Space Data Centers Achieve Communication Efficiency with OptiVote and Federated Learning

The increasing demand for interconnected satellite networks fuels the development of space data centres, but efficiently training machine learning models across these systems presents significant challenges due to limited bandwidth and unreliable communication links. Anbang Zhang from Shandong University, Chenyuan Feng from the University of Exeter, and Wai Ho Mow from The Hong Kong University of Science and Technology, along with colleagues, address this problem by introducing OptiVote, a novel communication framework that leverages free-space optical technology. Their research demonstrates a method for aggregating data without the need for precise signal alignment, a major hurdle in the dynamic space environment, and achieves this through a non-coherent approach combining a majority-vote principle with pulse-position modulation. By eliminating stringent phase requirements and incorporating a dynamic power control scheme, OptiVote significantly improves the resilience and efficiency of distributed federated learning in future space data centres, paving the way for more powerful in-orbit computing capabilities.

Distributed computing and learning infrastructures present unique challenges, particularly when enabling federated learning. Iterative training demands frequent aggregation of data across inter-satellite links, which are inherently bandwidth and energy constrained, and subject to highly dynamic conditions. Consequently, communication-efficient aggregation is crucial for achieving scalable in-orbit intelligence. This work explores the potential of over-the-air computation (AirComp) as an in-network aggregation primitive to address these limitations. Conventional AirComp requires precise phase alignment, a significant difficulty in space environments due to satellite jitter and Doppler effects, and the research focuses on overcoming this limitation, paving the way for robust and efficient distributed learning in space.

Federated Learning via Over-the-Air Computation in LEO

This research details a system for Federated Learning (FL) in a Low Earth Orbit (LEO) satellite network, utilizing Over-the-Air Computation (OATC). The core idea is to train a machine learning model across a decentralized network without directly exchanging training data, preserving privacy. Instead of transmitting model updates digitally, satellites simultaneously transmit analog signals representing their updates, which interfere constructively at a central receiver, effectively performing a weighted sum of the updates, a process much faster than traditional digital communication. The system leverages LEO satellite constellations to provide connectivity and computational resources, employing the SignSGD algorithm to reduce communication overhead by using the sign of the gradient. A digital twin, a virtual replica of the physical network, is used for simulation and optimization.

OptiVote Enables Robust Federated Learning in Space

Scientists have developed OptiVote, a new framework for federated learning in space data centers, overcoming significant challenges posed by bandwidth limitations and dynamic link conditions in satellite networks. This system exploits over-the-air computation (AirComp) for efficient in-network aggregation of data, but crucially addresses the difficulty of maintaining precise phase alignment in the space environment, a requirement of traditional AirComp methods. OptiVote achieves robust aggregation through a non-coherent free-space optical (FSO) AirComp system, integrating sign stochastic gradient descent with a majority-vote principle and pulse-position modulation. The method involves each satellite conveying its local gradient signs by activating orthogonal time slots using pulse-position modulation, allowing the aggregating node to combine optical intensity rather than relying on sensitive phase synchronization.

Researchers developed a dynamic power control scheme that balances received energies from each satellite, mitigating bias introduced by heterogeneous FSO channels without requiring additional signaling. Theoretical analysis characterizes the aggregate error probability under statistical FSO channels, confirming convergence guarantees for non-convex optimization problems. Testing demonstrates OptiVote’s scalability and resilience for in-orbit intelligence applications, delivering a practical solution for distributed learning in communication-constrained satellite networks and paving the way for advanced data processing and analysis directly in space.

OptiVote, Federated Learning in Space

Researchers have developed OptiVote, a new framework for federated learning in space data centers, addressing the challenges of bandwidth and energy constraints in satellite communications. This system utilizes over-the-air computation with free-space optical links, employing a non-coherent approach that eliminates the need for precise phase alignment, a significant difficulty in the dynamic space environment. OptiVote integrates a sign-based gradient descent method with a majority-vote aggregation principle and pulse-position modulation to achieve simultaneous data aggregation through energy accumulation. The team further enhanced OptiVote with a dynamic power control scheme, designed to balance received energies across heterogeneous communication links without requiring additional signaling. Theoretical analysis confirms the framework’s ability to minimize errors under realistic space communication conditions and guarantees convergence for complex, non-convex learning objectives. Testing demonstrates that OptiVote consistently improves learning accuracy and communication efficiency compared to existing methods, paving the way for scalable and resilient in-orbit intelligence.