Autonomic GNN-Based Shortest-Path Routing in LEO Networks via Imitation and Transfer Learning

Low Earth Orbit (LEO) satellite networks represent a key enabler for global connectivity, yet their highly dynamic topologies severely challenge classical routing algorithms that require static tables or frequent global recomputation. To enable continuous self-adaptation, we propose an autonomic, data-driven routing framework based on Graph Neural Networks (GNNs). Adopting an imitation learning approach, we formulate decentralized shortest-path routing as a supervised edge classification task: given a network graph, the model autonomously infers whether each inter–satellite link belongs to the optimal path. To overcome data scarcity and ensure resilient generalization across changing topologies, we introduce a transfer learning pipeline that pre-trains on synthetic spatial graphs before fine-tuning on real Iridium NEXT topologies. Furthermore, replacing 2D geodetic coordinates with a 3D Earth-Centered, Earth-Fixed (ECEF) representation eliminates wrap-around artifacts, substantially improving spatial reasoning. The best model identifies the shortest path in 97.1% of cases, while overall identifying a valid source-to-destination routing in 99.0% on unseen test sets, successfully enabling self-organizing routing.