Deep learning model based on Bayesian optimization for predicting the infinite dilution activity coefficients of ionic liquid-solute systems

Abstract

The infinite dilution activity coefficient (γ∞) is a crucial thermodynamic property that provides a measure of the affinity between the solute and the solvent. In this work, three deep learning models (deep neural network, convolution neural network, and convolution deep neural network) were proposed to predict the γ∞ for ionic liquid-solute systems. 52 ionic liquids and 114 organic solutes (7783 data points in total) contained in the γ∞ database over the temperature from 293.15 to 428.15 K were collected to construct the model. Molecular descriptor was generated using the RDKit. The Bayesian optimization algorithm was employed to determine the parameters for each model, and the stability of the model was increased by 10-fold cross validation. Finally, the method of SHapley Additive exPlanations was employed to explain our proposed model. The coefficient of determination, mean absolute error, mean square error, root mean square error, error, sum of squared error, relative root mean square error and the Wilcoxon signed-rank test were used to assess the model. The results indicated that the experimental values were in satisfactory agreement with the values predicted for convolution deep neural network. These predictive models can be used to predict the γ∞ for ionic liquid-solute systems to solve problems in phase equilibrium and process design.