Debonding detection on CFRP/Al-Honeycomb sandwich structures using Geometry-Aware GNNs
Quick Start โข Features โข Pipeline โข Citation โข ๐ Wiki โข Contributing
GNN-SHM is a research project that combines Graph Neural Networks (GNN) with Finite Element Method (FEM) to detect and localize skin-core debonding in the JAXA H3 rocket's CFRP/Aluminum Honeycomb payload fairing.
ๆฅๆฌ่ช: JAXA H3 ใญใฑใใใฎ CFRP ใใใซใ ใตใณใใคใใใใงใขใชใณใฐใซใใใฆใGNN ใจ FEM ใ็ตฑๅใใในใญใณ-ใณใข็้ขใใใณใใฃใณใฐไฝ็ฝฎ็นๅฎใทในใใ ใ้็บใ2025ๅนด F8 ไบๆ ใง้กๅจๅใใ CFRP/Al-HC ๆฅ็ๅฅๅ จๆงใขใใฟใชใณใฐใฎๅฎ็จๅใ็ฎๆใใ
| H-IIA/B, Epsilon (Legacy) | H3 (This Project) | |
|---|---|---|
| Skin | Al 7075 | CFRP (T1000, AFP) |
| CTE Mismatch | โ0 | Severe (โ0.3 vs 23 ร10โปโถ/ยฐC) |
| SHM Need | Low (40yr mature) | High (F8 accident, 2025) |
The H3 F8 accident (Dec 2025) identified CFRP/Al-Honeycomb interface debonding as a likely cause. This project aims to enable Condition-Based Maintenance (CBM) via guided-wave SHM + GNN-based defect localization.
H3 Rocket Test Flight No.2 Lift-off (Source: JAXA Digital Archives)
- Geometry-Aware Graph Construction: Surface normals, principal curvature, geodesic distance โ no UV-mapping distortion
- 4 GNN Architectures: GCN, GAT, GIN, GraphSAGE with Focal Loss for class imbalance
- H3-Spec FEM: Barrel + Ogive (ฯ5.2m), thermal load (CTE mismatch), debonding defects
- Cutting-Edge ML Roadmap: Graph Mamba, E(3)-Equivariant GNN, FNO surrogate, PINN
- Multi-Class Target: debond / delam / impact / healthy (2-year roadmap)
- JAXA Collaboration: Real PSS test data validation planned
flowchart LR
subgraph DataGen["๐ฆ Data Generation"]
DOE[generate_doe.py<br/>DOE params]
BATCH[run_batch.py<br/>Abaqus FEM]
DOE --> BATCH
end
subgraph FEM["๐ฌ FEM"]
ABAQUS[Abaqus FEM<br/>H3 Fairing<br/>Thermal + 120ยฐC<br/>Debonding]
end
subgraph Graph["๐ Graph"]
EXTRACT[extract_odb<br/>CSV]
BUILD[prepare_ml_data<br/>Curvature-Aware Graph]
EXTRACT --> BUILD
end
subgraph Train["๐ง GNN Training"]
GNN[GCN / GAT / GIN / SAGE<br/>Focal Loss]
end
subgraph Deploy["๐ Inference"]
INFER[Checkpoint]
HEATMAP[Defect Prob<br/>Heatmap]
API[FastAPI<br/>REST API]
INFER --> HEATMAP --> API
end
BATCH --> ABAQUS
ABAQUS --> EXTRACT
BUILD --> GNN
GNN --> INFER
# Clone
git clone https://github.com/keisuke58/Payload_gnn.git
cd Payload_gnn
# Install
pip install -r requirements.txt
# Train (existing data)
python src/train.py --arch gat --epochs 200 --cross_val 5
# Inference API
MODEL_CHECKPOINT=runs/<run>/best_model.pt uvicorn src.predict_api:app --port 8000python src/generate_doe.py --n_samples 50 --output doe.json
python src/run_batch.py --doe doe.json --output_dir dataset_output
python src/build_graph.py --data_dir dataset_output
python src/train.py --arch gat --epochs 200Payload2026/
โโโ src/ # Core pipeline
โ โโโ generate_fairing_dataset.py # Abaqus FEM
โ โโโ build_graph.py # Curvature-aware graph
โ โโโ train.py # GNN training
โ โโโ predict_api.py # FastAPI inference
โโโ wiki_repo/ # ๐ Full documentation
โ โโโ Home.md # Wiki index
โ โโโ Cutting-Edge-ML.md # Graph Mamba, Equivariant GNN
โ โโโ Vocabulary.md # Technical glossary
โ โโโ ...
โโโ .github/
โ โโโ ISSUES.md # Task index
โ โโโ ...
โโโ requirements.txt
| Item | Value |
|---|---|
| Graphs | 101 (train 81 + val 20) |
| Nodes/graph | ~10,897 |
| Node features | 16 (normal, curvature, stress, temp) |
| Edge features | 5 |
Defect size distribution: Small 30%, Medium 40%, Large 25%, Critical 5%.
**โ Full visualization (docs/DATASET_VISUALIZATION.md)
| Summary | Feature Distributions |
|---|---|
| Resource | Description |
|---|---|
| Wiki Home | Full project index, status, navigation |
| 2-Year Goals | 5K samples, 4-class, Sim-to-Real |
| Cutting-Edge ML | Graph Mamba, Equivariant GNN, FNO, PINN |
| Vocabulary | Technical terms (ENโJP) |
| Publication Venues | IWSHM, JSASS, Structural Health Monitoring journal |
We welcome contributions! See CONTRIBUTING.md for guidelines.
- ๐ Report a bug
- ๐ก Request a feature
If you use this work in your research, please cite:
@software{payload_gnn_2026,
title = {GNN-SHM: Graph Neural Networks for H3 Rocket Fairing Structural Health Monitoring},
author = {Payload2026 Contributors},
year = {2026},
url = {https://github.com/keisuke58/Payload_gnn}
}This project is licensed under the MIT License - see LICENSE for details.
- JAXA โ H3 specifications, collaboration
- Open Guided Waves โ Benchmark dataset
- PyTorch Geometric โ GNN framework
If this project helps your research, please consider giving it a โญ