Voorspelling van de transversale trekeigenschappen van vezelversterkte composieten: Een data en fysica gedreven aanpak

Unidirectional fibre-reinforced composites are widely used in high-performance engineering structures due to their superior strength-to-weight ratio and damage tolerance. A critical mode of early failure in these materials is transverse cracking, which can severely compromise their structural integrity and long-term durability. The microscale is the relevant scale to understand and predict transverse cracking in fibre-reinforced composites, as it captures the detailed interaction between fibres and the surrounding matrix. Predicting such damage initiation and evolution at the microscale is particularly challenging, given the intrinsic heterogeneity and complex architecture of the composite microstructure. Traditional experimental characterisation techniques are often limited by cost and scalability. Meanwhile, conventional microstructure generators fail to capture key microstructure features, such as resin-rich pockets, that influence damage mechanisms. To overcome these limitations, this thesis presents an integrated, data-driven, and physics-informed framework for extracting fibre geometries and microstructural features, generating realistic microstructures, and conducting high-fidelity finite element simulations to investigate transverse cracking mechanisms in fibre-reinforced composites. The first contribution of this work is to enhance the extraction of fibre geometries and microstructural features by overcoming the limitations of fast time-resolved synchrotron radiation computed tomography (CT). A novel deep-learning enabled image analysis pipeline is introduced that combines a 3D CycleGAN for unsupervised super-resolution and denoising with U-Net-id for high-accuracy segmentation of individual fibres. By comparing segmentation quality across fast- and slow-acquisition datasets, it has been demonstrated that the enhanced data from this pipeline achieves comparable performance to slow-scan images that required 200 times longer scan times. This pipeline, therefore, enables more robust data extraction from fast, time-resolved X-ray tomography, removing a critical shortcoming for the application of this technique to in situ studies. The second major contribution presents a Deep Convolutional Generative Adversarial Network (DCGAN) for generating realistic 2D transverse microstructures. These synthetic microstructures are evaluated using Fréchet Inception Distance, latent space exploration, and feature-matching techniques, confirming that they are statistically diverse and not replicas of their training data. Crucially, the generated images preserve key microstructural features observed in experimental CT scans, including the correct fibre diameters, volume fractions, spatial distribution, and resin-rich pockets (RRPs). Comparative analyses revealed that DCGAN significantly outperforms traditional generation methods in capturing these critical features, particularly for RRPs. Building upon this, the third contribution addresses the challenge of achieving geometric realism and periodicity in synthetic microstructure generation, which is essential for finite element simulations with periodic boundary conditions. To this end, a novel physics-informed, diffusion-based framework is introduced. This method integrates multiple experimental data sources, including fibre diameter measurements from laser diffraction systems and spatial topologies from SRCT scans. Spatial topologies were encoded by spatial coordinates and diameter as image channels, eliminating the reliance on pixel-based representations. Enhanced by a fine-tuning network and physics-based constraints, the developed model achieves overlap-free fibre arrangements, realistic diameter distributions, and periodic boundaries. Quantitative validation confirms that this approach accurately preserves fibre arrangement statistics and the geometry of resin-rich regions, producing simulation-ready statistical volume elements that reflect true microstructural variability. The final contribution investigates the effects of fibre arrangement on crack initiation and propagation using microstructures derived from CT, deep learning and traditional random-generated packings. Microscale finite element simulations are performed using Abaqus/Explicit on statistical volume elements with element-based periodic boundary conditions (PBCs). A triaxiality-insensitive epoxy resin constitutive model is assumed to capture its nonlinear failure strain at the microscale. An algorithm was developed to convert non-periodic image data into periodic representations, enabling the implementation of PBCs. A set of quantitative metrics, geometric characteristics of interface failure, local fibre volume fraction near failure sites, and transverse crack path morphology, is proposed to assess the failure behaviour. The results demonstrate that deep learning-generated microstructures can more accurately replicate thermal residual stress patterns and interfacial failure, as observed in models based solely on CT image data. Due to their random nature and the ability to generate large numbers of microstructures, these newly generated microstructures will support the statistical strength prediction of unidirectional composites based on their constituent properties. In conclusion, this thesis presents a robust, scalable framework that bridges experimental imaging, deep generative modelling, and physics-informed simulation for fibre-reinforced composite microstructures. It offers a significant advancement in the field of micromechanical analysis, enabling high-fidelity fibre segmentation, the generation of realistic, periodic microstructures, and the accurate prediction of complex damage behaviours, such as transverse cracking.