Publications Conference paper · Orbital computing · 2025

FedSat-LAM: Enabling Large AI Models on Resource-Constrained Satellites via Hierarchical Federated Learning

Hannaneh B Pasandi, Zhaowei Tan, Mohammad Hosseini, Sina Darabi, Franck Rousseau, and Juan A. Fraire

ACM International Workshop on Edge and Mobile Foundation Models (EdgeFM), pp. 19–24, 2025

Abstract

LEO satellites generate 5.2 PB/day but can only downlink 2.3 PB/day, leaving 57% of data inaccessible. We present FedSat-LAM, enabling 632M-parameter foundation models on heterogeneous satellite constellations through: (1) multi-hop offloading reducing latency 41%, (2) tree-based aggregation achieving O(log N) complexity, and (3) PEFT training only 0.13% of parameters at 10.45 J/sample. Validated on 100-satellite simulations and Jetson hardware, we achieve 87.6% accuracy, 5.2% above FedProx—while cutting communication 85% and energy 95%.

Cite

DOI 10.1145/3737902.3768355

BibTeX

@inproceedings{pasandi2025fedsat,
  author    = {Pasandi, Hannaneh B and Tan, Zhaowei and Hosseini, Mohammad and Darabi, Sina and Rousseau, Franck and Fraire, Juan A.},
  title     = {FedSat-LAM: Enabling Large AI Models on Resource-Constrained Satellites via Hierarchical Federated Learning},
  booktitle = {ACM International Workshop on Edge and Mobile Foundation Models (EdgeFM)},
  pages     = {19--24},
  publisher = {ACM},
  year      = {2025},
  doi       = {10.1145/3737902.3768355}
}

More on orbital computing

Dirty Bits in Low-Earth Orbit: The Carbon Footprint of Launching Computers Robin Ohs, Gregory F Stock, Andreas Schmidt, Juan A. Fraire, and Holger Hermanns ACM SIGENERGY Energy Informatics Review, 2025
Abstract
Low-Earth Orbit (LEO) satellites are increasingly proposed for communication and in-orbit computing, achieving low-latency global services. However, their sustainability remains largely unexamined. This paper investigates the carbon footprint of computing in space, focusing on lifecycle emissions from launch over orbital operation to re-entry. We present ESpaS, a lightweight tool for estimating carbon intensities across CPU usage, memory, and networking in orbital vs. terrestrial settings. Three worked examples compare (i) launch technologies (state-of-the-art rocket vs. potential next generation) and (ii) operational emissions of data center workloads in orbit and on the ground. Results show that, even under optimistic assumptions, in-orbit systems incur significantly higher carbon costs—up to an order of magnitude more than terrestrial equivalents—primarily due to embodied emissions from launch and re-entry. Our findings advocate for carbon-aware design principles and regulatory oversight in developing sustainable digital infrastructure in orbit.

arXiv
On the Latency Trade-off Between Space and Terrestrial Clouds in Non-Terrestrial Networks Camilo Rojas, Juan A. Fraire, Fabio Patrone, and Mario Marchese Advanced Satellite Multimedia Systems Conference and Signal Processing for Space Communications Workshop (ASMS/SPSC), 2025
Abstract
Non-Terrestrial Networks (NTN) are poised to revolutionize 5G and 6G networks by integrating terrestrial and space-based cloud systems, enabling dynamic task allocation for optimal performance. Despite their promise, understanding the trade-offs between terrestrial and non-terrestrial edge computing architectures remains an area for improvement. This paper presents a comprehensive latency-focused trade-off analysis using a novel real-time emulation platform that accurately models terrestrial and space cloud environments. By evaluating network latency across geodesic distances from a fixed ground gateway, we delineate scenarios where terrestrial clouds excel and identify conditions under which Space Cloud architectures surpass their terrestrial counterparts. Additionally, we analyze how server placement strategies in satellite constellations impact performance, revealing the critical interplay between server distribution and latency outcomes. These findings offer actionable insights for designing and operating hybrid cloud systems, emphasizing the need for tailored architectures to maximize the potential of NTN-based edge computing.

HAL Publisher
Dark Clouds Rising in Low-Earth Orbit: On Environmental Limits to Massive Orbital AI Robin Ohs, Gregory F. Stock, Andreas Schmidt, Juan A. Fraire, Jörg Ott, and Holger Hermanns ACM Workshop on LEO Networking and Communication (LEO-NET), co-located with ACM SIGCOMM, 2026
Abstract
In 2026, we saw rising numbers of proposals for orbital data centers to facilitate AI, e.g., by SpaceX (up to one million satellites), Blue Origin, and Google—yet none include rigorous lifecycle sustainability analysis. We present ESpaS-ODC, a lifecycle carbon model that accounts for the three subsystems physically unavoidable at GPU-class power densities but absent from prior work: a thermal radiator sized from ISS data, solar-array degradation and eclipse margin, and cold-standby spares for no-repair access. Applying the model to an edge data center in a 510 km orbit reveals that the service-overhead-scaled radiator weighs about eleven times the GPU it serves and that modeled power and thermal infrastructure dominate launched component mass—not compute. Using mission duration T as the denominator to amortize launch, re-entry, and manufacturing carbon, model-boundary parity with a global-average terrestrial DC occurs within the first two mission years in our parameter sweep. With respect to a renewables-powered green-energy DC (Finland grid intensity: 68 gCO2e/kWh, PUE: 1.2), parity requires multi-year missions with cold-standby spares compensating the shorter component lifetimes. Idealized dawn-dusk SSO (no-eclipse regime, β > βcrit) eliminates the eclipse battery, cutting modeled component mass significantly (e.g., 25% on the 1 kW-ODC for a 3 yr mission), but leaves the radiator unchanged: high-beta orbits solve the battery problem, not the thermal problem. Finally, carrying one full cold spare (r = 2) increases amortized carbon per GPU hour by 40% on Starship and 34% on Falcon-9 at T = 3 yr, while its dependability benefit remains to be quantified.

HAL Publisher

Orbital computing: the topic page All 15 papers on this topic