Viscosity of deep eutectic solvents: Predictive modeling with experimental validation

Abstract

Viscosity, the measure of a fluid's resistance to deformation, is a critical parameter in many industries. Being able to accurately predict viscosity is essential for the successful design and optimization of technological processes. In this research, regression models were created to predict the viscosity of deep eutectic solvents (DESs). Machine learning models were trained using a data set of 3440 data points for two component DESs. Different algorithms, such as Multiple Linear Regression, Random Forest, CatBoost, and Transformer CNF, were employed alongside a variety of structural representations like fingerprints, σ-profiles, and molecular descriptors. The effectiveness of the models was assessed for interpolation tasks within the training data and extrapolation outside of it. The results indicate that a rigorous splitting of the dataset into subsets is necessary to accurately evaluate the performance of the models. Two new choline chloride-based DESs were prepared and their viscosities were measured to evaluate the predictive capabilities of the models. The CatBoost algorithm with CDK molecular descriptors was chosen as the recommended model. The average absolute relative deviations (AARD) of this model exhibited fluctuations during 5-fold cross-validation, ranging from 10.8 % when interpolating within the dataset to 88 % when extrapolating to new mixture components. The open access model was presented in this study (http://chem-predictor.isc-ras.ru/ionic/des/).