A new artificial intelligence approach co-led by researchers at the University of Oxford and Google Cloud allows astronomers to identify significant celestial events from a massive firehose of nightly alerts using only a handful of examples for training. The system dramatically lowers the barrier for creating powerful cosmic classifiers, providing a tool that is not only accurate but can explain its reasoning, moving beyond the “black box” problem of many previous machine learning models.
Published in Nature Astronomy, the study demonstrates that a general-purpose large language model, Google’s Gemini, can be transformed into an expert astronomy assistant with minimal guidance. Using just 15 example images, the AI learned to distinguish genuine astronomical transients—such as supernovae and asteroids—from imaging artifacts with approximately 93% accuracy. This breakthrough is critical as observatories prepare for an unprecedented deluge of data from next-generation telescopes that will stream terabytes of information every 24 hours.
A New Paradigm for Astronomical Classification
Modern telescopes relentlessly scan the sky, generating millions of alerts each night that point to potential changes. The vast majority of these are not new discoveries but “bogus” signals caused by satellite trails, cosmic ray impacts, or other instrumental errors. Traditionally, astronomers have relied on specialized machine learning models to sift through this data, but these models often require massive, carefully curated training datasets and deep expertise in AI programming to build and maintain.
The new method circumvents these challenges using a technique known as “few-shot learning.” Instead of hundreds of thousands of examples, the AI was given a simple set of instructions and just 15 labeled image sets from each of three astronomical surveys: ATLAS, MeerLICHT, and Pan-STARRS. Each set consists of an image triplet: a “New” image showing the latest observation, a “Reference” image of the same patch of sky from a previous time, and a “Difference” image created by subtracting the reference from the new frame to highlight the change. This approach allows the model to quickly learn the visual patterns that distinguish a real astrophysical source from an artifact without complex retraining.
Impressive Accuracy from Minimal Training
The research team, which also included scientists from Radboud University, tested the model on thousands of alerts from the different surveys. The AI successfully identified a wide range of cosmic phenomena, including supernovae (exploding stars), tidal disruption events where a black hole tears apart a passing star, fast-moving asteroids, and short stellar flares. Despite the minimal training, the system achieved an initial accuracy of roughly 93% across the board, demonstrating a remarkable ability to generalize from a small sample size.
This performance proves that a versatile, general-purpose AI can be effectively adapted for highly specialized scientific tasks. According to Dr. Fiorenzo Stoppa, a co-lead author from Oxford’s Department of Physics, “It is striking that a handful of examples and clear text instructions can deliver such accuracy. This makes it possible for a broad range of scientists to develop their own classifiers without deep expertise in training neural networks – only the will to create one.”
Building Trust with Explainable AI
Transparent Reasoning
A significant challenge with many older AI systems in science is their inability to explain their decisions, operating as opaque “black boxes.” This forces scientists to either blindly trust the output or spend valuable time manually verifying it. The new system addresses this directly by generating a concise, plain-English explanation for every classification it makes. For each candidate, it describes the salient features in the image triplet and outlines its reasoning for labeling it as “real” or “bogus.”
Expert Verification
To validate the quality of these explanations, the researchers had a panel of 12 professional astronomers review 200 randomly selected classifications via the Zooniverse platform. The astronomers rated the AI’s reasoning on a scale from 0 (a complete hallucination) to 5 (perfectly coherent and accurate). The average score exceeded 4, indicating that the descriptions were overwhelmingly logical and useful for human experts. This transparency is a crucial step toward making AI-driven science more trustworthy and fostering a collaborative relationship between humans and machines.
Human-in-the-Loop for Enhanced Performance
Perhaps the most novel aspect of the research is the AI’s ability to assess its own confidence. The model can review its classifications and assign a “coherence score” to its own text explanations. The researchers found that outputs with low coherence scores were much more likely to be incorrect, meaning the system can effectively flag its own potential mistakes and uncertainties. This capability is vital for creating a reliable “human-in-the-loop” workflow, where the AI automates the vast majority of clear-cut classifications and forwards only the most ambiguous cases for human review.
The team leveraged this self-correction loop to further refine the model’s performance. By identifying the challenging examples that the AI flagged as uncertain, they added a small number of them back into the training set and reran the analysis. This simple feedback process boosted the model’s accuracy on the MeerLICHT dataset from 93.4% to an impressive 96.7%. This demonstrates how the system can learn and improve over time in partnership with human experts.
Preparing for an Era of Big Data
The development of this AI comes at a critical time for astronomy. Upcoming facilities, most notably the Vera C. Rubin Observatory in Chile, are poised to revolutionize the field by surveying the entire southern sky every few nights. The Rubin Observatory alone is expected to generate about 20 terabytes of data and issue millions of transient alerts nightly—a scale at which manual vetting is completely infeasible. Tools like the one developed by the Oxford and Google team will be essential for managing this data flow and ensuring that rare and scientifically valuable events are not lost in the noise.
Democratizing Scientific Discovery
Beyond its immediate application in astronomy, this research highlights a broader shift in how AI can be used in science. By showing that a general-purpose model can be adapted for a complex task with little data and no specialized programming, the work empowers researchers in other domains to build their own AI assistants. “As someone without formal astronomy training, this research is incredibly exciting,” said Turan Bulmus, a co-lead author from Google Cloud. “It demonstrates how general-purpose LLMs can democratize scientific discovery, empowering anyone with curiosity to contribute meaningfully to fields they might not have a traditional background in.”
