Guide to 3D Defect Localization Technology for Carbon Fiber Composite Materials Based on Graph Neural Networks

This article introduces a research project for the WCCM/ECCOMAS 2026 conference, which combines Finite Element Analysis (FEA) with Graph Neural Networks (GAT/GATv2) to achieve 3D defect localization in carbon fiber composite material (CFRP) segment structures with holes. Using the differential normalization DSPSS method, it provides an intelligent solution for non-destructive testing. Original Author/Maintainer: keisuke58 Source Platform: GitHub Original Title: wccm2026-cfrp-gnn Original Link: https://github.com/keisuke58/wccm2026-cfrp-gnn Publication Date: June 7, 2026 Related Paper: Nishioka et al., Frontiers in Materials 12, 1652484 (2025), DOI:10.3389/fmats.2025.1652484

Research Background and Problem Definition

Carbon Fiber Reinforced Polymer (CFRP) is widely used in aerospace, automotive manufacturing and other fields due to its excellent specific strength and stiffness. However, internal defects such as delamination and debonding are prone to occur during manufacturing and use, affecting structural safety. Traditional non-destructive testing (NDT) is difficult to accurately locate the 3D spatial position of defects. This study focuses on CFRP segment structures with holes (key load-bearing components of rockets and aircraft, where stress concentration around holes is prone to defects). The core challenge is to infer the position and depth of internal defects using only surface stress measurement data (simulating infrared stress measurement).

Core Technical Methods and Workflow

Finite Element Analysis (FEA) Data Generation

Generate a large amount of labeled data by simulating the mechanical response of CFRP structures with holes under different defect configurations (position, depth, shape) via FEA, avoiding expensive physical experiments.

Graph Neural Network Architecture

Adopt GAT/GATv2: Discretize the CFRP surface into graph nodes, with features as Differential Shear Stress in Principal Stress System (DSPSS) measurement values, and edges representing spatial adjacency relationships. The self-attention mechanism can learn position correlation, handle non-uniform stress fields, and capture local effects of defects.

Differential Normalization Strategy

• Differential Normalization: The difference between measured stress and defect-free baseline is normalized by z-score to offset stress concentration around holes and highlight defect disturbances;
• Original Normalization: Directly normalize DSPSS data without a baseline (closer to engineering applications); Experiments show that differential normalization has higher accuracy, and original normalization combined with geometric features (r, θ) can be used as an alternative.

Model Architecture and Training Optimization Strategies

Classification Task Definition

Defect localization is defined as a 19-class classification problem: 1 class for no defects, and 18 classes corresponding to defect positions in different depth layers (2 layers) and in-plane regions, converting continuous spatial localization into a classification task.

Network Variants and Extensions

Compare multiple GNN architectures: GAT/GATv2, MeshGraphNet (encode-process-decode, edge update), Transformer, SAGE, etc.

Geometric Feature Enhancement

Introduce radial distance r=√(x²+y²) and angle θ to help the model understand the geometric impact of holes and improve localization accuracy.

Training Regularization Techniques

• DropEdge: Randomly remove edges to prevent overfitting;
• Training Noise Injection: Add Gaussian noise to simulate measurement errors;
• Noise Curriculum Learning: Gradually increase noise intensity;
• Label Smoothing: Handle misclassification of adjacent layers;
• Mirror Augmentation: Left-right flipping to expand data and solve regional asymmetry.

Experimental Results and Performance Analysis

Baseline Performance

The pre-trained model on the validation set has a Macro F1 of 0.730, and about 0.66 on the test set, indicating overfitting and the need to improve generalization ability.

Comparison with Previous Work

Compared with the Frontiers paper (defect-free specimens, no differential normalization, Macro F1=0.61, TDPS=0.70), this study targets complex scenarios with holes and uses more advanced normalization and architecture.

Ablation Experiments

Verify the contribution of each component: differential vs. original normalization, impact of geometric features, performance of different GNN architectures, effects of noise injection and curriculum learning.

Engineering Application Prospects and Research Outlook

Engineering Application Prospects

• NDT Integration: Integrate with infrared thermal stress measurement to infer internal defects through surface stress, avoiding the high cost of X-ray or ultrasonic C-scan;
• Practical Deployment Considerations: Differential method requires a defect-free baseline, while original normalization does not; noise injection improves robustness; GNN forward propagation is fast, suitable for online detection.

Research Limitations and Future Directions

• Limitations: Training data only comes from FEA simulations, not verified with real experimental data; large performance gap between test set and validation set; simplified defect shape assumptions;
• Future: Integrate experimental infrared data for verification; model complex defect geometry; uncertainty quantification; extend to other composite materials and structures.

Research Summary and Outlook

This study combines physical simulation (FEA) with deep learning (GNN) to achieve 3D localization of internal defects in CFRP using only surface stress measurements, improving the automation of detection and providing a technical foundation for real-time health monitoring. With the maturity of deep learning and the popularization of computing resources, such intelligent detection technologies are expected to play an important role in aerospace, wind turbine blade monitoring and other fields.