That communication challenge sits at the center of two papers plus an earlier paper from the Optical Communications & Computing department at NEC Laboratories America, written in collaboration with California State University, Los Angeles, and Pomona College.

The first paper, Mobile Orbital Domain-based Hierarchical Routing in Satellite Networks, was presented at OFC 2026, the Optical Fiber Communication Conference. The authors are Zilong Ye, California State University, Los Angeles; Philip N. Ji, NEC Laboratories America; Deniz Bajin, Pomona College; Connor Wang, Pomona College; Yunxuan Wang, Pomona College; and Ting Wang, NEC Laboratories America.

The second paper, “Mobile Orbital Domain-based Hierarchical Routing with Joint Path and Gateway Selection,” was presented at OECC 2026, the 31st OptoElectronics and Communications Conference in Busan, South Korea. The authors are Zilong Ye, Philip N. Ji, and Ting Wang.

Where The Mobile Orbital Research Begins

The story actually starts with an earlier paper, Mobile Orbital Domains: Addressing Dynamic Topology Challenges in Satellite Networks, presented at Frontiers in Optics + Laser Science 2025 (FiO + LS 2025). Zilong Ye of NEC Laboratories America and California State University, Los Angeles, wrote it with Philip N. Ji, Giovanni Milione, and Ting Wang of NEC Laboratories America. This first paper asked a simple question: if satellites never stop moving, does anything in the network stay put?

To find out, the team followed 3,000 Starlink satellites for 12 hours. More than 2,500 of them maintained steady links with their neighbors throughout, forming what the authors call a satellite backbone network (SBN). It was the first time anyone had identified a backbone like this. This paper also sketched the idea of mobile orbital domains, groups of satellites that travel in the same direction at similar speeds. Those two findings set the stage for the routing work that was presented at OFC 2026 and OECC 2026.

Routing Through a Moving Sky

Low Earth orbit (LEO) satellites circle the planet in roughly 90 minutes. By 2025, the Starlink constellation alone had more than 7,000 active LEO satellites. Many of these satellites talk to each other directly through inter-satellite links (ISLs), using radio or free-space optical (FSO) communications that carry data via laser beams through open space. Data can then hop across many satellites before reaching the ground.

Deciding which hops to take is the job of routing, and each satellite keeps a routing table, a directory of directions to other destinations. Today’s approaches struggle with constant motion. In flat routing, every satellite stores directions to every other satellite. In a 3,000-satellite network, that means 2,999 entries each. Geographic domain routing groups satellites by location over Earth, which shrinks tables, but satellites still cross domain boundaries. Each crossing forces the network to reorganize, adding control overhead.

Grouping Satellites That Travel Together

The research team started with real orbital data. They tracked Starlink satellites over 12 hours using two-line element (TLE) data, a standard format for describing satellite orbits, published by CelesTrak. They counted two satellites as connected when they stayed within 1,000 kilometers of each other, about the typical range of an FSO link. A large share of satellite pairs stayed connected for the entire 12 hours. Among 3,000 tracked satellites, 2,505 belonged to this persistent satellite backbone network (SBN), and the backbone grew denser as more satellites were added, with the average number of stable connections per satellite rising from 1.44 at 1,000 satellites to 3.02 at 3,000.

Building on that backbone, the team grouped satellites with similar direction, speed, and orbital plane into mobile orbital domains (MODs). Think of a convoy on a highway. The convoy moves, but each vehicle’s position relative to the others holds steady. Using DBSCAN, a clustering algorithm that groups closely packed data points, the team identified 198 MODs. Each forms a snake-like chain that keeps its shape as it moves.

Routing then works in three stages. Data first travels inside its starting MOD to a border gateway satellite (BGS), a satellite with a link into a neighboring MOD. It then crosses a sequence of MODs through these gateways, and finally travels inside the destination MOD to its target. Because the structure stays stable, the routing tables can be computed ahead of time with Dijkstra’s algorithm, a classic method for finding shortest paths. Regular satellites store directions only within their own MOD, while gateways hold extra entries to reach other MODs.

The second paper tackles a weakness in that design. Gateways carry all traffic between MODs, so a few of them can become bottlenecks. The improved scheme, called MOD-H with Path and Gateway Selection (MOD-H-PGS), considers several candidate routes between MODs and every viable gateway pair along each route. It then picks the route whose busiest gateway carries the lightest load, much like a navigation app choosing both the highway and the on-ramp based on current traffic.

Smaller Tables, Faster Decisions

The team simulated 3,000 Starlink satellites with random traffic requests ranging from 10 megabits per second to 1 gigabit per second. The MOD-based approach, called MOD-H, cut the average routing table to 94.69 entries, compared with 2,999 for flat routing and 124 for geographic domain routing, a 96.81% reduction from flat routing. It also computed routes at least twice as fast as flat routing and significantly faster than geographic routing, since most requests resolve through a quick table lookup.

To capture both gains, the authors defined a Routing Efficiency Index (REI) that combines table size and computation time. MOD-H scored 2.14, compared with 0.71 for flat routing and 0.22 for geographic routing. The trade-off is that precomputed paths take more hops and can crowd gateways. Even with hop count and peak link load added to the index, MOD-H still led with 0.55, compared with 0.50 for flat routing and 0.21 for geographic routing.

The gateway selection work addressed that crowding directly. MOD-H-PGS reduced the load on the most congested link by 22% on average. Its 95th percentile gateway load, the level below which 95% of gateway loads fall, dropped to 1.53 gigabits per second from 1.99 gigabits per second, a 23% improvement. The advantage grew with traffic.

Real-World Applications

Satellite broadband operators and constellation operators stand to benefit most directly. Smaller routing tables and fewer updates mean less onboard memory, processing, and control traffic, which matters on spacecraft with limited power and computing.

Emergency management agencies, humanitarian organizations, and utilities rely on satellite links when ground networks fail after hurricanes, floods, or wildfires. Stable, precomputed routing can help keep those links predictable when demand spikes. Maritime and aviation connectivity providers, rural broadband carriers, and government agencies that need resilient communications face similar demands.

Satellites may look chaotic from the ground, but they actually move in highly predictable patterns. Once we stopped chasing every satellite’s position and started treating groups that travel together as stable units, routing became much faster and simpler. Our latest work shows that we can also spread traffic more evenly across the network without losing those advantages. – Philip N. Ji, NEC Laboratories America

Both studies simulated up to 3,000 satellites, limited by computing resources. Testing at full constellation scale and adding ground-station links are natural next steps as satellite networks carry more of the world’s communications.