Efficient mapping of CO adsorption on Cu_(1-x)M_x bimetallic alloys via Machine Learning

The electrochemical reduction of CO2 (CO2RR) has the potential to allay the greenhouse gas effect while also addressing global energy challenges by producing value-added fuels and chemicals (mostly C2 molecules such as ethylene and ethanol). However, due to the complicated chemical pathways involved, achieving high selectivity and efficiency towards specific reduction products remains challenging. In fact, the design of more selective and efficient catalysts often relies on trial-and-error approaches, which are very time consuming and resource intensive. In response, driven by the inherent importance of CO adsorption energy in the conversion of CO2 into C2+ hydrocarbons and alcohols, we propose a two-step approach employing machine learning classification and regression algorithms to predict CO binding energies on CuM(111)/(100) (M = Al, Ti, V, Fe, Co, Ni, Zn, Nb, Mo, Ru, Pd, Ag, Cd, Sn, Sb, Hf, W, Ir, Pt, Au) bimetallic surfaces. Firstly, we assess the stability of each adsorption site by utilizing classification algorithms. Subsequently, focusing exclusively on the stable sites, we employ regression models to predict the adsorption energies of CO. Remarkably, by employing a Gradient Boosting Classifier and Neural Networks for classification, together with a Gradient Boosting Regressor for regression, we predict CO binding energies with high level of robustness and accuracy for Cu bimetallic alloys with up to 17% of surface impurity concentrations. The accuracy of our models is demonstrated by F1 scores exceeding 98% and MSE below 0.05 eV^2 for the classification and regression parts, respectively. These remarkable results highlight the adaptability of our approach and its capability for efficiently screening Cu-based CO2RR electrocatalysts, enabling rapid evaluation of promising candidates for future in-depth explorations.