A new artificial intelligence model is capable of tracking satellites and detecting deviations in their behavior, reducing the risk of collisions in orbit. The study was published in the journal Expert Systems 27 September 2026 year.
The development was carried out within the framework of the AI4 Space Safety and Sustainability project with the support of the UK Space Agency. The consortium includes organizations from the FVEY countries: Australia, Canada, New Zealand, the United Kingdom and the United States.
The project is led by Professor Massimiliano Vasile, director of the Aerospace Centre of Excellence at the University of Strathclyde. Among the participants are the Alan Turing Institute, the University of Arizona, MIT, the University of Waterloo, as well as the companies GMV, Columbiad, LMO and Zendir.
The model, created at the Defence AI Research Centre (DARe) of the Alan Turing Institute, was the first to learn to predict anomalies and the movement of satellites from the light reflected off them — the so-called light curves.
Every year the orbit becomes increasingly crowded: in 2025 more than 4000 new satellites were launched, compared with 159 in 2000. The automated tool helps to cope with the monitoring and management of space traffic.
The model was trained on large volumes of data on satellite brightness collected by telescopes, and then fine-tuned on simulations from Strathclyde and GMV. In real time, it analyzes observatory data and flags anomalies for further study by specialists.
Testing showed an accuracy of detecting unusual light curves at the level of 88 %. The system also distinguishes such satellite behaviors as rotation and tumbling — this is important for on-orbit servicing and extending the service life of spacecraft.
How exactly will the combination of light, radar and other data sources change the approach to protecting satellites in the coming years?
Professor Vasile noted that with the help of AI4S3 it was possible to show how AI helps in analyzing the behavior of space objects even from a single pixel in the sky. The next steps of the project include the integration of multimodal data — radar, hyperspectral measurements and orbital information.
