Autoregressive modeling and synthetic generation of high-fidelity, statistically equivalent 3D microstructures for as-manufactured misalignments in fiber-reinforced composites
Abstract
Fiber misalignment governs the microscale failure of composites, yet experimentally resolved three-dimensional microstructures remain scarce, costly, and specimen-specific, limiting high-fidelity virtual testing. This study presents an end-to-end framework that turns X-ray-μCT fiber paths into statistically equivalent, physically admissible 3D microstructures. A developed stochastic model reproduces descriptors by combining slice-wise empirical marginals, Gaussian-copula in-plane dependence, latent autoregressive depth memory, gated joint-tail enrichment, and a library of rare high-misalignment “motifs”, calibrated by Bayesian optimization to a cumulative objective of about 12.6%. An ablation study isolates each mechanism, with depth memory reducing the trajectory-continuity mismatch from over 200% to below 25%. The model transfers to an independent sub-volume with matching agreement and 1%–2% run-to-run variability. Driven online by the calibrated model, a slice-by-slice growth engine with variable-radius seeding, Delaunay-based neighborhoods, and an ellipse-aware projected Gauss–Seidel solver generates about 2400 fibers, while a closed-form between-slice capsule check certifies non-overlap over the full depth. Parametric studies up to 70% fiber volume fraction and 2795 fibers remain strictly overlap-free, and a three-way comparison of experimental, modeled, and generated microstructures quantifies the fidelity cost of enforcing physical admissibility. The verified microstructures are exported as non-intersecting solids importable into finite-element software, opening a scalable route to virtual testing, uncertainty quantification, and data-driven composite design.
Mohamad Raja
Doctoral candidate
My research interests include computational micromechanics-driven analysis of composites.
