Deep Learning Models for Image-Based Damage and Microstructure Characterization in CFRP and CMC Composites

Abstract

As composite materials become more widely used in different industry fields other than automotive and aerospace, methods of investigating the microstructure are critical when testing the physical properties of the composite. Unlike metals, composites are made up of different material phases which cause them to fail and develop internal damage in unique ways. Micro computed tomography (micro-CT) has emerged as one of the most useful techniques to assess three dimensional (3D) microstructures and damage in composite materials. Machine Learning (ML) is becoming one of the essential tools used to analyze large volumes of experimental data in materials science. In this work, ML is used for automatic segmentation of micro-CT imaging data of low-velocity impact damage in carbon fiber reinforced polymer (CFRP) composites and of the microstructure and damage in ceramic matrix composites (CMCs). For CFRPs, deep learning models based on U-Net, BiSeNet, INet, and FC-DenseNet architectures were trained and refined to evaluate the accuracy of supervised ML compared to unsupervised ML. The unsupervised ML methods utilized statistical distances and greyscale threshold intensity segmentation to isolate damage in high-resolution image data. Results show that the Kullback-Leibler divergence is the most conservative and, thus, preferred statistical distance for unsupervised ML in CFRP composites. Furthermore, when comparing with the supervised ML models, FC-DenseNet provided the most accuracy but U-Net, which was the second most accurate, also provided much faster training and segmentation times. For CMCs, microstructure and damage in a minicomposite consisting of continuous silicon carbide (SiC) fibers embedded in a SiC matrix with boron nitride (BN) coating was investigated under a 40 N tensile load. Deep learning models based off of a 2.5D U-Net architecture were used to identify the different material phases and isolate cracks in the CMC minicomposite. A sequence of two-class semantic image segmentation models was created to isolate pores, fibers, fiber coating, and matrix, which were then used to build and train a five-class semantic image segmentation ML model. This model could identify fibers, fiber coating, matrix, pores, and background simultaneously. The segmentation results were analyzed to estimate volume fractions of composite phases and compared to literature values. Results show that prioritizing high contrast phases when training the U-Net model attained volume fractions that compared well to results reported in the literature. Furthermore, the use of deep learning models was investigated to determine their ability to identify and isolate matrix cracks within the microstructure.