GTopOpt: Gradient-Based Topology Optimization for Satellite Constellations

The growing number of satellites in Geostationary (GEO) and Low Earth Orbit (LEO) requires efficient network management for dynamic topology configuration. We propose a new satellite network optimization framework that optimizes topology over end-to-end latency, in the context of both propagation delay and hop count. This is normally a computationally intense network optimization problem, attributed to the binary nature of connection values, the dimensionality of the possible solution space, and the need for a globally connected topology. We develop Gradient Topology Optimization (GTopOpt), an end-to-end differentiable system utilizing auxiliary logits and functional constraints, enabling optimization of inter-satellite links (ISLs) with respect to the end-to-end objective. A key feature of our solution is a temperature-based cooling method, allowing initially for relaxed constraints and then tightening them over the course of the gradient descent procedure. Comparison against baselines demonstrates a 60% reduction in weighted number of hops in larger constellations and 30% reduction in end-to-end latency in smaller constellations.