Data-driven insight into CO2 electroreduction: Ensemble machine learning for CO Faraday efficiency prediction in single-atom catalysts

This study presents an ensemble machine-learning (ML) framework to predict the Faraday efficiency of CO formation (FE CO ) for electrocatalytic CO 2 reduction (CO 2 RR) on single-atom catalysts (SACs). A literature-derived dataset of 795 experimental entries from 42 primary studies was compiled, representing 24 carrier classes, 12 dopant atoms, and H-cell, flow-cell, and membrane-electrode-assembly (MEA) configurations. Eighteen regressors from nine algorithm families were systematically evaluated. A stacking ensemble combining gradient boosting decision tree, random forest, extremely randomized trees, support vector regression, k -nearest-neighbor regression, and CatBoost base learners with a kernel-based meta-regressor achieved the highest record-level performance, with a test R 2 of 0.9106 and an RMSE of 8.41 percentage points. Repeated five-fold cross-validation confirmed robust within-domain performance across random partitions (mean R 2 = 0.882 ± 0.006 ). Permutation importance, SHAP, partial-dependence analysis, and descriptor ablation identified dopant atom, applied potential, catalyst identity, and carrier as the dominant model-associated descriptors. Screening recovered established associations with Ni-based SACs, with graphene showing the highest mean model-predicted FE CO among the represented carrier classes. A discrete metal-potential response surface identified a Ni-associated region with a mean predicted FE CO of approximately 85% near -0.77 V versus RHE and a maximum of approximately 99% near -0.97 V. More stringent catalyst-, publication-, and Ni-restricted grouped validation showed weaker transfer to unseen systems. Thus, record-level performance primarily reflects interpolation within the represented experimental domain, while screening identifies experimentally grounded, dataset-conditioned trends for further validation.