Machine Learning‐Driven Prediction and Optimization of Cu‐Based Catalysts for CO 2 Hydrogenation to Methanol

ABSTRACT China targets carbon peak by 2030 and neutrality by 2060. Catalytic CO 2 hydrogenation to methanol using renewable H 2 is a promising route for carbon recycling and emission reduction. However, developing efficient catalysts is hindered by complex structure–performance relationships. This study addresses this by constructing a dataset (112 samples and 28 features: composition, processing, and properties) and applying machine learning to predict Cu‐based catalyst performance. Six models—random forest, XGBoost, LightGBM, gradient boosting, SVR, and DNN—were compared for predicting CO 2 conversion, CH 3 OH selectivity, and yield. Optimal models varied as follows: SVR best predicted CO 2 conversion (test R 2 = 0.245), XGBoost excelled for CH 3 OH selectivity (test R 2 = 0.922 and MSE = 46.508), whereas CH 3 OH yield prediction was poor (LightGBM best and R 2 = 0.006). SHAP analysis revealed key nonlinear feature contributions (e.g., Cu content, GHSV, and temperature). Focusing on CH 3 OH selectivity, XGBoost was optimized as follows: multialgorithm voting and stepwise elimination identified 12 key features (e.g., Zn salt type, In content, and drying time). Bayesian hyperparameter tuning boosted performance (test R 2 = 0.9352). SHAP provided interpretability and design guidance. Bootstrap resampling validated reliability (95% CIs). An online prediction platform enhanced screening efficiency. Despite strong test performance, cross‐validation R 2 (0.7473) indicates a need for larger datasets. This work provides a robust data‐driven framework for optimizing CO 2 ‐to‐methanol catalysts, demonstrating ML's potential in catalysis research.