Simultaneous Prediction of Density, Viscosity and Heat Capacity of Ionic Liquids- A Deep Learning Approach

Abstract

Estimation of properties of ionic liquids with artificial neural networks have been successful in overcoming the challenges posed by experimental predictions and equation of state models. With the advent of deep learning and affordable GPU-computing, there is potential to accelerate the predictions with better accuracy. In this work, different deep neural network architectures have been designed to determine the density, viscosity and heat capacity of ionic liquids (ILs) as single task models in the Python package Keras. The features for the ILs include temperature, pressure and molecular descriptors of the cation and anion. Additionally, multi-task neural networks were designed to simultaneously predict all three properties. The root mean squared error and R-squared have been used to evaluate the various models. The performance of the best multi-task model is compared to the best single task model for each property. Viscosity, heat capacity perform better in single task and density performs better in multi-task. Overall, multi-task learning shows to be promising and can be further improved by including more properties.