Predicting the CO₂ Capture Capability of Deep Eutectic Solvents and Screening over 1000 of their Combinations Using Machine Learning

Abstract

Deep eutectic solvents (DESs) are a new class of environmentally friendly solvents that have attracted the attention of many researchers. Since DESs have several practical applications in CO2 capture, knowledge of their CO2 solubility is crucial. In this study, the CO2 solubility of DESs was predicted via a multilayer perceptron (MLP) using molecular descriptors derived from the Conductor-like Screening Model for Real Solvents (COSMO-RS). An extensive database of 2327 data points was created from 94 unique DES mixtures made from 2 anions, 17 cations, and 39 hydrogen bond donors (HBDs) at 150 different compositions and operating conditions of temperatures and pressures. Several statistical tests were performed, and after thorough hyperparameter tuning, it was found that the best MLP architecture predicted the CO2 solubility with an R2 value of 0.986 ± 0.002 and an average absolute relative deviation (AARD) of 4.504 ± 0.507. The MLP has also been loaded into an accessible Excel spreadsheet included in the Supporting Information. Thereafter, in order to guide DES molecular design for achieving high CO2 solubilities, the MLP was utilized for the high-throughput screening of 1320 DES combinations. This model encourages the creation of robust and accurate models to predict the CO2 solubility of novel DESs, which will minimize the need for conducting costly and time-consuming experiments.