Deep-learning-based image-driven modelling of woven ceramic matrix composites: incorporating real void morphology and damage-induced behaviour

An image-driven method for generating high-fidelity finite element models for woven ceramic matrix composites is presented in this paper. The model incorporates the real morphologies of fibre tows, matrix, and inter-tow voids. With high-resolution X-ray CT scans of plain woven C_f/SiC composites, a deep-learning-based image segmentation method was employed to accurately segment the fibre tows from the CT images. A convex hull-based algorithm, alongside a thorough tow trajectory tracking method, was developed to handle fibre tow intersections and categorise the cross-sections of the fibre tows in the segmented images. For void identification, a rapid approach based on the watershed transform was implemented. The resulting geometric model demonstrates a high degree of morphological fidelity with the 3D CT image of the material. Furthermore, a progressive damage-induced nonlinear stress-strain relation for fibre tows was developed and incorporated into the CT image-driven finite element model. The predicted stress-strain curve has excellent correlations with experimental results, demonstrating higher accuracy compared to the ideal model, and therefore more accurately reflects the stress distributions.