Dynamic Predictive Routing for Satellite Megaconstellations

ABSTRACT Emerging low‐Earth orbit (LEO) satellite mega‐constellations introduce new challenges in traffic engineering due to their scale and highly dynamic topologies. This paper presents dynamic predictive routing (DPR), a machine learning approach that leverages graph neural networks (GNNs) and gradient‐free optimization to enhance routing performance in constellations with intersatellite links (ISLs). Unlike traditional heuristics such as Dijkstra or Bellman–Ford, which react to current topology states, DPR anticipates short‐term network evolution and proactively redistributes traffic to reduce congestion. Through simulation and emulation, we show that DPR outperforms baseline routing, achieving up to ~9% congestion reduction in highly loaded networks, with scalability to larger constellations. A GPU‐optimized path algorithm further accelerates computation, enabling feasibility in resource‐constrained environments. The vision at the heart of all mega‐constellations is a very wide area network (VWAN) with latency competitive to terrestrial fiber. While routing challenges exist in any mesh network, they are particularly critical in LEO systems, where orbital dynamics, scarce spectrum, and limited gateways intensify performance pressure. As capacity becomes a premium resource, dynamic optimization is essential to extract maximum value from the orbital real estate. An intelligent routing fabric can unlock greater throughput without proportional hardware expansion, support differentiated services, and integrate operational constraints directly into planning and decision‐making. In this paper, we describe the training methodology, including live interaction with simulated and emulated networks, and present results showing consistent congestion reduction, modest latency trade‐offs, and scalability. Importantly, the alignment of performance gains across simulation and emulation provides a strong basis for operational deployment. Further optimization remains feasible when tailoring DPR to specific network constraints.