The increasing demand for interconnected satellite networks is driving the development of space data centers, but limited bandwidth and unreliable communication links pose significant challenges to efficiently training machine learning models across these systems. Anbang Zhang of Shandong University, Chenyuan Feng of the University of Exeter, and Wai Ho Mow of the Hong Kong University of Science and Technology, along with colleagues, are tackling this problem by introducing OptiVote, a new communications framework that leverages free-space optical technology. Their work demonstrates how to aggregate data without requiring precise signal conditioning, a major hurdle in dynamic space environments, and accomplishes this through a noncoherent approach that combines majority voting principles and pulse position modulation. By eliminating strict phase requirements and incorporating a dynamic power control scheme, OptiVote significantly improves the resiliency and efficiency of distributed federated learning in future space data centers, paving the way for more powerful on-orbit computing capabilities.
Distributed computing and learning infrastructures present unique challenges, especially when enabling federated learning. Iterative training requires frequent aggregation of data across intersatellite links, which are inherently bandwidth and energy constrained and subject to highly dynamic conditions. Therefore, communication-efficient aggregation is essential to achieve scalable on-orbit intelligence. This study explores the potential of wireless computing (AirComp) as an in-network aggregation primitive to address these limitations. Traditional AirComp requires precise phase alignment, which poses great challenges in the space environment due to satellite jitter and Doppler effects. Our research focuses on overcoming this limitation and paving the way for robust and efficient distributed learning in space.
Federated Learning with Wireless Computing in LEO
This study details a system of federated learning (FL) in low earth orbit (LEO) satellite networks using over-the-air computing (OATC). The core idea is to train machine learning models across a distributed network without directly exchanging training data, preserving privacy. Instead of transmitting model updates digitally, the satellite simultaneously transmits analog signals representing the updates. This signal interferes constructively at the central receiver, effectively performing a weighted sum of updates. This process is much faster than traditional digital communication. The system leverages the LEO satellite constellation to provide connectivity and computational resources, and employs the SignSGD algorithm to reduce communication overhead using the sign of the gradient. Digital twins, which are virtual replicas of physical networks, are 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 that overcomes significant challenges posed by satellite network bandwidth limitations and dynamic link conditions. This system leverages over-the-air computation (AirComp) to efficiently aggregate data within a network, but largely addresses the challenge of maintaining accurate phase alignment in a space environment, a requirement of traditional AirComp techniques. OptiVote achieves robust aggregation through a non-coherent free-space optical (FSO) AirComp system that integrates coded stochastic gradient descent with majority voting principles and pulse position modulation. In this method, each satellite conveys a local gradient code by activating orthogonal time slots using pulsed position modulation, allowing aggregation nodes to combine optical intensities rather than relying on sensitive phase synchronization.
The researchers developed a dynamic power control scheme that balances the received energy from each satellite and reduces the bias introduced by disparate FSO channels without the need for additional signaling. Theoretical analysis characterizes the total error probability in the statistical FSO channel and confirms the convergence guarantee for non-convex optimization problems. The test demonstrates OptiVote's scalability and resilience for on-orbit intelligence applications, providing a practical solution for distributed learning on 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 that addresses the challenges of bandwidth and energy constraints in satellite communications. The system utilizes wireless computation with free-space optical links and employs a non-coherent approach that eliminates the need for precise phasing, which is a major challenge in dynamic space environments. OptiVote integrates code-based gradient descent with majority aggregation principles and pulse position modulation to achieve simultaneous data aggregation with energy storage. The team further enhanced OptiVote with a dynamic power control scheme designed to balance received energy across disparate communication links without the need for additional signaling. Theoretical analysis confirms the framework's ability to minimize errors under realistic space communication conditions and ensures convergence of complex non-convex learning objectives. Tests demonstrate that OptiVote consistently improves learning accuracy and communication efficiency compared to existing methods, paving the way for scalable and resilient on-orbit intelligence.
👉 More information
🗞 OptiVote: Non-coherent FSO wireless majority voting for communication-efficient distributed federated learning in space data centers
🧠ArXiv: https://arxiv.org/abs/2512.24334
