Modeling study on the density and viscosity of ionic liquid-ionic liquid-water ternary mixtures

The vast diversity of ionic liquids (ILs) necessitates the development of accurate predictive models to support their industrial applications. This study combines machine learning (ML) algorithms with group contribution (GC) methods to model the density and viscosity of IL-IL-H₂O ternary mixtures. Three ML algorithms (i.e., ANN, XGBoost, and LightGBM) were employed to develop robust predictive models, which were trained on a large experimental dataset. The effect of dataset partitioning on the prediction results is analyzed, and the generalizability of the models is validated through 5-fold cross-validation. The ANN-GC model performs well in predicting both density and viscosity properties, with a mean absolute error (MAE) of 1.7909 and a correlation coefficient (R²) of 0.9933 for density, and an MAE of 0.0329 and an R² of 0.9813 for viscosity. Furthermore, hyperparameters for the ANN model were optimized using Bayesian optimization, while XGBoost and LightGBM were optimized via grid search. After optimization, the prediction accuracies of all three models improved, with ANN-GC maintaining the highest prediction accuracy. Specifically, the optimized ANN-GC model achieves an MAE of 1.5834 and an R² of 0.9963 for density prediction, and an MAE of 0.0279 and an R² of 0.9924 for viscosity prediction. Further insights were obtained through SHAP (SHapley Additive exPlanations) analysis, which clarified the contributions of different features to the model predictions. Additionally, the validity of the density and viscosity prediction models was confirmed by calculating the fluid flow unit process case.