DEVELOPMENT OF CONVOLUTIONAL NEURAL NETWORK MODELS FOR FEATURE EXTRACTION OF PVDF MEMBRANE SURFACES

Abstract

 Previous models of membrane filtration processes required simplification of the membrane structures; thus, it was difficult to characterize the actual porous membranes. Therefore, this study aimed to develop a model that can delineate features of the actual porous membranes and deposited particles. Convolutional neural networks (CNN) were used as the architecture of the deep-learning models developed in this study. After tuning hyperparameters, the CNN models were able to classify the SEM images of polyvinylidene difluoride (PVDF) flat sheet membranes of different pore sizes, i.e. 0.1 µm and 0.45 µm, with high accuracy. The accuracy improved with the epoch size and reached the 100% accuracy after 50 epochs. The CNN model was modified and applied to classification of 0.45 µm pore-sized membranes filtered with feed waters containing 2, 5 and 8 mg/L of nanoparticles. Although the overall classification accuracy was 93%, the classification of 2 and 5 mg/L was less accurate (84% and 88%, respectively) due to small differences of particles deposited on the membranes. The heatmaps of the membranes were drawn by gradient class activation mapping (Grad-CAM) of the CNN models. The Grad-CAM heatmaps output by the CNN model with a kernel size of 3x3 was able to clearly extract detailed features of the membranes, such as aggregates of PVDF spherical crystalline structures and aggregates of particles deposited on the membranes, whereas a 7x7 model featured wide areas including deep pores, top and intermediate layers. Thus, it was found that the CNN models can extract the features of both intact and fouled membranes without any simplification of the membrane structures. By adjusting the model structure, the kernel size and epochs, the CNN models can be used for feature extraction and comparison of different membranes and for improving our understandings of membrane fouling processes.