Bhargav Joshi, Digantara; Subramanian Arumugam, Digantara; Siri L., Digantara; Thamim Ansari, Digantara; Tanveer Ahmed, Digantara
Keywords: SSA, SDA, Breakup Events, Space Debris, AI, Graph Neural Networks (GNN)
Abstract:
With the increasing population of RSOs around Earth, collision risks have heightened significantly. In addition, variations in safety standards, manufacturing quality, and end-of-life management have increased the risk of explosions and fragmentation, as some satellites lack fail-safes or proper passivation. Accurate and timely classification of such events is essential for Space Domain Awareness (SDA) to assess risks and mitigate potential threats to other Resident Space Objects (RSOs). Traditional heuristic approaches, such as rule-based classification and manual feature engineering, face significant challenges in handling the complexity and dynamics of debris fields generated by breakup events. These approaches are often inefficient, requiring extensive fine-tuning and struggle to adapt to dynamic scenarios with a large and fluctuating number of fragments, leading to increased false positives, reduced accuracy, and increased computational time.
To overcome these challenges, a novel, efficient, and accurate model leveraging a Graph Neural Network (GNN) is proposed, tailored for classifying such breakup events. The model utilizes orbital elements of the fragments as an input. The GNN represents these inputs as nodes, while edges capture relationships such as velocity correlation and spatial proximity. For training, synthetic breakup events were generated by NASA’s EVOLVE 4.0 simulation tool, modeling collision or explosion scenarios based on initial state vectors and physical attributes such as mass and surface area. Satellite objects are assigned randomized mass and area to generalize the characteristics of realistic breakup scenarios, while the minimum characteristic length threshold for debris generation is set at 0.1 m, aligning with cataloged objects. Random fragments were chosen from each event to make the model outcome independent of the total number of debris pieces generated in the breakup events. The model parses the data through a combination of Graph Convolution Layers and Feed Forward layers and outputs a score ranging from 0 to 1, indicating whether an event is classified as an explosion or a collision respectively.
Preliminary testing in the Low Earth Orbit (LEO) regime demonstrated that the GNN-based model maintains high accuracy while minimizing false positives, even when faced with variations in debris kinematics and fragmentation patterns. Its ability to adapt to debris fields with varying densities and object sizes marks a significant improvement over fixed-rule classification methods, which often struggle under such conditions. The study further investigates the impact of the number of debris objects included in the input on the model’s performance across different orbital regimes.
Automating breakup classification with a GNN-based model enhances SDA by enabling fast, accurate, and scalable identification of collision and explosion events. This reduces reliance on manual analysis and improves real-time decision-making. Furthermore, it enhances debris mitigation strategies, contributing to a safer and more sustainable space environment. The insights from this study lay the groundwork for integrating advanced AI-driven classification models into broader SDA frameworks, ensuring more effective monitoring and management of the increasingly congested orbital environment.
Date of Conference: September 16-19, 2025
Track: Space Debris