Application of an Explainable Machine Learning to CO₂ Methanation for Optimal Design Nickel-Based Catalysts

The hydrogenation of carbon dioxide to produce clean fuels is one of the important means to achieve carbon neutrality, among which CO2 methanation has attracted much attention due to its thermodynamic advantages and environmental friendliness. However, the large-scale application of the technology still faces bottlenecks such as insufficient low-temperature activity of the catalyst and poor resistance to carbon deposition. The multiscale catalyst design method based on machine learning (ML) provides an innovative way to break through the limitations of traditional trial and error methods and achieve precise control of the energy conversion. In this work, reliable methods are first employed to remove the data with excessive outliers and missing values. Subsequently, a cross-validation method based on catalyst composition-stratified sampling is proposed to avoid overfitting. Finally, ensemble learning methods are utilized to construct ML models. The categorical boosting (CatBoost) model demonstrates superior performance with R2 values of 0.77 and 0.75 for CO2 conversion and CH4 selectivity, respectively. The results show that descriptors based on the catalyst performance and reaction conditions have a significant effect on the prediction of CO2 conversion and CH4 selectivity. By analyzing the most influential descriptors, the optimal reaction conditions were determined as T of 250–350 °C, gas hourly space velocity below 15,000 cm3 g–1 h–1, BET specific surface area between 50 and 200 m2 g–1, and CNi of more than 5%. This study provides new perspectives and methods for the development and optimization of CO2 methanation catalysts.