Vishanti

Predicting Chinese Satellite Constellation Maneuvers in the Year 2051

Why This Matters

By the early 2050s, China is on track to operate a fleet of roughly 50,000 satellites. That is fifty times what they fly today. No analyst workforce can track that many satellites by hand. No heuristic rule set built for the 2010s will scale to it. The tools that defense, intelligence, and space operations rely on right now were designed for a sky that no longer exists, and they will not survive the one that is coming.

For a defense or intelligence analyst, the payoff is seven days of warning before a Chinese satellite maneuvers. Vishanti looks across the full fleet and ranks every satellite by how likely it is to maneuver in the next week. The analyst sees a sorted list and decides where to point collection assets. At seven days out, the system is right about 96 percent of the time, with a precision of 97 percent. False alarms are rare. Missed maneuvers are rare. The whole forecast for a 50,000-satellite fleet refreshes every six hours on a single modern GPU.

For space agencies and commercial operators, the same seven-day warning replaces reactive conjunction screening with proactive collision avoidance. Operators can schedule ground stations around expected orbital shifts. Mission planners get advance notice of orbital congestion. The system works across every major orbital regime, from low Earth orbit through GEO and HEO, with consistent accuracy in each one. It runs on standard hardware.

For the broader space community, earlier warning means fewer surprises in an orbital environment that is filling up fast. The compute budget is small. The architecture scales as the satellite population grows. Inference is fast enough to plug into existing space operations workflows without rebuilding them.

Background

About 1,000 Chinese satellites are in orbit today. Publicly filed Chinese constellation programs project that number toward 50,000 or more by the early 2050s. Guowang, Qianfan, the December 2024 CTC filings for two constellations of 96,714 satellites each, and the broader PLA Aerospace Force fleet all contribute to that trajectory. Knowing when each of those satellites is about to maneuver matters for collision avoidance, mission planning, indications and warning, and space domain awareness.

Today, most of this work is reactive. A maneuver gets confirmed after it happens. The analyst updates the tracking data and assesses the risk after the fact. That approach falls apart when there are tens of thousands of satellites to track and when the operator of those satellites does not announce intent. Traditional orbit determination can confirm that a maneuver happened by spotting the trajectory change. It cannot forecast the next one.

Earlier machine learning work in this area reports about 70 percent accuracy on three-day prediction. Those studies looked at hundreds of satellites, not tens of thousands. At a fleet of 50,000, manual triage stops working. Heuristic rules cannot keep up.

Two-Line Element data is the foundation for the new approach. TLEs are public, refreshed often, and they capture the orbital parameters that shift when a satellite adjusts its trajectory. The hard part has always been pulling the faint pre-maneuver signals out of the noise. Drift patterns. Deadband approaches. Multi-day phasing setups. Older methods miss these signals. So do conventional neural networks.

Our Solution

Vishanti predicts when a satellite will maneuver, up to 14 days ahead. On a held-out test set of 43,380 windows drawn from a simulated 50,000-satellite fleet, the system hits 96.3 percent F1 and 98.1 percent accuracy at the seven-day horizon. At three days, F1 is 88.6 percent. At fourteen days, 98.6 percent. Every one of those numbers reflects how the model performs on satellites it never saw during training.

The system uses three deep learning models built on a quantum-inspired architecture. Two of the models read orbital telemetry from different angles and capture different aspects of satellite behavior. The third combines what the first two see into a single calibrated prediction. The combined result beats any single model on its own, across every prediction window and every orbital regime tested.

We grounded the training data in real orbital behavior. The patterns came from Space-Track.org, the U.S. government's public catalog of satellite tracking data. We studied how current Chinese satellites actually move, station-keep, and maneuver. We then extrapolated those patterns forward to model what a 50,000-satellite Chinese fleet would look like in the year 2051. The synthetic fleet inherits the orbital mechanics, drift behavior, and maneuver cadence observed in the real data. That keeps the training grounded while letting the model learn at a scale that does not exist yet.

Each satellite is run through the full ensemble. The inputs are recent orbital state, derivative behavior, deadband proximity, and per-satellite history. The outputs are calibrated probabilities at 24 hours, three days, seven days, and fourteen days.