3D U-Net-based segmentation of voids and delamination in composite materials using micro-CT data

Carbon Fibre Reinforced Polymers (CFRPs) are commonly used in aerospace because of their high strength and stiffness to weight ratios, but their performance can be affected by defects such as voids and delamination. High-resolution micro-computed tomography (micro-CT) provides detailed three-dimensional (3D) images of these internal features, although manually segmenting the data is both labor-intensive and often inconsistent. This study presents an automated methodology for voxel-wise segmentation of micro-CT scans of CFRP composites using a 3D U-Net convolutional neural network. A novel approach for generating pixel-level masks from scans was developed, enabling the creation of high-quality training datasets without manual annotation. To address class imbalance between background and defect voxels, a hybrid loss function combining Dice and Focal loss was used, along with Gaussian-weighted patch reconstruction for smooth volumetric predictions. The trained model reached a voxel-wise accuracy of 99.99%, with Dice coefficients of 99.996% for background, 99.54% for delamination, and 74.98% for voids. These results show the potential for automatically detecting voids and delamination in composite materials using 3D U-Net. The results of this study can be extended to improve predictive maintenance practices, enhance composites manufacturing processes, and to advance the development of usable digital twin models for aerospace structures.