Automated generation of labeled synthetic training data for machine learning based segmentation of 3d-woven composites

Abstract

A novel pipeline for the generation of synthetic tomograms of woven composite materials, to be used for training of machine learning based segmentation algorithms is presented. The pipeline is completely based on open source software and heavily utilizes the graphical processing unit for fast data generation. The proposed method generates a surface mesh of the woven geometry, scans it, reconstructs the scan, and generates a voxel labeling of the generated tomogram. It is demonstrated that the method can generate images that show good agreement with experimentally produced x-ray computed tomography images of a 3D-woven carbon fiber reinforced polymer composite.