activity relationship in literature data.