Modern astronomy resembles a cosmic treasure hunt, as telescopes worldwide search for fleeting events like supernovae to glean insights into the universe. These surveys produce millions of alerts, but many are “bogus” signals from satellite trails or cosmic rays. To tackle this, astronomers have utilized specialized machine learning models, such as convolutional neural networks (CNNs), to filter data. However, these “black box” models only offer a simple “real” or “bogus” label, leading to potential roadblocks with next-gen telescopes, like the Vera C. Rubin Observatory, which will generate 10 million alerts nightly. Our research, published in Nature Astronomy, explores whether a general-purpose multimodal model can not only match—but exceed the specialized models in accuracy while providing explanations. We demonstrate that Google’s Gemini model can classify cosmic events and articulate its reasoning in plain language, using few-shot learning with just 15 examples per survey, revolutionizing the interpretation of astronomical data.
