Analysis of the Compaction Behavior of Textile Reinforcements in Low-Resolution In-Situ CT Scans via Machine-Learning and Descriptor-Based Methods

TL;DR

This work addresses the challenge of characterizing nesting in textile reinforcements during compaction using low-resolution in-situ CT. It combines a tailored 3D-UNet for semantic segmentation of matrix, weft, and fill with descriptor-based statistical analysis, notably the two-point correlation S₂, to extract average layer thickness and nesting factors. The approach yields high segmentation accuracy at early stages and robust nesting quantification that agrees with micrograph-based benchmarks, even under limited resolution and noise. The integrated pipeline provides a non-destructive framework for process–structure–property investigations in textile-reinforced composites and supports future enhancements in reconstruction, synthetic data generation, and orientation-aware microstructure modeling.

Abstract

A detailed understanding of material structure across multiple scales is essential for predictive modeling of textile-reinforced composites. Nesting -- characterized by the interlocking of adjacent fabric layers through local interpenetration and misalignment of yarns -- plays a critical role in defining mechanical properties such as stiffness, permeability, and damage tolerance. This study presents a framework to quantify nesting behavior in dry textile reinforcements under compaction using low-resolution computed tomography (CT). In-situ compaction experiments were conducted on various stacking configurations, with CT scans acquired at 20.22 $μ$m per voxel resolution. A tailored 3D{-}UNet enabled semantic segmentation of matrix, weft, and fill phases across compaction stages corresponding to fiber volume contents of 50--60 %. The model achieved a minimum mean Intersection-over-Union of 0.822 and an $F1$ score of 0.902. Spatial structure was subsequently analyzed using the two-point correlation function $S_2$, allowing for probabilistic extraction of average layer thickness and nesting degree. The results show strong agreement with micrograph-based validation. This methodology provides a robust approach for extracting key geometrical features from industrially relevant CT data and establishes a foundation for reverse modeling and descriptor-based structural analysis of composite preforms.

Figures (8)

• Figure 1: Photograph of a plain-weave carbon fiber fabric sample and its corresponding idealized unit cell geometry modeled in TexGen long_modelling_2011 (dimensions are in mm).
• Figure 2: Experimental set-up (left) and drawings of the load train (right) in situ CT test set-up (all dimensions are in mm)
• Figure 3: Schematic illustration of 3D-UNet's architecture for semantic segmentation of textile reinforcement as a three-class semantic segmentation problem (Illustration was prepared using code from iqbal_harisiqbal88plotneuralnet_2018)
• Figure 4: Illustration of compactation force $F$ and tamp gap $\delta$ over the measuring time $t$ as well as center slice of the shape $[600,1,256]$ of each CT-stage with estimated FVC $\phi=$[]<30;50;55;60
• Figure 5: Prediction results on a 10-layer evaluation dataset (stage IV) at the center $xz$-slice (cropped in $z$ for clarity): \ref{['fig:eval_prediction_results:a']},\ref{['fig:eval_prediction_results:c']} show predicted probabilities $p(c)$ for classes 1 and 2; \ref{['fig:eval_prediction_results:b']},\ref{['fig:eval_prediction_results:d']} show differences to the ground truth probability $\Delta p(c)$ for each class; all results are overlayed to the raw CT-data ;\ref{['fig:eval_prediction_results:e']} displays the segmented volume with the highlighted $xz$-slice