A novel self-attention mechanism-based machine learning model for high-accuracy prediction of CO2 solubility in natural deep eutectic solvents.

Natural deep eutectic solvents (NADESs) exhibit inherent eco-friendly characteristics and negligible pollution compared to traditional absorbents. However, identifying optimal NADESs for CO2 capture efficiently requires extensive experimental investigation, leading to significant time and resource consumption. Therefore, developing a robust predictive model to estimate CO2 solubility in NADESs accurately is crucial for advancing carbon capture technology. In this study, a novel machine learning model incorporating a self-attention neural network (SANN) based on the quantitative structure-property relationship (QSPR) method was proposed to predict CO2 solubility in NADESs. Unlike traditional models, the SANN model not only delivers high-precision predictions but also quantitatively evaluates the contributions of hydrogen bond acceptors (HBAs) and hydrogen bond donors (HBDs), enabling rational solvent design with minimal experimental effort. Evaluation on a test dataset shows that the SANN model achieves an average absolute relative deviation (AARD) of 4.41 % and a mean squared error (MSE) of 0.011, indicating high accuracy. Furthermore, the SANN model was employed to screen potential NADES candidates. The computational screening identified a potential NADES system comprising acetylcholine chloride and diethylene glycol, whose experimental CO2 solubility measurements exhibited remarkable agreement with model predictions, demonstrating the model's predictive accuracy. This approach significantly enhances sustainability by reducing experimental demands and encouraging the use of green solvents for efficient carbon capture.