Topology-Aware Learning for Routing In Satellite–Terrestrial Integrated Networks: A Review of Graph Neural Network and Reinforcement Learning Approaches

Abstract

Satellite–Terrestrial Integrated Networks (STINs) are a key part of 5G-Advanced and new 6G non-terrestrial networks. They connect the world by combining LEO, MEO, and GEO satellite constellations with ground-based infrastructure. But routing is very hard because of highly dynamic topologies, different link characteristics, and large-scale networks. This makes traditional protocols and topology-agnostic learning methods less useful. This review analyses topology-aware learning-based routing for STINs, concentrating on Graph Neural Networks (GNNs) and hybrid GNN–Reinforcement Learning (GNN–RL) frameworks. By modelling STINs as graphs that change over time, these methods clearly show how relationships and multi-hop interactions work, which are important for routing that can grow and change. Comprehensive analyses are conducted on classical routing, non-topology-aware reinforcement learning, purely GNN-based methodologies, and hybrid GNN–RL architectures, emphasising their merits and drawbacks in dynamic satellite–terrestrial contexts. We also look at hierarchical and multi-agent extensions, as well as current datasets and evaluation methods. Finally, important open problems related to scalability, non-stationarity, and real-world use are found, and future research directions that fit with new 6G non-terrestrial network standards are laid out.