CroSatFL: Energy-Efficient Federated Learning with Cross-Aggregation for Satellite Edge Computing

A research team led by Nan Yang has developed CroSatFL, a breakthrough framework for running federated learning (FL) on satellite constellations while dramatically reducing energy consumption. The system addresses the fundamental challenge of training AI models in space where power budgets are strict and connectivity is dynamic. By keeping the ground station out of the iterative training loop, CroSatFL requires only two communication phases with Earth—one for initialization and one for final model collection—instead of constant back-and-forth exchanges.

CroSatFL integrates three innovative mechanisms: StarMask forms clusters based on laser inter-satellite link feasibility and matches data volume to heterogeneous onboard hardware, Skip-One mitigates slow satellites by allowing clusters to skip at most one straggler per round, and random-k cross-aggregation enables lightweight mixing between clusters. Using a realistic Walker-Delta constellation simulation, the team demonstrated that CroSatFL reduces ground station communication count by over 100x while maintaining competitive model accuracy and achieving faster convergence than traditional approaches.

The framework represents a significant step toward sustainable AI in space, enabling continuous model training across satellite networks without draining limited power resources. This could enable real-time Earth observation analytics, space-based IoT device management, and autonomous satellite swarm coordination—all while minimizing the energy footprint of space-based computing infrastructure.