Graph neural network-driven prediction of high-performance CO2 reduction catalysts based on Cu-based high-entropy alloys

Abstract

High-entropy alloy (HEA) offer tunable composition and surface structures, enabling the creation of novel active sites that enhance catalytic performance in renewable energy application. However, the inherent surface complexity and tendency for elemental segregation, which results in discrepancies between bulk and surface compositions, pose challenges for direct investigation via density functional theory. To address this, Monte Carlo simulations combined with molecular dynamics were employed to model surface segregation across a broad range of elements, including Cu, Ag, Au, Pt, Pd, and Al. The analysis revealed a trend in surface segregation propensity following the order Ag > Au > Al > Cu > Pd > Pt. To capture the correlation between surface site characteristics and the free energy of multi-dentate CO₂ reduction intermediates, a graph neural network was designed, where adsorbates were transformed into pseudo-atoms at their centers of mass. This model achieved mean absolute errors of 0.08–0.15 eV for the free energies of C₂ intermediates, enabling precise site activity quantification. Results indicated that increasing the concentration of Cu, Ag, and Al significantly boosts activity for CO and C₂ formation, whereas Au, Pd, and Pt exhibit negative effects. By screening stable composition space, promising HEA bulk compositions for CO, HCOOH, and C₂ products were predicted, offering superior catalytic activity compared to pure Cu catalysts.