Identification of machine learning neural-network techniques for prediction of interfacial tension reduction by zein based colloidal particles

Abstract

In this work, five different models of the artificial neural networks (ANNs) called feedforward, cascadeforward, timedelay, layrec and narx were used for IFT prediction as a function of biopolymers (zein and water-soluble portion of Zedo gum (WZG)) contents, pH, and time based on the experimental data. The architecture of the ANN models consisted of four inputs, one hidden layer and one output layer by considering the range of 1–60 neurons and the tansig activation function in the hidden layer. The experimental IFT data obtained from pendant drop method and drop shape analyzer (DSA) instrument were used to train the ANN models. The relative root mean square error (rRMSE) for the mentioned models were in the range of 0.68%− 2.51% and the overall score of timedelay model was higher than other models. Moreover, the relative importance index of the input data for more detailed assessment of IFT was determined for time delay model as the best model to predict IFT. Analysis showed that the impacts of time and zein content on IFT were greater than other factors. The results of this research showed the considerable ability of timedelay model (a data-based and knowledge-based combined approach) to predict IFT. This model can be used by food industry engineers as a very useful and cost-effective tool for process optimization and control the stability of emulsions.