An artificial neural network to calculate pure ionic liquid densities without the need for any experimental data

Abstract

In this study, a feed-forward artificial neural network, with three layers and seventeen neurons, was constructed to estimate the densities of a wide range of ionic liquid families including those based on the imidazolium, ammonium, pyridinium, pyrrolidinium, and isoquinolinium cations, together with various anions, as well as varying lengths of alkyl side chain lengths. The model is a function of the molecular weight and structure of the ionic liquid, and the system condition of temperature and pressure, which covers a temperature range of (273.15 to 393.17)K and a pressure range of (0.1 to 100)MPa. Therefore, no additional experimental data on the ionic liquid is required as input parameter(s), which makes this technique quite versatile. It was observed that the estimated values of densities of pure ionic liquids have very good agreement with the experimental data. The training correlating coefficient (R), the training performance (MSE), and the average absolute error on the training dataset were 0.99997, 6.04×10−6, and 0.019%, respectively. The average absolute error value on the test dataset is 0.014%.