Graph Neural Networks Empower Structural Health Monitoring of Rocket Fairings: JAXA H3 Rocket CFRP/Aluminum Honeycomb Debonding Defect Detection System

This article introduces an open-source project called Payload_gnn, which combines Graph Neural Networks (GNN) and Finite Element Method (FEM) to detect skin-core interface debonding defects in the carbon fiber composite (CFRP) fairing of JAXA's H3 rocket. Developed by the keisuke58 team and released on GitHub under the MIT license, this project provides an intelligent solution for spacecraft Structural Health Monitoring (SHM). Its core innovation lies in the detection scheme that combines geometry-aware graph construction with FEM simulation data.

In the December 2025 accident of JAXA's H3 rocket F8 mission, interface debonding between CFRP and aluminum honeycomb sandwich structure was listed as one of the possible main causes, highlighting the urgent need for SHM of composite materials. The H3 rocket uses T1000-grade CFRP skin (automated fiber placement process) instead of traditional aluminum alloy, which reduces weight but introduces a CTE mismatch issue: the CTE of CFRP is approximately -0.3×10⁻⁶/°C, while that of aluminum honeycomb is 23×10⁻⁶/°C, making it prone to debonding defects under thermal cycle loads.

The core technical solution of the project is the combination of GNN and FEM:

1. Geometry-Aware Graph Construction: To avoid UV mapping distortion, graph structures are built using 3D geometric information (surface normal vectors, principal curvature, geodesic distance). Each graph contains approximately 10,897 nodes (16-dimensional features) and 5-dimensional edge features;
2. FEM Simulation Data Generation: Simulations are performed using Abaqus based on the real specifications of the H3 fairing (5.2m diameter cylindrical section + oval head). The load condition includes a maximum thermal load of 120°C. A dataset of 101 samples (81 for training, 20 for validation) is generated via DOE, covering four defect size levels: small, medium, large, and critical.

Model Architecture & Training Strategy:

• Implements four GNN architectures for comparison: GCN (spectral convolution), GAT (attention mechanism), GIN (graph isomorphism network), GraphSAGE (inductive learning);
• Uses Focal Loss to address class imbalance (healthy areas are far more than defect areas), improving the detection sensitivity of small defects;
• Supports 5-fold cross-validation for training, with 200 training epochs, based on the PyTorch Geometric framework.

Inference Deployment & Application Interface:

• Provides model checkpoints to save training parameters;
• Generates defect probability heatmaps to visualize detection results;
• Develops a FastAPI REST interface for easy integration into existing monitoring systems, laying the foundation for data validation of JAXA's Ground Test System (PSS).

Technical Roadmap & Cutting-Edge Exploration: Short-Term Goals: Expand the dataset to 5,000 samples, implement multi-class damage detection (debonding/ delamination/ impact damage/ healthy), and complete validation with real PSS data; Cutting-Edge Methods: Explore directions such as Graph Mamba (long-range dependencies), E(3)-equivariant GNN (spatial equivariance), FNO (FEM surrogate replacement), and PINN (integration of physical constraints).

Engineering Significance & Outlook:

• Addresses issues like delayed detection and low positioning accuracy in traditional SHM, enabling Condition-Based Maintenance (CBM), fast damage localization (in seconds), and full lifecycle management;
• The open-source nature (MIT license) promotes industry-university-research collaborative innovation;
• With the advancement of the H3 rocket project and validation with real data, it is expected to move from laboratory to engineering practice, providing a reference for the field of structural health monitoring.