Designing deep eutectic solvents for efficient CO2 capture: A data-driven screening approach

The continuous increase in carbon dioxide emissions into the atmosphere necessitates the exploration of new, efficient, and environmentally friendly systems for CO2 capture. Deep eutectic solvents (DESs), known for their unique physicochemical properties, have shown promising potential for replacing traditional absorbents due to their strong CO2 absorption capacity. In this study, we propose a two-step screening process, employing a data-driven approach, to design novel DESs as CO2 absorbents. At the initial screening stage, machine learning methods were used to develop models capable of predicting the CO2 absorption capacity of DES. To enhance the chemical diversity within the training set, we combined data on the CO2 absorption capacity of DESs (162 structures) and ILs (232 structures). Among the models developed, two demonstrated superior performance. The first one, called transformer convolutional neural fingerprint (TransCNF), and the second one, Random Forest (RFR) with extended connectivity fingerprint (ECFP), outperformed the others. To gain insights into the RFR/ECFP model, we employed the SHAP method and identified 30 significant descriptors. By comparing the contributions of these descriptors with the experimental observations, we found that the developed model accurately represented the influence of the structure of DESs on their CO2 absorption capacity. Thus, the model exhibited reliable performance. At the second stage of the screening process, we employed the Redlich-Kister thermodynamic model together with the machine learning model to predict the melting temperature of the DESs. Based on the screening results, we found that 1447 DESs with high CO2 absorption ability remained liquid at room temperature, which made them promising candidates for CO2 capture.