Deep Learning-Based 3D Local Fibre Orientation Estimation in Composite Computed Tomography: From Volumetric Convolutions to Real-Time Direct Projection Inversion

The mechanical integrity, load-bearing efficiency, and damage tolerance of short- and continuous-fiber-reinforced polymer (SFRP/CFRP) composites depend intrinsically on their local 3D fiber-orientation distribution. While X-ray micro-computed tomography (μCT) non-destructively reveals microstructural morphology, standard volumetric extraction relies on classical 3D structure-tensor eigen-analysis. These classical approaches require dense angular projection sets (180+ views), computationally demanding Feldkamp-Davis-Kress (FDK) reconstructions, manual filter tuning, and exhibit severe degradation under Poisson photon noise. This paper introduces an end-to-end deep learning framework for local fiber orientation estimation that operates across two practical regimes: high-precision offline volumetric inspection and real-time inline scan-to-orientation inspection. To overcome the scarcity of physical CT ground truth, we develop a physics-grounded synthetic pipeline coupling directional sampling via the von Mises-Fisher (vMF) distribution (κ∈[0.5,25]) with collision-free Random Sequential Adsorption (RSA) packing (V_f∈[10%,35%]) and GPU-accelerated forward projection with realistic Poisson noise. Ground truth is formulated analytically through second-order orientation tensors, mathematically resolving the unoriented fiber director ±180^∘ sign ambiguity. We benchmark three paradigms across N=7,889 held-out sub-volume patches drawn from 150 independent composite phantoms (70/15/15 parent split): (1) classical 3D structure tensor eigen-analysis, (2) a 3D volumetric convolutional neural network (Model B), and (3) a novel Direct Projection Network (Model C) that infers 3D orientation directly from only 8 raw 2D sinogram projection views, completely bypassing 3D volumetric reconstruction. Model B achieves an overall mean angular error of 9.35 ∘ (4.78 ∘ on aligned tape; a 67% error reduction relative to the 28.27 ∘ classical baseline) at an inference throughput of 140.9 patches/s on GPU. Model C demonstrates that 3D orientation can be recovered directly from sparse 2D projection shadows with 22.57 ∘ overall error (11.35 ∘ on aligned composites), achieving 174.0 patches/s with zero reconstruction computational overhead. Furthermore, industrial tolerance evaluation demonstrates an 84.57% in-spec acceptance rate at automotive tolerance (τ≤15^∘) for Model B. This work demonstrates that deep learning provides both a superior offline alternative to classical image filtering and a viable pathway for inline, reconstruction-free composite manufacturing quality assurance.