Andreas Grießer

• In the last two decades, micro-computed tomography (micro-CT) scanning technology has become affordable and widely available for industrial and research applications. Imaging technology has evolved from requiring synchrotrons for high-resolution images in to desktop scanners providing submicron resolution or 4D dynamic imaging in a lab-based system. This popularized the application of micro-CT scans in material science. Fibers, by nature, have a very high aspect ratio, requiring high resolution to resolve their cross-section while needing enough field of view to capture their trajectory. Extracting properties of the microstructure from micro-CT scans opens up a wide range of possibilities, from comparing different manufacturing methods to creating digital models of the materials and changing them to investigate how specific microstructural properties impact overall performance. A method was developed to analyze the fibers in large micro-CT scans of nonwoven materials. Nonwoven fibers can be multiple millimeters long, typically in a diameter range of 15-30 micrometers. To overcome typical issues of skeletonization-based approaches, I present a machine learning-based method to find the centerlines of each fiber in the scan. These centerlines are then processed into a graph to correct errors in the neural network’s output. Training data for the neural network was obtained using the structure generator ”FiberGeo” in ”GeoDict”. I demonstrated the approach on a large nonwoven micro-CT. Fiber-based structures like carbon paper also play an important role as gas diffusion layers in fuel cells. Carbon paper is a porous composite made from carbon fibers and carbonized binder. When imaging these materials in a micro-CT, the binder and the fibers do not have any contrast. The distribution of binder and fibers influences the physical properties of the gas diffusion layers. To segment fibers from binder in these images, a machine learning-based approach is presented. Similar to before I created artificial models of carbon paper. For these models, we know which voxels are binder and which are fibers. The models are used to train a neural network that segments the binary images (pore and solid) into three-phase images (pore, fiber, and binder). I presented the approach on multiple scans and also validated it against a small cutout that was labeled manually. For injection-molded composites, the length of the fibers after the injection process can significantly influence the overall strength of the materials. To measure the fiber length, I enhanced the fiber identification method from before with a larger training dataset. I applied our enhanced method to a sample of a glass fiber composite and compared the results to experimental lab measurements. The method presented gave good agreement, enabling the measurement of fiber length without destruction of the sample and the experimental work.