Machine Learning-Based Prediction of Catalytic Activity for CO₂ Hydrogenation Reactions Using Physicochemical Descriptors

The rational design of efficient heterogeneous catalysts for CO₂ hydrogenation remains a central challenge in catalysis research, traditionally demanding extensive and resource-intensive experimental campaigns. This study proposes a data-driven machine learning (ML) framework for the predictive modeling of catalytic activity and product selectivity in CO₂ hydrogenation reactions, utilizing physicochemical and structural descriptors extracted from published literature and open catalysis databases. A curated dataset comprising catalyst formulations — including active metals (Ni, Cu, Fe, Co, Ru), support materials (Al₂O₃, TiO₂, CeO₂, ZrO₂), and promoters along with corresponding reaction conditions (temperature, pressure, H₂/CO₂ ratio, space velocity) and performance outcomes (CO₂ conversion, CH₄ and methanol selectivity) is assembled and preprocessed. Feature engineering incorporates metal d-band center values, surface area, metal dispersion, electronegativity, and metal–support interaction descriptors. Multiple supervised learning algorithms including Random Forest, Gradient Boosting (XGBoost), and Artificial Neural Networks are trained, validated using k-fold cross-validation, and benchmarked against one another in terms of predictive accuracy (R², RMSE, MAE). Feature importance analysis via SHAP (SHapley Additive exPlanations) values is employed to elucidate the relative contributions of catalyst properties and process variables to model predictions. Preliminary results demonstrate that ensemble tree-based models achieve R² values exceeding 0.90 for CO₂ conversion prediction, with reaction temperature and metal dispersion emerging as dominant predictors. The proposed ML pipeline offers a computationally efficient, resource-accessible approach to catalyst screening and process monitoring, bridging the gap between big data analytics and heterogeneous catalysis for CO₂ valorization applications.