Quantifying Categorical Catalyst Descriptors for CO2 Methanation Design via CatBoost-Based Interpretable Machine Learning

Abstract

CO2 methanation is a key route for CO2 valorization and renewable hydrogen storage. Machine learning (ML) provides an analytical framework for understanding catalytic systems. However, conventional ML models usually use one-hot encoding, which splits intrinsically related catalytic components into independent sparse binary variables and limits the understanding of complex relationships among catalysts, reaction operating conditions, and CO2 conversion. In this study, an interpretable Categorical Boosting model using native categorical encoding (CatBoost-NCE) was applied to a dataset containing 4026 experimental entries. By preserving native categorical descriptors, the model enabled direct quantitative analysis of active metals, supports, promoters, and preparation methods, while maintaining reliable predictive performance across three independent data splits (mean test R2 = 0.918 ± 0.003). PDP and SHAP analyses showed that operating conditions were the dominant factors, while catalyst composition and preparation descriptors also made substantial contributions. Among categorical catalyst descriptors, the relative influence followed the order support > active metal ≈ preparation method > promoter. Beyond global feature interpretation, active metal-conditioned temperature analysis revealed distinct temperature response patterns among different catalyst systems. Metal-conditioned support substitution analysis further showed that support effects were closely coupled with active metal. For Ni-based catalysts, CeO2 exhibited a positive and directionally stable substitution effect. The model guided X-Ni–CeO2 catalysts showed good agreement with the predicted conversion trend, with an external validation R2 of 0.876. Characterization and in situ DRIFTS measurements further supported the interactions between active metal and support, and elucidated the evolution of reaction intermediates. ReaxFF molecular dynamics simulations performed on bare Ni clusters provided atomistic insights into plausible reaction events and pathways over intrinsic Ni active sites during CO2 methanation. This study converts accumulated catalyst data into actionable design rules and provides a reliable tool for the data-driven discovery of heterogeneous catalysts.