Segmentation-Free Estimation of Fibre, Matrix and Void Fractions from Simulated X-ray CT of Fibre Composites with a 3D Residual Network
Year: 2026
Type: preprint
Abstract: Fibre volume fraction and void content govern the stiffness and interlaminar strength of fibre-reinforced polymers, but measuring them from X-ray computed tomography (CT) normally requires voxel-wise segmentation, which is fragile when carbon fibre and epoxy differ little in attenuation and when fibres are only a few voxels across. We instead regress the three volume fractions of a CT sub-volume directly using a 3D residual network (3.6 million parameters, with a softmax output ensuring that the fractions sum to one). Training and test data are simulated: 66 synthetic composites with fibre contents ranging from 40% to 66% and void contents ranging from 0% to 5% are generated on a supersampled grid, ensuring that every crop has exact ground truth. The composites are scanned using a physics-based CT simulator incorporating parallel-beam projection, Poisson noise at three dose levels, detector blur, and filtered back-projection. On 216 test crops from 9 held-out composites, the network achieves a mean absolute error (MAE) of 0.84 percentage points (pp) [95% confidence interval: 0.66–1.05] for fibre and 0.17 pp [0.11–0.25] for void, compared with 6.06 pp and 0.28 pp, respectively, for the best threshold- or histogram-based methods. The network is more accurate on all 9 test composites for fibre and matrix (Wilcoxon p = 0.0039, the minimum attainable value for n = 9). Its fibre error increases only from 0.69 to 1.07 pp as the contrast-to-noise ratio decreases from 5.2 to 1.2. For void estimation, the network does not significantly outperform a Gaussian-mixture histogram fit (p = 0.36) and tends to overestimate void content below 1%. A mean-grey-value estimator calibrated on validation data achieves comparable performance (1.09 vs 0.87 pp; p = 0.20). However, a 2% calibration error increases its estimation error to 8.4 pp, whereas the network, whose input is normalized independently for each crop, remains completely unaffected.
Source: Research Square
Authors: Abhinav Swaran Singh, Mayank Goswami
Institution: Indian Institute of Technology Roorkee
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225812021