⚡ electrical engineering

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

This paper presents a physics-grounded deep learning framework that enables both high-precision offline and real-time, reconstruction-free 3D fiber orientation estimation in composite materials by leveraging synthetic training data and a novel direct projection network to significantly outperform classical methods in accuracy and speed.

Original authors: Swetank Singh, Ankit Singh Bhadauriya, Mayank Goswami

Published 2026-11-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Swetank Singh, Ankit Singh Bhadauriya, Mayank Goswami

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Modern machines, from the wings of airplanes to the frames of high-performance cars, increasingly rely on a special class of materials known as fiber-reinforced polymers. These are not uniform blocks of plastic but rather complex mixtures where tiny, incredibly strong fibers are embedded within a softer resin matrix. The strength and stiffness of the final part do not come from the material itself alone, but from how those fibers are arranged. If the fibers align with the direction of the force, the part can be dozens of times stiffer than if they are scattered randomly. To ensure these parts are safe, engineers must look inside them without breaking them apart, a process called non-destructive testing. For decades, the gold standard for this has been micro-computed tomography, a technique that uses X-rays to build a detailed 3D map of the material's interior. However, reading the orientation of millions of tiny fibers within that 3D map has been a slow, difficult, and often inaccurate task, especially when the images are noisy or the fibers are densely packed.

A team of researchers at the Indian Institute of Technology Roorkee has developed a new approach that changes how this internal inspection is performed. Instead of relying on traditional mathematical rules to guess the fiber direction, they trained a computer to learn the patterns directly. The challenge was that real-world scans do not come with a perfect answer key; engineers cannot know the exact orientation of every fiber in a physical object without destroying it. To solve this, the team built a virtual factory. They wrote a computer program that generates thousands of perfect, synthetic 3D models of these composite materials. In this virtual world, they placed millions of tiny cylinders to represent fibers, arranging them with precise mathematical rules to mimic everything from randomly scattered short fibers to perfectly aligned continuous strands. Because they created the models, they knew the exact orientation of every single fiber. They then simulated the X-ray scanning process on these models, adding realistic noise and imperfections to make the digital images look just like real scans. This provided the perfect training data: images with a known, exact truth.

Using this vast library of synthetic data, the researchers tested two different ways to teach a computer to find the fiber directions. The first method involved feeding the computer the full 3D volume of the reconstructed scan, similar to how a human might look at a solid block of material. The second, more radical method skipped the 3D reconstruction entirely. Instead, it looked only at a handful of flat, 2D shadow images taken from different angles. In traditional scanning, a machine must take hundreds of these shadow images and spend minutes or hours combining them into a 3D volume before any analysis can begin. The new method, however, learned to read the 3D orientation directly from just eight of these 2D shadows, bypassing the slow reconstruction step completely.

The results showed a significant improvement over the old ways of doing things. When the researchers compared their new computer models against the standard mathematical techniques used in industry today, the new models were far more accurate and much faster. The standard method, which relies on calculating gradients in the image, struggled with noise and often produced errors of nearly 30 degrees, a margin too large for precise engineering. The new 3D computer model reduced this error to less than 10 degrees on average, and even lower for well-aligned fibers. More importantly, it processed the data nearly four times faster than the traditional method running on standard computer processors. The model that looked directly at the 2D shadows was even faster, processing data at a rate that could keep up with a moving production line, while still maintaining an accuracy that was far superior to the old methods.

The study also tested how well these models held up under different conditions, such as when the fibers were very sparse or extremely dense. The computer models proved robust, maintaining high accuracy across a wide range of material densities that are common in real manufacturing. When the researchers applied a strict industrial standard for what counts as an acceptable part, the new models passed the test for the vast majority of samples, whereas the traditional method failed most of them. This suggests that deep learning can not only provide a more precise way to inspect materials in a laboratory but also offers a viable path for real-time quality control on factory floors. By learning directly from the raw data and understanding the physics of how light travels through the material, these new tools can see the hidden structure of composite materials with a clarity and speed that was previously impossible.

Technical Summary: Deep Learning-Based 3D Local Fibre Orientation Estimation in Composite Computed Tomography

1. Problem Statement

The mechanical integrity and load-bearing efficiency of short- and continuous-fiber-reinforced polymer (SFRP/CFRP) composites are intrinsically linked to their local 3D fiber orientation distribution. While X-ray micro-computed tomography ( $\mu$ CT) provides non-destructive volumetric visualization of these microstructures, extracting directional information via classical methods faces four critical industrial limitations:

Computational Bottleneck: Classical 3D structure tensor eigen-analysis requires hours of CPU processing for high-resolution volumes, precluding inline manufacturing inspection.
Noise Sensitivity: Gradient-based methods are highly susceptible to Poisson photon noise and partial-volume effects, leading to systematic angular errors exceeding $25^\circ\text{--}30^\circ$ .
Mandatory Full Reconstruction: Classical approaches strictly require dense angular projection sets (180+ views) and computationally expensive Feldkamp-Davis-Kress (FDK) filtered backprojection, creating latency and memory bottlenecks.
Parameter Heuristics: Filter scales must be manually tuned for specific fiber diameters and volume fractions, limiting generalizability.

