Density, viscosity and CO2 solubility modeling of deep eutectic solvents from various neural network approaches

Abstract

Background

Deep eutectic solvents (DESs) have gained attention as innovative green solvents, but accurate prediction of their thermophysical properties is essential for practical applications. This work explored the potential of different deep learning approaches to model density, viscosity, and CO₂ solubility over a wide range of temperature and pressure conditions.

Methods

A comprehensive dataset was compiled, consisting of 2218 data points for density, 148 points for viscosity, and 144 points for CO₂ solubility, covering a range of DES compositions. Deep neural network (NN) architecture was employed for density prediction, while simpler artificial neural network (ANN) architectures were used for viscosity and CO₂ solubility predictions.

Significant findings

The deep NN model exhibited an excellent performance in predicting the density, achieving an average absolute relative deviation (AARD%) of 0.13 % and R² value of 0.9998, indicating high accuracy and robust generalization. The ANN models for viscosity and CO₂ solubility also demonstrated promising results, with AARD% values of 1.44 % and 1.11 %, respectively. The comparison with semi-empirical models further highlighted the superiority of NN approaches for characterizing these innovative solvents. This work showcases the capability of deep learning in accurately modeling the thermophysical properties of DESs, providing valuable tools for applications of these green solvents.