Intelligent Modelling of Moisture Sorption Isotherms in Milk Protein-Rich Extruded Snacks Prepared from Composite Flour

Abstract

In this paper, connectionist models have been investigated empirically to predict adsorption isotherms of milk protein-rich extruded snacks prepared from composite flour, at different temperatures (i.e., 28, 37 and 45 °C) and water activities (i.e., in the range: 0.112–0.971). These models were based upon error back propagation learning algorithm supplemented with Bayesian regularization optimization mechanism as well as with various combinations/settings of network parameters. In all simulation experiments, the connectionist models with single hidden layer were found to fit the best to the adsorption isotherms data. The best configuration of the connectionist models comprised 10 neurons in the hidden layer with tangent-sigmoid transfer function; which attained accuracy in the range of 0.467–0.958 root mean square percent error (%RMS). Also, several conventional mathematical sorption models including two-parameter models, viz., Lewicki-I, Mizrahi and Modified BET; and three- and four-parameter models, i.e., Ferro-Fontan, GAB, Lewicki-II, Modified GAB, Modified Mizrahi and Peleg were developed for the purpose. The Ferro-Fontan and Peleg were the best similar models among the conventional sorption models, with %RMS lying in the ranges: 1.63–1.89 and 1.41–3.33, respectively, for the same temperatures and water activities range. Evidently, the connectionist sorption models developed in this study were found to be superior over conventional sorption models, to efficiently and intelligently predict adsorption isotherms of milk protein-rich extruded snacks prepared from composite flour.