Reaching machine learning leverage to advance performance of electrocatalytic CO₂ conversion in non-aqueous deep...

Deep eutectic electrolytes (DEEs) show promise for future electrochemical systems due to their adjustable buffer capacities. This study utilizes machine learning algorithms to analyse the carbon dioxide reduction reaction (CO₂RR) in DEEs with a buffer capacity of approximately 10.21 mol/pH. The objective is to minimize undesired hydrogen evolution reactions (HER) and render CO₂RR dominant in a membrane cell. The CO₂RR process was found to be non-adiabatic, as the time of nuclear motion for CO₃^2- in K₂CO₃ product, through CO₂^●- trapping, is 0.368 femtoseconds shorter than the 1.856 × 10^-3s charge transfer relaxation time. Microkinetic analysis reveals that the rate of CO₂RR to CO₂^●- is 2.14 × 10³ mol/cm²/s² with a rate constant of 2.1 × 10¹⁰ cm/s. Our findings demonstrate that ensemble and k-Nearest Neighbours algorithms learn the CO₂RR dataset, achieving a prediction accuracy of over 99%. The models were verified visually and quantitatively by overlaying predicted and experimental dataset. Diagnostic and SHAP analyses highlighted the gradient boost ensemble algorithm, predicting asymptotic current densities of -4.114 mA/cm² or -13.340 mA/cm², with high turnover frequencies (TOF) of 3.79 × 10¹⁰ h^-1 or 12.30 × 10¹⁰ h^-1 for CO₂^●- or K₂CO₃ generation on silver electrodes, respectively. These results consider both accuracy and robustness against overfitting, providing an opportunity to optimize future non-aqueous electrolytes for convenient TOF measurements at industrially relevant current densities.