Feedforward and cascade forward networks for viscosity prediction for binary mixtures of ammonium-based ionic liquids and water

Abstract

Ionic liquids have been attracting great interest due to some properties, such as low vapor pressure, thermal and chemical stability, non-flammability and the possibility of designing ionic liquids for specific purposes. The development of new processes and the substitution of traditionally used solvents, however, depend on the availability of data for the substances involved, either experimental or modeling data. Neural networks are robust models that can be used for optimization, pattern classification and recognition, property prediction and many other applications. They can use experimental data to, after enough training, learn functional relationships between variables and then generalize them for unseen cases. In this work, neural networks were trained and applied in the prediction of viscosity for binary mixtures of ammonium-based ionic liquids and water using a group contribution approach. The input variables were the temperature, the ionic liquid mass fraction and the amount of each group used to describe the ionic liquids structures. The influence of the hidden neuron number and the addition of an extra hidden layer was investigated. Furthermore, 13 training algorithms were tested. Feedforward and cascade forward networks were compared. The average absolute relative deviations for the testing set varied from almost 200% to 0.1999% depending on the selected parameters. The results showed neural networks can be used successfully to predict viscosity data, but a careful selection of architecture and training algorithm is required.