A new paper in Nature Astronomy says Google’s Gemini model can be adapted to classify transient astronomy images and explain its reasoning in plain language. The study, “Textual interpretation of transient image classifications from large language models”, looks at how a general-purpose multimodal model could help astronomers handle the huge volume of alerts generated by modern sky surveys.
Modern astronomy surveys produce millions of alerts about possible discoveries, but most are not real cosmic events. The source text says many are “bogus” signals caused by satellite trails, cosmic ray hits, or other instrumental artefacts. Astronomers have long used specialised machine learning models such as convolutional neural networks (CNNs) to sort these alerts, but these systems usually return only a “real” or “bogus” label with no explanation.
That lack of transparency can force scientists to either trust the output or manually check candidates. The source text says this is becoming a growing bottleneck as next-generation telescopes approach, including the Vera C. Rubin Observatory, which is expected to generate 10 million alerts per night.
According to the paper, Gemini was tested as a more flexible alternative that can work with both text and images. The researchers say the model matched the accuracy of specialised systems and could also explain its classifications in plain language. They used few-shot learning, giving Gemini just 15 annotated examples per survey along with concise instructions to classify and explain cosmic events.
The result matters for astronomy teams because it could reduce the time spent on manual verification while making automated classifications easier to interpret. The study presents Gemini as an “expert astronomy assistant” for transient image classification, rather than a replacement for existing survey workflows.
Source: research.google.
Companies can share verified announcements through Newz9’s international press release submission page.
