Analysis of CO selectivity during electroreduction of CO2 in deep eutectic solvents by machine learning

Abstract

In this work, supervised and unsupervised machine learning approaches were applied to determine routes to high CO selectivity during the electroreduction of CO2 in deep eutectic solvents (DES) utilizing the molecular, chemical, and physical characteristics of hydrogen bond donors and acceptors, as well as the properties of different electrodes and DES solvents. In addition, effective data visualization and machine learning techniques were employed to identify relationships between descriptor variables and CO faradaic efficiency. First, SHAP (Shapley Additive exPlanations) analysis was applied to determine the positive and negative effects of the descriptor variables on the target, and it was found that urea in HBD (hydrogen bond donor) has the greatest impact on the target. Then, principal component analysis (PCA) was used to identify the combinations that lead to low, medium, and high levels of the target. PCA indicated that high-level clusters may be linked with HBA (hydrogen bond acceptor) molecular properties rather than HBD in addition to choline chloride-type HBA, HBA/HBD ratio, HBD density, HBD melting point, and urea-type HBD. Finally, decision tree classification was used to discover the variables leading to very high levels of the target. The decision tree revealed one pathway with very high CO faradaic efficiency and two pathways with high CO faradaic efficiency. To conclude, future researchers will be able to design new experiments with less effort and time while analyzing the effect of new DES components for high-performance CO2 electrolyzers as a result of the machine learning study and exploratory data analysis performed in this study.