Estimating Viscosity of Low Sugar Apple Marmalade Using Backpropagation Neural Network

Abstract

In this paper, a backpropagation neural network (BPNN) model was developed for the prediction of viscosity values of apple marmalade using experimental data collected from several measurements. Stevioside or Sucralose sweetener was used instead of sugar (sucrose) in some of the formulations. In the BPNN architecture, the shear stress, and shear strain with mass concentrations of the Stevioside, Sucrolase, and Sucrose were utilized as input, whereas the viscosity value of apple marmalade was used as an output to be estimated. The Stochastic gradient descent algorithm (SGD) was used to minimize the loss of the BPNN based on the experimental data set. The Mean squared error (MSE), and the coefficient (𝑅 2 ) were employed to assess the performance of the BPNN. The number of hidden neurons was found to be 20 using the adaptive hidden neuron algorithm. With 20 hidden neurons, the least MSE and the highest R 2 value were attained. Furthermore, the predicted viscosity values were found to be within 1% of the experimental viscosity values. The developed BPNN model can, therefore, be effectively utilized to predict the viscosity of any fruit marmalade using the same input and output parameters in the data range where the new data is normalized with the experimental data used in this paper.