Machine learning accelerated exploration of CO2 reduction electrocatalyst

Abstract

CO2 reduction is proved to be a possible way to recycle excessive atmospheric CO2 while producing alternative fuels and useful chemicals. Catalysts for electrochemical CO2 reduction reaction (CO2RR) is the key to an efficient conversion towards desired products. Despite the endeavours of many researchers, the progress in discovering high-performing electrocatalysts has been limited by time-consuming experiments and DFT calculations. Recently, Machine Learning (ML) emerges as a powerful tool in catalysis research due to its high efficiency and accuracy. Therefore, we explore the feasibility of using ML to accelerate the discovery of CO2RR catalysts by predicting their selectivity. A Crystal Graph Convolutional Neural Network (CGCNN) is implemented to predict the adsorption energy of key intermediate on the surface of certain catalysts. Then, the adsorption energies of CO and OH are used to qualitatively predict the major product according to a thermodynamical analysis. Using this method, the selectivity of five single metal catalysts and two bimetallic systems with various compositions are analysed. Experiments are conducted to verify the ML result. As a result, it shows that ML prediction is able to provide useful insight into electrochemical CO2RR catalyst selection.