Physics-informed parameterised generative model for unidirectional fibre-reinforced microstructures with periodic boundaries

Accurate generation of realistic statistical volume elements is essential for predicting and optimising the mechanical behaviour of fibre-reinforced composites. Deep generative models that rely solely on imaging data often lack geometric accuracy and periodicity. We introduce a physics-informed deep learning framework based on diffusion models, which integrates multi-source experimental data: laser diffraction system measurements for ground-truth fibre diameter and computed tomography for image-derived spatial topology. The methodology features a novel parameterised microstructure grid that encodes spatial and dimensional attributes at each pixel, resolving fibre-order ambiguity. Augmented with a refinement module and physics-based constraints, our model generates realistic, overlap-free fibre arrangements that reflect representative diameter distributions and enforce periodic boundaries. Quantitative comparisons with real microstructures show that our method accurately captures key geometric features, including fibre spatial distributions and resin-rich regions. These results highlight the potential of physics-guided diffusion models to generate reliable microstructural representations, paving the way for accelerated microscale materials design and analysis.