Interpretable Machine Learning-Aided Discovery of Multicomponent Catalysts for Propane Dehydrogenation

Propylene serves as a crucial primary feedstock in the production of a wide range of chemical derivatives, including polypropylene, propylene oxide, and acrylonitrile. With the increasing demand for propylene, there is a growing interest in propane dehydrogenation (PDH), a selective reaction that produces propylene. In this study, we trained sure independence screening and sparsifying operator (SISSO) regression models to discover new catalysts exhibiting high performance for the PDH reaction. A database was established by compiling the performance of alumina-supported catalysts in PDH, obtained from our own experimental investigations. The SISSO model was trained using input variables that included the type, loading, and elemental properties of up to four active components, as well as reaction conditions, to predict propylene yield. The performance of the trained SISSO models was evaluated in order to select the optimal model, by comparing the predictive accuracy and simplicity of the resulting formulas. The formula derived from the optimal model was employed to predict the propylene yields of a total of 4,080 candidate catalysts, each consisting of three active components. Based on the predicted PDH performance, the catalysts were classified into categories of high-, moderate-, and low-performance. To experimentally validate these predictions, the actual performance of carefully chosen representative catalysts from each category was measured and compared with the predicted propylene yield. The superior performance of the catalysts categorized in the high-performance group, compared to those in the other groups, demonstrated the reliability of the developed SISSO model. Most importantly, the model identified highly efficient PDH catalysts that outperformed the catalysts tested during the construction of the database. The successful discovery of new and high-performing catalysts demonstrates the benefits of utilizing interpretable machine learning models in the development of multicomponent catalysts for heterogeneous catalysis.