A Machine Learning Approach for Prediction of Faradaic Efficiency in Electrochemical CO2 Reduction on Nitrogen-Doped Carbon.

Nitrogen-doped carbon materials are promising catalysts for electrochemical CO₂ reduction, yet achieving high Faradaic efficiency for CO production remains challenging due to the competing hydrogen evolution reaction . To accelerate catalyst design, a machine learning-based stacked model is developed, integrating random forest and XGBoost (XGB) as base models with linear regression as a meta-model. This approach mitigates overfitting, achieving superior predictive performance (R² = 0.98 train, 0.91 test) compared to XGB alone (R² = 0.99 train, 0.86 test). SHapley Additive exPlanations (SHAP) analysis identifies pyridinic nitrogen (N) as a key driver of CO selectivity but reveals that its influence varies with different carbon substrates. SHAP interaction analysis uncovers a strong synergy between pyridinic-N and graphitic-N, where their combined impact on CO production exceeds their individual effects. Furthermore, the optimal pyridinic-N content depends on the carbon structure with distinct SHAP clustering for materials like graphene and carbon black. These insights provide a data-driven strategy for optimizing N-doped carbon catalysts, enabling targeted material selection to enhance CO₂ reduction to CO.