Application of machine learning in designing and understanding effective catalyst for selective catalytic reduction of nitrogen oxide

Abstract

Traditional catalyst design includes trial-and-error and orthogonal methods. However, these processes usually require large number of experiments to get an optimized formula. Machine learning was applied in designing effective catalyst for selective catalytic reduction (SCR) of nitrogen oxide (NOx). Catalyst formulas and their activities in previous reports were collected and fitted by extreme gradient boosting algorithm and explanatory algorithm-SHapley Additive exPlanations. Mn-Cr coupling was predicted to be the most effective for SCR among various couplings, which was then proved by experimental results. SCR activity of Mn catalyst was increased from 50.3 % to 85.0 % at 150°C after the catalyst was loaded by Cr. Furthermore, machine learning and experimental characterizations revealed that the big total electronegativity of Cr resulted in bidentate nitrate bonding one cation with two oxygens, which was the most active NOx-derived intermediate during SCR. This work is in favor of catalyst design and catalytic-species recognition at the same time.