Electrochemical CO2 reduction: Predicting the selectivity

Abstract

Electrochemical CO₂ reduction can lower the global carbon footprint while producing value-added products. The success of this approach is dependent on the development of highly selective electrocatalysts. Recently, descriptor-based approaches have been able to determine the selectivity of the major product groups. This work expands on the descriptor-based selectivity approach by using machine learning to create a mapping for experimentally determined product distributions. We report to accurately be able to predict product distributions based on Density Functional Theory (DFT) -based descriptors. Using our model, we predict areas of high ethanol faradaic efficiency and using an ensemble of models we quantify the model uncertainty in this area. Post hoc model analysis allows for model interpretation and determining feature importance, which gives a chemical insight into what determines the selectivity of CO₂ reduction reaction. The descriptor-based machine learning approach allows for accurate screening of selective catalyst candidates without a complete understanding of the complex reaction mechanistics.