Graduate Thesis Or Dissertation


Relating Damage to Microstructure in SiC-SiC Ceramic Matrix Composites With Micro Computed Tomography and Deep Learning Image Segmentation Public Deposited

Downloadable Content

Download PDF
  • The increasing need for high temperature structural materials in aerospace applications is driving the development of new Ceramic Matrix Composites (CMCs). CMCs mitigate the inherent brittleness of ceramics through fiber reinforcements. A weak fiber-matrix interface allows for crack bridging and fiber pullout to increase their toughness; however, their design complexity is limiting the understanding of CMC behavior. In-situ Synchrotron micro Computed Tomography (µ-CT) experiments are conducted to characterize microstructural variations in CMCs and their effect on damage. Limitations in µ-CT experiments arise due to the difficulty of analyzing large quantities of images, especially for CMCs with SiC fibers and SiC matrix which lack grey level contrast between material phases. The overall goal of this thesis is to relate damage to microstructure in CMCs. The first step towards this objective is to facilitate the extraction of quantitative information from challenging SiC-SiC CMC CT images using automated image segmentation methods. An automated Deep Learning model for segmenting fiber reinforced composites is developed and validated using three methods. Automated segmentations of CMC images remove the current bottleneck and limitation in extracting information from CT images owing to extremely large amounts of data. Image segmentations of CMCs are shown to be an effective tool for measuring internal fiber architectures and microstructural variations. This provides means for material quality control as well as correlating damage to microstructures. Damage in the form of matrix cracks and fiber breaks in unidirectional samples are related to microstructural variations in fiber alignment, spatial distribution of fibers, thicknesses of fiber coatings and defects in the matrix. Matrix cracks rendered in 3D are more complex than previously described in literature due to bifurcations and deviation in orientation. The effect of non-uniform spatial distribution and variations in coating thicknesses of fibers on fiber breaks was investigated. Finally a study on the volume of training data required for a successful deep learning model to segment images of fiber reinforced composite images was conducted. In this study it was found that the volume needed for training deep learning models are dependent on the complexity of the images. It was also shown that synthetic augmentations from manually annotated ground truth images provide an advantage to cases with limited training volume.

Date Issued
  • 2021-11-17
Academic Affiliation
Committee Member
Degree Grantor
Commencement Year
Last Modified
  • 2022-06-23
Resource Type
Rights Statement