Toward Scalable SDN for LEO Mega-Constellations: A Graph Learning Approach

Sivaram Krishnan, Bassel Al Homssi, Zhouyou Gu, Jihong Park, Sung-Min Oh, Jinho Choi · 2026

Read Toward Scalable SDN for LEO Mega-Constellations: A Graph Learning Approach by Sivaram Krishnan, Bassel Al Homssi, Zhouyou Gu, Jihong Park, Sung-Min Oh, Jinho Choi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-27478.txt.

Introduction

Krishnan and five colleagues propose a scalable, layered software network framework for mega-constellations of thousands of low Earth orbit satellites. Their design uses graph neural networks to map orbit layout, and Koopman theory to make nonlinear dynamics linear. A Graph Koopman Autoencoder forecasts space-time moves within a per-shell linear subspace. A controller then combines shell forecasts for control. On Starlink tests, it lifts space compression by at least 42.8 percent and time forecasts by about 11 percent over baselines, while using a smaller footprint.

Catalog record

CREATORSivaram Krishnan, Bassel Al Homssi, Zhouyou Gu, Jihong Park, Sung-Min Oh, Jinho Choi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30955
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.27478

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Toward Scalable SDN for LEO Mega-Constellations: A Graph Learning Approach Sivaram Krishnan, Bassel Al Homssi, Zhouyou Gu, Jihong Park, Sung-Min Oh, Jinho Choi Abstract Terrestrial network limitations drive the integration of non-terrestrial networks (NTNs), notably mega-constellations comprising thousands of low Earth orbit (LEO) satellites. While these satellites act as interconnected network switches via inter-satellite links (ISLs), their massive scale creates severe bottlenecks for network management. To address this, we propose a scalable, hierarchical software-defined networking (SDN) framework. Our architecture leverages graph neural networks (GNNs) to compactly represent the constellation topology, and Koopman theory to linearize nonlinear dynamics. Specifically, a Graph Koopman Autoencoder (GKAE) forecasts spatio-temporal behavior within a linear subspace for each orbital shell. A central SDN controller then aggregates these shell-level predictions for globally coordinated control. Simulations on the Starlink constellation demonstrate that our approach achieves at least a 42.8% improvement in spatial compression and a 10.81% improvement in temporal forecasting compared to established baselines, all while utilizing a significantly smaller model footprint. I Introduction Thanks to inter-satellite links (ISLs), low Earth orbit (LEO) satellites can form a network that can scale into mega-constellations [ 1 , 13 ] , enabling seamless interconnection among satellites as they orbit the Earth. Such mega-constellations have the potential to support truly global connectivity with high data rates and low latency, offering significantly improved coverage and service quality even in remote or underserved regions. By leveraging dense satellite deployment and flexible routing through ISLs, these networks can provide resilient and scalable communication services that integrate seamlessly with terrestrial systems as part of non-terrestrial networks (NTNs) [ 9 ] , thereby complementing and enhancing existing terrestrial infrastructure. Software defined networking (SDN) offers centralized, software-based control for networks, providing greater flexibility and scalability than traditional distributed architectures [ 5 ] . It can align well with the structured nature of LEO constellations, which are organized into shells composed of distinct orbital planes whose predictable motion enables accurate modeling of network dynamics and supports the global, time-evolving view required by SDN. Prior studies show that SDN can improve handover performance [ 15 ] and enable more efficient routing through global network awareness [ 8 ] . SDN-based control becomes particularly advantageous in complex scenarios where multiple LEO constellations span different shells or operate across many orbital planes, requiring coordinated management of rapidly changing topology. In parallel, there has been a growing rise of artificial intelligence (AI) and machine learning (ML) techniques applied to SDN [ 14 ] and satellite networking [ 6 ] . These approaches can enable data-driven optimization, adaptive decision-making, and intelligent control mechanisms that more effectively manage the complexity and dynamics of mega-constellations. However, managing a large number of LEO satellites presents significant challenges for fully centralized SDN control. While distributed mechanisms, such as end-to-end flow control used in the Internet, can be applied to mega-constellations, they fail to exploit the networks’ highly regular structure and predictable orbital dynamics, which could enable more efficient centralized management. Unlike the Internet, where heterogeneity necessitated distributed control at the cost of efficiency, mega-constellations offer the opportunity to achieve both scalability and efficiency. Addressing these challenges