Machine learning assisted stability and CO₂ reduction reaction activity prediction of single atom alloys

Single-atom alloy catalyst design requires a synergistic understanding of stability and activity. Herein, density functional theory (DFT) and machine learning (ML) were integrated to investigate the stability and CO₂ reduction activity of 1131 SAA configurations across the periodic table. XGBoost regression model was developed to predict SAA stability, achieving performance with R² of 0.93 and MAE of 0.21 eV. Recursive feature elimination (RFE) and Shapley additive exPlanations (SHAP) analysis identified the bulk cohesive energy difference (ΔCEbulk), electronegativity difference (Δχ), and atomic radius difference (Δr) as key stability descriptors, where Δr quantifies geometric compatibility, Δχ modulates electronic distribution, and ΔCEbulk directly reflects the combined effect of geometric and electronic factors, aligning with the Hume-Rothery theory. For CO₂RR activity investigation, 514 CO adsorption energy (ΔE_CO) data points were collected, and Random Forest Regression (RFR) model was built with R² = 0.97 and MAE = 0.09 eV. Cohesive energy of the dopant atom (CEbulk_b), C-O bond length (l_C-O), and C-dopant bond length (l_C-N) were identified as dominant descriptors for ΔE_CO. Guided by the Sabatier principle (ΔE_CO∈[-0.77, -0.57 eV] and ΔE_H∉[-0.47, -0.07] eV), 26 promising SAA candidates were screened. DFT validation confirmed Al₁/Cu (111) and Au₁/Pd (111) reduce the potential-determining step barrier of CO₂RR to CH₄/CH₃OH, enhancing the reaction kinetics. This work provides a comprehensive framework for the efficient screening of stable and active SAA catalysts for CO₂RR.