Automated Feature Engineering and Model Aggregation for Data-Driven Oxidative Coupling of Methane Catalyst Design

The identification of active catalytic species and reaction mechanisms remains a challenging obstacle in catalyst design for the Oxidative Coupling of Methane (OCM), as reaction pathways are dependent on a catalyst's chemistry, structure, and reaction conditions, and characterization during operation is not feasible in most cases. Machine Learning (ML) has emerged as a recent addition in the catalyst design toolbox, since the construction of regression models able to predict catalytic activity as a function of catalyst composition and operating conditions facilitates the design of potential catalyst compositions. This work introduces the use of engineered compositional features reflecting both catalyst active metal and support information to aggregate multiple regression models to discover potential metal-support combinations with high OCM activity. As a result, three combinations are found to hold a <i>C</i>₂ yield greater than 20%: (Na, K, W)/CeO₂, (Cs, Ba, W)/TiO₂, and (Na, Cs, W)/SiO₂. These results show how an automated framework to generate and search for suitable catalyst features can discover active catalyst formulations within a large materials space.