Furthermore, supervised deep learning is hindered by the scarcity of physical CT ground truth; acquiring exact point-wise orientation labels without relying on the imperfect classical algorithms one seeks to improve is impossible.

2. Methodology

Physics-Grounded Synthetic Data Generation

To overcome the lack of ground truth, the authors developed a procedural simulation pipeline that generates paired CT data with exact analytical labels:

Phantom Synthesis: 3D microstructures are generated using Random Sequential Adsorption (RSA) for collision-free fiber packing (volume fractions $V_f \in [10\%, 35\%]$ ) and von Mises-Fisher (vMF) distribution sampling for directional alignment (concentration $\kappa \in [0.5, 25]$ ).
Forward Projection & Noise: The voxelized phantoms undergo forward projection based on the Beer-Lambert law, with realistic Poisson photon shot noise ( $I_0 = 10^5$ photons/ray) injected.
Reconstruction: Standard FDK filtered backprojection is applied to generate 3D volumes.
Exact Ground Truth: Orientation labels are derived analytically from the known 3D cylinder coordinates using second-order Advani-Tucker orientation tensors ( $T = \sum w_i v_i v_i^T$ ). This mathematical formulation resolves the unoriented fiber director $\pm 180^\circ$ sign ambiguity without image-derived heuristics.

Deep Learning Frameworks

The study benchmarks three paradigms:

Model A (Classical Baseline): Standard 3D Gradient Structure Tensor (GST) analysis using Gaussian-smoothed gradients and eigen-analysis.
Model B (3D Volumetric CNN): A 3D convolutional neural network that takes $32^3$ voxel sub-volumes from reconstructed CT data as input. It predicts the 6 independent components of the orientation tensor.
Model C (Direct Projection Network): A novel architecture that infers 3D orientation directly from 8 raw 2D projection views (sinograms), completely bypassing 3D volumetric reconstruction.

Loss Function and Symmetry

To handle the $\pm 180^\circ$ ambiguity, the network outputs a symmetric tensor. The training utilizes a composite loss function combining:

Tensor Frobenius Loss: For global component convergence.
Geodesic Spherical Arc Loss: Optimizes the angular deviation on the unit sphere $S^2$ , enforcing direction invariance via $|\hat{d} \cdot d_{true}|$ .
Data Augmentation: Exact 3D spatial rotations are applied analytically to the tensor labels to enforce rotational equivariance.

3. Key Results

The models were evaluated on 7,889 held-out sub-volume patches from 150 independent composite phantoms.

Accuracy and Throughput

Model B (3D CNN): Achieved a mean angular error of $9.35^\circ$ (median $7.14^\circ$ ), representing a 67% error reduction compared to the classical baseline ( $28.27^\circ$ ). On highly aligned composites ( $\kappa \ge 10$ ), the error dropped to $4.78^\circ$ . Throughput reached 140.9 patches/s on GPU.
Model C (Direct Projection): Despite using only 8 projection views (vs. 180) and zero reconstruction overhead, Model C achieved a mean error of $22.57^\circ$ overall, and $11.35^\circ$ on aligned composites. It achieved the highest throughput at 174.0 patches/s.
Classical Baseline (Model A): Suffered from high dispersion ( $28.27^\circ$ mean error) and low throughput (36.0 patches/s on CPU).

Sensitivity Analysis

Alignment ( $\kappa$ ): Model C is highly sensitive to alignment; performance degrades significantly in random 3D bulk regimes ( $\kappa < 3$ ) due to projection ambiguity but excels in aligned regimes. Model B maintains robustness across all regimes.
Volume Fraction ( $V_f$ ): Both deep learning models exhibit a U-shaped sensitivity curve, with peak accuracy in the 18%–25% volume fraction range, which corresponds to typical industrial SFRP parts.

Industrial Tolerance

At the automotive tolerance threshold of $\tau \le 15^\circ$ , Model B achieved an 84.57% in-spec acceptance rate, outperforming the classical baseline (24.80%) by a factor of 3.4.

4. Significance and Claims

The paper claims to demonstrate two primary advancements in composite non-destructive testing (NDT):

Superior Offline Inspection: Deep learning (Model B) provides a significantly more accurate and faster alternative to classical image filtering for laboratory characterization, effectively mitigating noise-induced errors inherent in gradient-based methods.
Viable Inline Inspection: The "Direct Projection" approach (Model C) proves that 3D orientation can be recovered directly from sparse 2D projection shadows without 3D reconstruction. This offers a pathway for real-time, reconstruction-free quality assurance in high-throughput manufacturing environments (e.g., automotive stamping or automated fiber placement), where full 3D reconstruction is computationally prohibitive.

The authors emphasize that their synthetic pipeline, grounded in first-principles physics and exact analytical tensors, allows for rigorous benchmarking without the "garbage-in, garbage-out" limitation of training on classical algorithm outputs. They conclude that while sim-to-real transfer requires careful validation against physical artifacts (e.g., beam hardening), the framework establishes a new accuracy benchmark for composite NDT.