Neural Network Modeling for The Microfiltration Process of Bacteria Harvesting From Fermentation Liquid

Abstract

Microfiltration is considered useful processes in the field of food industry separation. In the present study, neural network modeling for laboratory data was performed using a ceramic microfiltration membrane to isolate Kocuria rhizophila bacteria from the feed stream. The independent parameters of this process are the back pressure of the membrane, speed of the current entering the membrane, contact time, and the dependent parameter of the flux passing through the membrane. The neural network used was a multilayer perceptron (MLP) system. The data are divided into three main parts of education, validation, and evaluation with a distribution percentage of 15-15-70. The variable of the number of hidden layer neurons in this study was changed from 1 to 20, and the value of 17 neurons was selected as the optimal number according to the results. for verify the prediction performance of the data in the neural network, two basic parameters of the determination coefficient (R2, mean square error (MSE)) were used. The R2 values for the training data, validation, evaluation, and all data using the learning function and the optimal transfer function of Levenberg-Marquardt and Tansing were 0.9995, 0.99966, 0.999111, and 0.9994, respectively. In addition, an MSE value of 0.008 was obtained, indicating a very low error rate. The results show that the neural network used and the optimal function obtained had the least errors in the data calculations and predictions.