In A Sky Full Of Cosmic Noise, Oxford's New AI Tool Helps Astronomers Find Supernovae With Ease

Hyderabad: Artificial Intelligence (AI) and Machine Learning (ML) have been continuously helping researchers with their work, enabling them to sort through massive amounts of data and reach results quickly than ever. These technologies are now helping astronomers find supernovae in a sky full of cosmic noise.

Supernovae are rare, bright explosions that mark the death of massive stars. However, identifying them is not so easy, as they must be spotted quickly before they fade, while differentiating them from other instrumental errors and known objects. A team of researchers led by Oxford University and Queen’s University Belfast searches for these using the Asteroid Terrestrial Impact Last Alert System (ATLAS), a project funded by NASA and led by the University of Hawaii.

Originally built as an asteroid impact early warning system, ATLAS scans the entire visible sky every 24 to 48 hours using five telescopes located around the globe. Oxford processes the data for high explosions beyond our galaxy, which yields millions of potential alerts nightly. Researchers then apply standard filtering and automated image analysis techniques, leaving them with 200-400 candidate signals that need to be checked manually and take several hours each day.

To tackle this problem, Oxford has made an AI tool, called the Virtual Research Assistant (VRA), that can do the heavy lifting—probably using an ML approach—filtering through thousands of data alerts to identify the few genuine signals caused by supernovae. According to the study, published in The Astrophysical Journal, it reduces astronomers' workload by as much as 85 per cent.

The new tool is a collection of automated bots that mimics the human decision-making process by ranking alerts based on their likelihood of being real, extragalactic explosions. Unlike data-hungry deep learning models that require vast training data and supercomputers, the VRA uses a leaner approach and uses smaller algorithms based on decision trees, looking for patterns in selected aspects of the data. This allows scientists to use their expertise for the training and guide it to look for key features required for the task.

"The surprising thing is how little data it took. With just 15,000 examples and the computing power of my laptop, I could train smart algorithms to do the heavy lifting and automate what used to take a human beings hours to do each day," says lead researcher Dr Héloïse Stevance from the Department of Physics, University of Oxford.

The VRA also updates its assessment each time a telescope revisits the same patch of sky, meaning a signal is automatically re-checked and re-scored over several nights. This leaves human astronomers to review only the most promising candidates.

In its first year of use, the VRA successfully filtered over 30,000 alerts, while missing fewer than 0.08 per cent of real supernovae alerts and retaining more than 99.9 per cent of genuine supernovae candidates. In December 2024, a researcher linked the tool to the South African Lesedi Telescope for an automated "promising signal" alert, even before a human reviews the data.

Researchers say that VRA comes just in time as the Vera Rubin Observatory’s Legacy Survey of Space and Time (LSST) prepares for a launch in early 2026. Over ten years, the LSST will survey the entire southern hemisphere sky every few days, ultimately generating over 500 petabytes (1 PB = 1,024 TB) of images and data.