Machine Learning-Guided Design of Catalysts with SO2 Resistance for Low-Temperature NH3-SCR Reaction.

The design of efficient catalysts to mitigate SO2 poisoning in low-temperature Selective Catalytic Reduction of NOx (LT-SCR) is challenging. Herein, machine learning (ML) was employed to design catalysts with SO2 resistance. A multidimensional data set containing 242 data points was constructed, including elemental descriptors, catalyst structures, reaction conditions, and SO2 poisoning conditions. Regression models such as XGBoost (XGB) were trained to predict NOx conversions and SO2 resistance over different catalysts. It was found that the electronegativity descriptor (EN.) was the critical factor influencing NOx conversions in the presence of SO2, with approximately a monotonic positive correlation trend with NOx conversions within the range of 0.7-0.8. Based on these findings, quaternary CeMoFe/Ti catalysts were further synthesized via an inverse design. It was observed that this catalyst could still maintain 60% NOx conversion in the presence of SO2 after 6 h at 250 °C, wherein the SO2 resistance was significantly improved compared to ternary CeMo/Ti (∼40%) and binary Ce/Ti (∼4%) catalysts. In addition, the ML revealed the core roles of EN. in the design of SO2 resistance catalysts, breaking through the limitations of traditional ternary catalyst systems. This study provided a data-driven paradigm for precise design of SO2 resistance catalysts for the LT-SCR reaction, holding promise to accelerate the research and development of effective catalysts.