Micro-CT Based Automated Reconstruction of 3D Textile Composite Microgeometry Using Segmentation-Based Machine Learning

3D woven polymer matrix composites (PMCs) consist of a polymer resin matrix reinforced with fiber bundles (yarns) woven in three dimensions, resulting in a complex heterogenous and anisotropic microstructure. This intricacy of the weave architecture presents significant challenges for accurate modeling and analysis, particularly in capturing local variations and defects within specific regions of interest. Traditional methods for generating realistic computational models of these materials are often labor-intensive and time-consuming. In this study, we propose an automated framework that combines machine learning and geometry processing techniques to reconstruct 3D digital representations of woven composite domains from μCT scans. The framework employs instance segmentation by fine-tuning a pretrained Detectron2 model to generate segmented yarn masks from 2D μCT image slices, enabling automated detection of individual yarns. Subsequently, point clouds are extracted from the segmented masks, and slice-to-slice entity tracking is performed to assemble a continuous 3D surface of the reinforcement structure. The framework is applied to PMCs with three yarn types (weft, warp and binder) and validated using both quantitative and qualitative metrics. This approach significantly reduces manual intervention, streamlines the modeling process, and enhances the accuracy of computational models for downstream analysis.