Automated CT-Based Segmentation and Modelling of Textile Composite Materials

Abstract

Textile-reinforced carbon fibre composites are increasingly considered for aerospace applications due to their high specific stiffness and strength together with their ability to be tailored towards specific loading conditions. However, the mechanical behaviour of such materials is strongly influenced by the as-manufactured textile architecture, making accurate geometrical representations important for reliable prediction of homogenised material properties. X-ray computed tomography (XCT) combined with computational homogenisation provides a potential route for generating representative volume elements directly from physical composite samples. The present work investigates the applicability of an existing XCT-to-FE workflow, previously developed for 3D layer-to-layer angle interlock composites, to a 2D five-harness satin (5HS) woven carbon fibre reinforced polymer composite architecture representative of aerospace applications. The workflow combines XCT imaging, machine learning-based image segmentation, and voxelbased finite element homogenisation in order to estimate elastic material properties from XCT reconstructions of as-manufactured composite samples. Both virtual and physical investigations were carried out. A virtual 5HS composite sample generated in TexGen was utilised to evaluate XCT scan parameters and quantify segmentation performance against known ground truth data. In addition, physical XCT scans were performed on both resin-injected and dry fibre samples in order to assess the applicability of the workflow to manufactured materials. Experimental tensile testing combined with digital image correlation was further conducted in the principal material directions to provide reference elastic properties for comparison with the numerical predictions. The results showed that the pre-trained segmentation model was capable of identifying the general textile architecture, but that additional transfer learning using synthetic 5HS training data was required to obtain segmentation quality suitable for reliable voxel-based finite element modelling. Following transfer learning, a volume-wide pixelwise agreement value of 94% was obtained for the evaluated virtual sample. The study further showed that attenuation contrast significantly influences the segmentation performance, with the dry fibre samples visually exhibiting substantially improved material constituent separation compared with the resin-injected sample. Computational homogenisation of the segmented XCT reconstruction produced elastic stiffness predictions showing reasonable agreement with the experimentally measured tensile stiffness values, with a deviation of less than 12%. Although discrepancies between experimental and numerical results remained, the work demonstrates that the XCT-to-FE workflow can be extended to previously unseen 2D woven textile architectures through the introduction of additional target architecture-specific training data. Overall, the results demonstrate the potential of combining XCT imaging, machine learningbased segmentation, and computational homogenisation for non-destructive characterisation of as-manufactured textile-reinforced composite samples