Prediction of the clogging of ultrafiltration membranes by neural networks

Abstract

One of the important aspects affecting the overall effectiveness of membrane filtration systems is pore clogging. Pore blockage results in significant filtration resistance and a dramatic reduction in the filtrate flux rate at constant pressure conditions and a significant rise in pressure for membrane filtration operation under constant flux conditions clogging filtration. This work investigates the use of neural networks in modelling the clogging of ultrafiltration membranes. A feed-forward neural network (NN) model characterized by a structure (three neurons in the input layer, ten neurons in the hidden layer, and one neuron in the output layer) are constructed with the aim of predicting the clogging of the ultrafiltration membrane.

A set of 175 data points was used to test the neural networks.  70%, 15%, and 15% of the total data were used respectively for the training, the validation, and the test. For the most promising neural network model, the predicted clogging values were compared to measured rejections clogging ultrafiltration membrane values, and good correlations were found (R =0.99095 for the training phase, R =0.99180 for the validation phase, and R =0.9844 for testing phase). The mean squared errors were (MSE =0.0577938 for the training phase, MSE =0.063348 for the validation phase, and MSE =0.109035 for the testing phase).