requires efficient representations that leverage the regular structure and dynamics of the constellation, along with learning in a low-dimensional latent space to support effective centralized control and resource management. In this article, we address the challenge of effectively representing and managing satellite mega-constellations. By treating a mega-constellation as a dynamic graph—where nodes represent LEO satellites and edges represent ISLs—we employ graph neural networks (GNNs) [ 11 , 4 ] as a foundational tool to learn topological structures. However, relying solely on standard GNNs is insufficient to manage the massive scale and complex, time-varying dynamics of multi-orbit networks. To overcome this, we propose a hierarchical learning framework designed to equally address both spatial and temporal scalability. First, to achieve spatial scalability , we decompose the global constellation into distinct orbital shells. This shell-specific representation drastically reduces the dimensionality of the state space, allowing localized topologies to be processed independently. Second, to achieve temporal scalability , we model the evolution of each individual shell using a dedicated Graph Koopman Autoencoder (GKAE) [ 7 ] . At its core, Koopman operator theory offers a powerful mathematical paradigm: it posits that any complex, non-linear dynamical system can be equivalently represented by a linear operator acting on a higher-dimensional space of observable functions. Building on this, the GKAE merges the spatial compression of GNNs with Koopman theory to dynamically ’lift’ (to a higher-dimension) the highly non-linear, discrete network states into a continuous, linear latent space. This linearization is the key to temporal scalability. By transforming a non-linear control problem into a linear one, the framework bypasses the compounding errors of traditional recurrent models, enabling long-term predictions that are mathematically stable, highly interpretable, and computationally efficient. Within this dual-focused framework, we envision a logically centralized controller that aggregates the compact, linearized embeddings from each shell’s GKAE, including their forecasted state trajectories. Using these predictive representations, the controller coordinates inter-shell communication and establishes a pathway to optimize control sequences directly within the latent space, ensuring global coherence. II Software Defined Satellite Networking: Opportunities and Challenges Software Defined Satellite Networking : As illustrated in Fig. 1 , the software defined satellite networking (SDSN) framework adapts the traditional terrestrial SDN paradigm—decoupling the control plane (decision logic) from the data plane (packet forwarding)—to the dynamic environment of space. In this architecture, LEO satellites function as programmable switches in the data plane, forwarding traffic based on flow tables populated by a logically centralized controller via a Southbound Interface (e.g., space-adapted OpenFlow [ 10 ] ). To ensure resilience, these control channels can be mapped out-of-band using dedicated telemetry frequencies, or multiplexed in-band over ISLs. To circumvent the strict size, weight, and power (SWaP) constraints of space hardware, foundational architectures like OpenSAN [ 3 ] traditionally host the management plane on the ground. However, relying on ground stations for continuous global orchestration introduces debilitating ground-to-space control loop latency, fundamentally limiting the network’s responsiveness. Driving Applications : Operating through a Northbound Interface, the application plane translates high-level business logic into executable network policies. The core advantage of SDSN is its ability to multiplex diverse, mission-critical services over a single physical infrastructure. Through a central controller, a mega-constellation can dynamically provision secure, low-latency links for defense operations (e.g., Starshield), deliver fiber-like maritime broadband (e.g., Starlink Maritime), and prioritize the rapid downlink of time-critical remote sensing data. Fig. 1 : An illustration of the SDSN framework. The Data Plane consists of LEO satellites acting as forwarding switches. The Control Plane centralizes network intelligence for effective control and management. The Application Plane defines high-level network policies. (Generative AI was used for editing this schematic) Challenges in SDSN : These diverse applications—from broadband connectivity to real-time sensing—underscore the need for scalable, adaptive, and high-fidelity network control: • The Curse of Dimensionality : As mega-constellations scale to tens of thousands of satellites, the global state space explodes. This renders traditional, centralized routing optimizations NP-hard and computationally infeasible. Managing this requires a hierarchical, learnable framework capable of compressing immense topological data into a tractable, lower-dimensional space, while accounting for the distinct orbital mechanics and traffic loads of different shells. •

Read the complete text