Machine Learning–Driven Mapping of the Surface Structure and Activity Landscape in Cu–Zn Catalysts for CO₂ Electroreduction

Electrocatalytic CO2 reduction using bimetallic Cu-based catalysts offers a promising route for carbon-neutral carbon utilization. However, a lack of an atomic-scale understanding of active sites hinders the rational design of high-performance catalysts. In this work, we develop a machine-learning cluster expansion (CE) model, trained by density functional theory (DFT) calculations, to explore structure–activity relationships on Zn-doped Cu(111) surfaces for CO2 to CO conversion. By incorporating a Bayesian machine learning approach with leave-one-out cross-validation into the CE model fitting, we achieve high predictive accuracy while lowering the overfitting risk, even with a relatively small training set. Metropolis Monte Carlo simulations based on the CE model predict thermodynamically stable surface configurations, *CO adsorption energies, and turnover frequencies (TOF) across a broad range of Zn compositions. Our results show that Zn into Cu(111) significantly enhances catalytic activity, with an optimal Zn doping level of ∼15%, yielding a TOF approximately 28 times higher than that of pure Cu(111). This enhancement results from Zn surface segregation and the formation of Cu active sites modulated by Zn coordination. Specifically, the number of neighboring Zn atoms, such as three or four first-nearest-neighbor (first-NN) Zn atoms, fine-tunes *CO adsorption energies on Cu, placing them within the optimal activity window. This work provides atomic-level insights into the role of the local alloy structure in catalytic performance and offers a generalizable strategy for active site engineering in bimetallic electrocatalysts.