MODELING DEAD-END ULTRAFILTRATION OF APPLE JUICE USING ARTIFICIAL NEURAL NETWORK

An artificial neural network (ANN) was developed to model the dead-end ultrafiltration process of apple juice. Molecular weight cutoff, transmembrane pressure, gelatin–bentonite concentration and time were the input variables, while filtrate flux and filtrate volume were the output variables of the ultrafiltration process. According to error results and correlation values for two types of network (one or two hidden layer configurations), configurations with two hidden layers had comparatively better performance. The highest correlation coefficient with the minimum prediction error was calculated for two hidden layers with 6-5 nodes configuration. Trained ANN (4-6-5-2) predicted filtrate flux and filtrate volume with 2.33 and 1.38% mean relative error, respectively. The results suggest that the ANN modeling can be effectively used to optimize filtration process.

PRACTICAL APPLICATION

Membrane separation processes including ultrafiltration have gained importance in the food industry. Today, fruit juices are widely clarified by means of ultrafiltration process instead of tedious and laborious conventional clarification treatments. Membrane fouling which results in flux decline is the main problem associated with the ultrafiltration of fruit juices. In order to perform an efficient ultrafiltration process, optimization is required to obtain maximum filtrate volume per unit time. Artificial neural network (ANN) modeling offers great advantage on improving the performance of ultrafiltration process by accounting the effects of different variables, i.e., feed properties, transmembrane pressure and membrane pore size on filtrate volume as the main output of the filtration process. ANN modeling of ultrafiltration may be an alternative to previously proposed empirical and semiempirical models.