Prediction of CO2 reduction reaction intermediates and products on transition metal-doped r-GeSe monolayers:A combined DFT and machine learning approach

Abstract

The electrocatalytic CO2 reduction reaction (CO2RR) is a complex multi-proton-electron transfer process that generates a vast network of reaction intermediates. Accurate prediction of free energy changes (G) of these intermediates and products is essential for evaluating catalytic performance. We combined density functional theory (DFT) and machine learning (ML) to screen 25 single-atom catalysts (SACs) on defective r-GeSe monolayers for CO2 reduction to methanol, methane, and formic acid. Among nine ML models evaluated with 14 intrinsic and DFT-based features, the XGBoost performed best (R2 = 0.92 and MAE = 0.24 eV), aligning closely with DFT calculations and identifying Ni, Ru, and Rh@GeSe as prospective catalysts. Feature importance analysis in free energy and product predictions highlighted the significance of CO2 activation with O-C-O and IPC-O1 as the key attributes. Furthermore, by incorporating non-DFT-based features, rapid predictions became possible, and the XGBoost model retained its predictive performance with R2 = 0.89 and MAE = 0.29 eV. This accuracy was further validated using Ir@GeSe. Our work highlights effective SACs for CO2RR, and provides valuable insights for efficient catalyst design.