Density and speed of sound prediction for binary mixtures of water and ammonium-based ionic liquids using feedforward and cascade forward neural networks

Abstract Ionic liquids have attracted a lot of attention in the past years because of some of their properties that distinguish them from the classic solvents. Thus, the need for models that can represent their properties without new experimental efforts arises, as experiments are frequently expensive and time-consuming. Neural networks are processing systems capable of simulating biological learning and generalizing the learned functional relations to new cases never seen before. They have been used with success in several areas, like optimization, pattern recognition and function approximation. Therefore, they can be an important asset for properties prediction. This work is focused on designing, training and studying feedforward and cascade forward neural networks for density and speed of sound prediction for binary mixture of water and ammonium-based ionic liquids, using the temperature, mass fraction of ionic liquid and the structural groups of the reagents used to synthesize the ionic liquid as input variables. Besides the synaptic paradigm, some network parameters were also evaluated, namely the hidden neuron number and the number of layers. Also, 13 training algorithms were tested and had their performance evaluated. It was verified a superiority of the Levenberg-Marquardt method and the Bayesian regularization in the training. The proposed neural networks, two 12-10-10-1 cascade forward networks trained with Bayesian regularization, achieved an average absolute relative deviation of 0.0107% for density prediction and 0.1% for speed of sound prediction. The error dispersions showed the networks did not develop trends in prediction.