Machine learning-driven rapid screening and performance prediction of NH3-SCR catalyst

Traditional research and development of catalysts mainly rely on experimental trial-and-error methods, which struggle to meet the demand for efficient catalysts in the energy and environmental fields. Applying machine learning to catalyst design can achieve rapid screening of catalysts, thus overcoming the limitations of the traditional trial-and-error methods and solving the problems of low efficiency and high cost affecting catalyst research and development. This paper reports the development of a catalytic database containing 751 sets of data, with 30 input features divided into three categories, such as catalyst composition, preparation conditions, and catalytic reaction conditions, along with NO_x conversion and N₂ selectivity serving as output features. A machine learning approach was used to assist in the formulation screening and performance prediction of selective catalytic reduction (SCR) catalysts. The Categorical boosting (CatBoost) model was found to exhibit the best performance among the 10 models explored, with R², Mean absolute error (MAE), and Root mean square error (RMSE) values of 0.999, 0.006, and 0.008, respectively. The importance of each input feature for model prediction was obtained using the SHapley Additive exPlanations (SHAP) method, which showed that the reaction temperature and Mn elemental molar ratio had the greatest effect on NO_x conversion and N₂ selectivity. This study provides guidance for the screening, reaction condition optimization, and performance prediction of SCR catalysts. The successful integration of machine learning with experimental methods represents an effective strategy that can accelerate the discovery of catalysts.