Efficient Machine-Learning-Based New Tools to Design Eutectic Mixtures and Predict Their Viscosity

The development of models that accurately predict the formation of eutectic mixtures (EMs, including the well-known deep eutectic solvents) and their viscosity is imperative to save time in synthesizing new solvents. We developed reliable machine-learning-based classifiers able to discern between eutectic and noneutectic (non-EM) mixtures and regressors able to predict the viscosity of an EM. A new experimental data set of 219 EMs, 384 non-EMs, and 1450 viscosity points at different temperatures and water contents is provided and used to challenge several models, defined both by an algorithm and by descriptors. The top-performing EM/non-EM classifier yields an accuracy of 92%, and the best regressor achieves viscosity predictions with a mean absolute error of 2.2 mPa·s; the extrapolation capabilities of the latter were assessed on additional measurements at temperatures and water contents outside the range of the training data set, revealing good accuracy at low viscosities. The SHapley Additive exPlanations (SHAP) algorithm was employed in several models as an eXplainable Artificial Intelligence (XAI) technique to quantify input feature contributions to the model output. These results represent a significant step forward in developing robust and highly accurate models for determining eutectic mixtures and their viscosity.