Machine learning and text mining approaches to design selective catalyst reduction synthesis routes

Abstract

The development of selective catalytic reduction (SCR) catalysts is often hindered by the complexity of experimental processes and the time-consuming trial-and-error approaches. Machine learning offers a promising solution by enabling more efficient and data-driven catalyst design. In this study, information was automatically extracted from the SCR-related scientific literature, including catalyst synthesis and catalyst properties, using rule-based techniques. These extracted data were then structured through feature engineering to build a machine learning-ready dataset. Models such as extreme gradient boosting regression (XGBR) and random forest (RF) were employed to predict catalyst performance and identify key factors influencing selectivity and conversion rates. To optimize synthesis routes, the designed synthesizable space was combined with the machine learning models to optimize key parameters and predict synthesis routes for SCR catalysts. Finally, synthesis information for SCR catalysts with high-performance was recommended. This work demonstrates the potential of using machine learning to accelerate SCR catalyst development, providing a scalable method for designing more efficient catalysts.