🚀 GNN-SHM: Graph Neural Networks for H3 Rocket Fairing Structural Health Monitoring

Debonding detection on CFRP/Al-Honeycomb sandwich structures using Geometry-Aware GNNs

Quick Start • Features • Pipeline • Citation • 📚 Wiki • Contributing

---

🌟 Overview

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.

日本語: JAXA H3 ロケットの CFRP ハニカムサンドイッチフェアリングにおいて、GNN と FEM を統合したスキン-コア界面デボンディング位置特定システムを開発。2025年 F8 事故で顕在化した CFRP/Al-HC 接着健全性モニタリングの実用化を目指す。

Why This Matters

	H-IIA/B, Epsilon (Legacy)	H3 (This Project)
Skin	Al 7075	CFRP (T1000, AFP)
CTE Mismatch	≈0	Severe (−0.3 vs 23 ×10⁻⁶/°C)
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 Launch (JAXA)

H3 Rocket Test Flight No.2 Lift-off (Source: JAXA Digital Archives)

---

✨ Features

• Geometry-Aware Graph Construction: Surface normals, principal curvature, geodesic distance — no UV-mapping distortion
• 4 GNN Architectures: GCN, GAT, GIN, GraphSAGE with Focal Loss for class imbalance
• H3-Spec FEM: Barrel + Ogive (φ5.2m), 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

---

🏗 Pipeline

flowchart LR
    subgraph DataGen["📦 Data Generation"]
        DOE[generate_doe.py<br/>DOE params]
        BATCH[run_batch.py<br/>Abaqus FEM]
        DOE --> BATCH
    end

    subgraph FEM["🔬 FEM"]
        ABAQUS[Abaqus FEM<br/>H3 Fairing<br/>Thermal + 120°C<br/>Debonding]
    end

    subgraph Graph["📊 Graph"]
        EXTRACT[extract_odb<br/>CSV]
        BUILD[prepare_ml_data<br/>Curvature-Aware Graph]
        EXTRACT --> BUILD
    end

    subgraph Train["🧠 GNN Training"]
        GNN[GCN / GAT / GIN / SAGE<br/>Focal Loss]
    end

    subgraph Deploy["🚀 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

Loading

---

🚀 Quick Start

# 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 8000

Full Pipeline (with Abaqus)

python 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 200

---

📁 Project Structure

Payload2026/
├── src/                    # Core pipeline
│   ├── generate_fairing_dataset.py   # Abaqus FEM
│   ├── build_graph.py               # Curvature-aware graph
│   ├── train.py                     # GNN training
│   └── predict_api.py              # FastAPI inference
├── wiki_repo/              # 📚 Full documentation
│   ├── Home.md             # Wiki index
│   ├── Cutting-Edge-ML.md  # Graph Mamba, Equivariant GNN
│   ├── Vocabulary.md       # Technical glossary
│   └── ...
├── .github/
│   ├── ISSUES.md           # Task index
│   └── ...
└── requirements.txt

---

📊 Dataset

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%.

Dataset Progress Visualization

**→ Full visualization (docs/DATASET_VISUALIZATION.md)

Summary	Feature Distributions

---

📚 Documentation

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 (EN↔JP)
Publication Venues	IWSHM, JSASS, Structural Health Monitoring journal

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

• 🐛 Report a bug
• 💡 Request a feature

---

📄 Citation

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}
}

---

📜 License

This project is licensed under the MIT License - see LICENSE for details.

---

🙏 Acknowledgments

• JAXA — H3 specifications, collaboration
• Open Guided Waves — Benchmark dataset
• PyTorch Geometric — GNN framework

---

If this project helps your research, please consider giving it a ⭐