Interpretable machine learning framework for catalyst performance prediction and validation with dry reforming of methane

Abstract

Conventional methods for developing heterogeneous catalysts are inefficient in time and cost, often relying on trial-and-error. The integration of machine-learning (ML) in catalysis research using data can reduce computational costs and provide valuable insights. However, the lack of interpretability in black-box models hinders their acceptance among researchers. We propose an interpretable ML framework that enables a comprehensive understanding of the complex relationships between variables. Our framework incorporates tools such as Shapley additive explanations and partial dependence values for effective data preprocessing and result analysis. This framework increases the prediction accuracy of the model with improved R2 value of 0.96, while simultaneously expanding the catalyst component variety. Furthermore, for the case of dry reforming of methane, we tested the validity of the catalyst recommendation through dedicated experimental tests. The outstanding performance of the framework has the potential to expedite the rational design of catalysts.