3D Carbon Fiber Tow Segmentation Using ML Methods

Author(s):

Jennifer Sietins, Xiongjun Wu, and Andrew T. Gaynor

Abstract:

X-ray CT image segmentation of carbon fiber (CF) tow orientations is challenging due to the sample consisting of the same material phases having the same gray-scale contrast, so traditional thresholding methods are not successful. Advancements in ML and deep learning are overcoming this challenge by utilizing feature size, shape, and orientation information in segmentation algorithms. Two segmentation software programs (DragonFly and TomoSAM) were used to segment a 3D orthogonal CF preform. The specific procedures and methods to overcome segmentation challenges are described. The segmented datasets were then incorporated into Porous Microstructure Analysis (PuMA) software for quantification of the effective thermal conductivity in various directions. Sensitivity analysis was conducted by reassigning overlapping voxels from the DragonFly segmentation, which found only minor differences in the thermal conductivity results. Results from DragonFly and TomoSAM segmentations showed similar results and were in general agreement with one another. Both segmentation methods and subsequent PuMA simulation demonstrate the feasibility for quantifying thermal properties on real microstructures with ML overcoming prior fiber orientation segmentation challenges.