Data-driven and interpretable machine learning for performance-determining interactions governing C–C coupling and C2+ selectivity in Cu-catalyzed CO2 electroreduction

Abstract

Efficient electrochemical conversion of CO₂ into multi-carbon C₂⁺ products remains limited by the complex interplay between catalyst morphology, electrochemical environment, and reaction conditions. Here, we present an interpretable machine-learning framework to analyze C₂⁺ selectivity trends on Cu-based electrocatalysts using a literature-curated dataset of 380 experimental entries covering applied potential, morphology, particle size, support, electrolyte composition, and membrane type. Multiple regression models were benchmarked, with LightGBM giving the best performance for the C₂ : Product target, reaching R² ≈ 0.78. SHAP-based feature attribution and interaction analysis identified applied potential, particle size, and morphology as the dominant descriptors associated with C₂⁺ selectivity, while electrolyte and membrane-related variables provided secondary but non-negligible contributions. ANN-based response-surface projections and Pareto analysis further suggested a model-inferred C₂⁺-selective region near moderately negative potentials and sub-100 nm catalyst dimensions. Because the dataset is heterogeneous and literature-derived, these regions should be interpreted as data-supported trends rather than experimentally validated universal optima. Overall, this work converts fragmented CO₂RR literature data into an interpretable trend-analysis framework, providing hypothesis-guiding descriptor relationships for future catalyst and operating-condition design.