Exploiting the use of deep learning techniques to identify phase separation in self-assembled microstructures with localized graphene domains in epoxy blends

Abstract

In this study for the first time, convolutional neural networks (CNN) are applied to identify phase-separated microstructures in a novel nano-modified polymer composite. The input data consist of a mixed labelled dataset of homogeneous microstructure and phase-separated microstructure images collected using an optical microscope; the problem is modelled as a binary classification. The initial dataset consisted of microstructural images collected in house from past experiments, expanded in number using the data augmentation method, and upon testing, the model showed an accuracy of 65.4% and a 0.5 F1 score. Additional experiments were performed to generate more ground truth images and the CNN model built on this second data set showed improved accuracy of 80.1% and a 0.9 F1 score; The CNN classifier mimicked the receiver operating characteristic (ROC) curve of perfect classifier. Using the trained model, the phase separated microstructure is distinguished from homogeneous microstructure in a matter of minutes, saving hours of manual screening and cost intensive characterisation.