A new study demonstrates the power of a general-purpose artificial intelligence to identify exploding stars with remarkable accuracy, using a training dataset thousands of times smaller than previously required. The breakthrough suggests a transformative shift in how astronomers can sift through the immense volumes of data generated by modern sky surveys, potentially accelerating the pace of cosmic discovery.
In a paper published in the journal Nature Astronomy, an international team of researchers detailed how they successfully repurposed Google’s large language model, Gemini, for the highly specialized task of astrophysical analysis. By providing the AI with just 15 labeled examples, they taught it to distinguish between significant celestial events, such as supernovae, and unimportant image artifacts or variable stars. This “few-shot learning” approach stands in stark contrast to traditional machine learning models that must be trained on tens of thousands of examples, marking a significant step toward solving the data processing bottleneck in observational astronomy.
A New Method for Cosmic Event Detection
The core of the research involved testing a novel application for a tool not originally designed for scientific image analysis. Instead of building a specialized neural network from the ground up, the scientists leveraged the advanced pattern-recognition capabilities inherent in a pre-existing large language model. This method sidesteps a major hurdle in computational astronomy: the years of effort and vast, curated datasets required to train bespoke AI systems for specific tasks.
The process was elegantly simple. The researchers prompted Gemini with a tiny set of 15 images, instructing it to classify celestial phenomena into three distinct categories. Images containing explosive events like supernovae were labeled “High interest.” Those showing known variable stars were marked “Low interest,” and images with camera-related artifacts or other “bogus signals” were designated “No interest.” The model then applied this minimal training to analyze vast collections of images from professional observatories, demonstrating an ability to generalize from a handful of examples with surprising effectiveness.
Performance Across Major Sky Surveys
To validate their findings, the researchers tested Gemini’s performance on real-world data from three prominent astronomical survey projects: the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS), the Dutch MeerLICHT telescope, and the Asteroid Terrestrial-impact Last Alert System (ATLAS). These automated telescopes scan the sky nightly, generating a torrent of data on transient events—celestial objects that change in brightness over short timescales.
Accuracy on Professional Datasets
The results were consistently strong across all three datasets, underscoring the model’s robust capabilities. Gemini achieved an accuracy of 94.1% on data from Pan-STARRS, 93.4% with MeerLICHT images, and 91.9% for the ATLAS survey. The team also performed a follow-up analysis six months later, after Gemini had received algorithmic updates, confirming the durability of the model’s classification skills. The complete set of examples, prompts, and instructions used in the study has been made publicly available for other researchers to build upon.
Solving the Data Deluge Problem
Modern astronomy faces a fundamental challenge: telescopes are now so powerful that they generate more data than scientists can manually inspect. Dr. Stephen Smartt, a professor of astrophysics at the University of Oxford and a co-author of the study, highlighted the decade-long struggle to efficiently process this information. “We have spent years training machine learning models, neural networks, to do image recognition,” he noted. The primary difficulty lies in “weeding out the real events from the bogus signals.” According to Smartt, the LLM’s ability to achieve high accuracy with such minimal guidance was “remarkable” and could be a “total game changer for the field” if scaled up, serving as another key example of AI enabling fundamental scientific discovery.
Broader Implications for Astronomy
This study is part of a larger, rapidly growing trend of integrating artificial intelligence into nearly every facet of astronomy and planetary science. AI tools are already proving indispensable in a wide array of applications, from discovering new exoplanets to analyzing the surfaces of planets in our own solar system. For instance, AI was instrumental in identifying Kepler-90i, the eighth planet found orbiting a star 2,767 light-years away, by spotting a faint signal that had been missed in previous analyses.
The techniques pioneered in this study could soon be applied to other large-scale data challenges in astronomy. AI is currently being used to identify a menagerie of cosmic phenomena, including fast radio bursts, gamma-ray bursts, and the faint ripples in spacetime known as gravitational waves. By reducing the need for massive, task-specific training datasets, general-purpose models like Gemini could democratize the use of AI, allowing smaller research teams to tackle complex classification problems that were once the domain of highly specialized groups. As future observatories prepare to come online, this efficient and adaptable approach to data analysis will be crucial for managing the coming wave of cosmic information and ensuring no discovery is missed.