New insights into sulfur dioxide absorption in deep eutectic solvents

Abstract

This study intends to accurately predict the absorption capacity of sulfur dioxide (SO₂) in deep eutectic solvents (DESs) by developing robust machine learning (ML) models. A comprehensive dataset comprising 1382 experimental and 924 calculated data points was constructed, covering 156 unique DESs derived from 22 hydrogen bond acceptors (HBAs) and 42 hydrogen bond donors (HBDs). Input variables included sigma profile descriptors (s1–s10), water content, pressure, and temperature. Fifteen ML algorithms were evaluated, including Support Vector Regression (SVR), Convolutional Neural Networks (CNN), Random Forest, XGBoost, and LightGBM. Dataset reliability was confirmed using a Monte Carlo outlier detection algorithm. Among all models, SVR and CNN attained the highest level of performance on the testing dataset, with R² values of 0.9872 and 0.9904, mean squared errors (MSE) of 0.0015 and 0.0011, and Mean relative deviation (MRD) of 9.11% and 8.95%, respectively. The analysis identified pressure and structural descriptors (particularly s8) as key variables influencing SO₂ absorption. These results highlight the effectiveness of ML techniques in modeling gas absorption behavior in complex DES systems and support their application in the design of high-efficiency solvent systems.