General prediction of transport and thermodynamic properties of deep eutectic solvents based on choline chloride using machine learning

Abstract

This study delves into the potential of machine learning (ML) to predict the properties of choline chloride-based deep eutectic solvents (DESs), focusing on density, viscosity, and ionic conductivity. Utilizing a dataset derived from an extensive review of scientific literature, we applied a variety of ML models, including tree-based algorithms and neural networks, to establish correlations between DES features and their physical properties. Our findings demonstrate that models like Random Forest, XGBoost, LightGBM, and Stacked Models excel in predictive accuracy, particularly for density and viscosity, as evidenced by high R2 and Pearson correlation values. Our exploration into ionic conductivity revealed that despite initial assumptions, the size of the dataset did not limit the predictive capability. Learning curve analysis illustrated that models like LightGBM performed consistently well across varying dataset sizes, maintaining accuracy in their predictions. Notably, the Stacked Models emerged as the most effective, suggesting the benefit of combining different ML approaches for such predictions. The study also employed SHAP analysis to discern the impact of specific molecular features on the predictive outcomes, providing deeper insights into DES behavior. In essence, our results confirm that ML can be a powerful tool in the predictive modeling of DES properties, which has significant implications for the design and optimization of these solvents in various industrial applications.