Data-driven catalyst design for oxidative dehydrogenation of propane with CO2 using decision tree regression

• Published in: Applied catalysis. B, Environmental Environmental, 2025-02, Vol.361, p.124622, Article 124622
• Main Authors: Park, Jisu, Chung, Iljun, Jeong, Hyunjun, Lee, Dongmin, Yun, Yongju
• Format: Article
• Language: English
• Subjects: Carbon dioxide; Decision tree; Heterogeneous catalysts; Machine learning; Oxidative dehydrogenation; Propane
• ISSN: 0926-3373
• DOI: 10.1016/j.apcatb.2024.124622

Interpretable machine learning (ML) techniques have made significant progress in the design of high-performance materials. Here we employ decision tree (DT) models to discover and optimize multicomponent catalysts for the oxidative dehydrogenation of propane with CO2 (ODPC). While DT models have been used to improve a single target variable, this study leverages interpretable models by simultaneously considering the criteria for improving all the multiple target variables of ODPC, namely C3H8 conversion, CO2 conversion, and C3H6 yield. Instead of using literature data, we establish a consistent laboratory-scale database, and DT models are trained on multiple target variables representing catalytic performance. The trained DT models identify key input variables, including active metals, compositions, and support materials and provide criteria in the form of inequalities for the design of new ODPC catalysts. Despite using a database of binary and ternary metal oxides only, the fulfilment of all criteria suggests quaternary metal oxides with Cr, Ni, Mo, and ZrOx. The DT-proposed multicomponent catalysts show superior ODPC performance to the database catalysts, with a synergistic effect between the active elements. This exploitation of the relationship information between catalyst components and performance highlights the benefit of interpretable ML techniques in the design of high-performance multicomponent catalysts. [Display omitted] •DT models are used for the design of multicomponent ODPC catalysts.•A consistent laboratory-scale database has been established for model training.•Key input variables are identified and criteria are provided by trained DT models.•The catalysts suggested by the DT models outperform those in the database.