TL;DR

Researchers at the Alan Turing Institute’s defence research centre have built a model that identifies satellites behaving abnormally using nothing but how sunlight glints off them. Published in Expert Systems, it flags unusual readings 88% of the time and can tell a spinning spacecraft from a tumbling one — a distinction that matters for anyone planning to service or refuel it.

Learning what normal looks like

The technique borrows its structure from language models. Rather than text, the system consumes vast quantities of “light curves”, the brightness traces telescopes record as an object passes overhead, and learns the patterns that constitute ordinary behaviour. Curated simulation data from Strathclyde’s aerospace centre and the firm GMV then sharpens it for specific tasks. In operation it takes feeds from ground-based observatories and raises anomalies for a human to examine.

The scale problem it addresses is real and getting worse. Roughly 4,000 satellites went up in 2025; the figure for the year 2000 was 159 launches worldwide. Starlink alone accounts for more than 10,000 objects in orbit, with authorisation sought for four times that. Human analysts cannot read the daily output any more, which is the actual bottleneck.

Why a British institute is doing space safety

Victoria Nockles, who heads the centre and co-authored the paper, made the dependency argument rather than the space one: orbital collisions are rare but catastrophic, and the infrastructure at risk delivers remote communications, satellite positioning and the precision timing that global financial markets settle against. Framed that way this is critical national infrastructure monitoring that happens to be conducted from the ground.

Lead author Ian Groves described inferring behaviour purely from reflected light as a new technique. The work sits inside AI4S3, a UK Space Agency-funded consortium spanning Five Eyes countries, with MIT, Waterloo and Arizona among the academic partners.

Looking forward

Next comes multimodal input — radar returns, hyperspectral data and orbital tracks alongside the light curves. It is also a neat illustration of the argument the same institute made this morning about sovereign capability: evaluation and assurance in a domain Britain can credibly lead, rather than an attempt to compete on frontier model scale. The centre has separately won funding with Birmingham for an in-orbit radar imaging project.