Predicting of membranes water fluxes using artificial neural network

摘要

Water fluxes of polyetherimide (PEI) membranes are predicted by a neural network model which takes moment invariants of the cross-sectional scanning electronic microscope (SEM) images as inputs. The prediction includes two steps, the first is to construct the backpropagation neural network (BPNN) model, and the second is using the model to predict water flux of another group of PEI membranes. To construct the BPNN model, moment invariants of cross-sectional SEM images used as pattern features to represent the PEI membranes, are calculated and normalized. Meanwhile, water fluxes of the PEI membranes are determined by experiments. These moment invariants together with the water fluxes are utilized to train the BPNN model, and to obtain the best network architecture. When another group of PEI membranes' moment invariants are inputted to the BPNN model, their water flux are predicted. Compared the experimental water flux with the predicted results, the standard deviation is less than 6.7%.