LAUSR

Nitrogen-doped carbon materials are promising catalysts for electrochemical CO2 reduction (ECR), but achieving high Faradaic efficiency (FE) for CO production is challenging due to the competing hydrogen evolution reaction (HER). Conventional catalyst optimization approaches are often slow and labor-intensive. To overcome these limitations, we develop a machine learning-based stacked model to predict the Faradaic efficiency (FE) of CO in ECR, combining random forest (RF) and XGBoost (XGB) as base models with linear regression as a meta-model. This stacked approach significantly reduces overfitting, achieving R2 values of 0.98 for training and 0.91 for testing, compared to 0.99 (train) and 0.86 (test) using XGBoost alone with default parameters. SHAP (SHapley Additive exPlanations) analysis shows that pyridinic nitrogen plays a major role in increasing CO selectivity among the different nitrogen forms, while graphitic nitrogen is crucial for predicting hydrogen production during the HER. These findings highlight how machine learning accelerates catalyst design by providing insights into the roles of nitrogen configurations in tuning product selectivity. This study presents a novel, data-driven approach to electrocatalyst optimization, offering a pathway toward more effective CO2 reduction technologies.