‘Why give 10 million images to an astronomer?’ UM project trains AI on radio sky

A University of Malta-led project is using artificial intelligence to help astronomers analyse vast numbers of radio telescope images, in a bid to uncover patterns and anomalies in the universe that would be impossible for humans to sift through alone, the project's principal investigator, Andrea DeMarco, said. The project, known as Strada, is developing a foundation AI model designed specifically for radio astronomy. It has been trained on hundreds of thousands of radio images so that it can learn how galaxies and other radio sources appear across different telescopes, resolutions and image qualities. In an interview with The Malta Independent , DeMarco, the project's principal investigator, said the aim is not to replace astronomers, but to give them a tool capable of handling the first layer of analysis at a scale the human eye cannot match. "Why give 10 million images to an astronomer to find that one image?" he said, explaining that the real value of AI lies in making it possible to identify what is interesting before passing it on for scientific interpretation. The project is being led by the University of Malta in collaboration with Italy's National Institute for Astrophysics, INAF, through the Osservatorio Astrofisico di Catania. It has also received financial backing from Xjenza Malta under the Research Excellence programme. The team includes Dr DeMarco, Dr Ian Fenech Conti and Hayley Camilleri from the University of Malta, as well as Dr Simone Riggi from INAF and Ardiana Bushi, now at the University of Edinburgh, who continues to collaborate on the project. Radio astronomy looks at the universe in a way that is very different from the optical telescopes most people are familiar with. Optical telescopes rely on visible light, allowing people to observe bright objects such as stars and planets. Radio telescopes, however, detect radio waves - a different part of the electromagnetic spectrum that cannot be seen with the naked eye. DeMarco explained that the basic principle is similar to the technology behind a car radio or a television dish, except that the source is not a radio or TV station, but signals coming from space, often from distant galaxies. "The advantage is that you're looking at the sky in a completely different spectrum and you do not depend on day or night," he said. "You can look at the radio sky 24/7." That constant stream of data has created a major challenge for modern astronomy. Radio surveys are already producing millions of images, and that number is expected to grow dramatically as next-generation instruments, particularly the Square Kilometre Array, come online. The SKA, which is being built across South Africa and Australia, will be made up of hundreds of thousands of dishes and antennas. These locations were chosen partly because of their radio silence, as mobile phones, Wi-Fi signals and other forms of interference can disrupt radio astronomy observations. DeMarco said the astronomy community has known for years that the amount of data generated by such instruments would be too much for people to analyse manually. "There's no one human person, or even a team, that can look at all this stuff," he said. "At the very least, you need computation. But that is not enough in its own right, you need intelligent computation. This is where the AI aspect comes in." What makes Strada different from previous attempts to use AI in astronomy, he explained, is that it is being built as a foundation model for radio astronomy. In everyday AI, foundation models are systems trained on large amounts of data so that they can then be adapted to many different tasks. Large language models, for example, are foundation models for text, while other systems are trained to work with ordinary photographs and videos. But radio astronomy images are not ordinary photographs. They are usually black and white, often noisy, and can vary significantly depending on the telescope, survey, resolution and background conditions. A general image model trained on real-world photographs may recognise cats, cars or buildings, but radio galaxies are far more niche. DeMarco said that when standard vision models are adapted to radio astronomy, they can work to a degree, but eventually "break down" or produce unreliable results. Strada is being trained to understand radio astronomy from the start. Rather than being told that an image came from one telescope or another, it learns from the images themselves. It has to infer that some images have different backgrounds, different levels of clarity, different resolutions and different object sizes. The model is not being built to perform only one fixed task. Instead, it acts as a starting point that astronomers can adapt for what are known as downstream tasks. These can include classifying galaxies by shape, detecting objects in larger images, or eventually flagging unusual images that may deserve closer human attention. One of the tasks already tested is morphology classification, which looks at the shape of radio galaxies. In some images, the model has to distinguish between different numbers of cores, peaks or connected structures. In other cases, it may need to recognise more diffuse, cloud-like features linked to star-forming galaxies. DeMarco said that on one dataset, Radio Galaxy Zoo, Strada improved the state of the art by around 14%, a jump he described as "huge" in a field where improvements are often closer to 1 or 2%. The model was trained through self-supervised learning, a method that allows AI to learn from large amounts of data without needing each image to be manually labelled by a human. Traditionally, AI systems were trained using labelled data. If a system was learning the difference between cats and dogs, it would be shown many images clearly labelled as one or the other. In astronomy, the same principle has been used to label different types of radio objects. The problem is that astronomy has vast image archives, but not enough labelled data. No one has gone through every image and classified every object, because the scale is simply too large. To work around this, Strada uses learning techniques that allow the image itself to become part of the training process. One method is similar to giving the AI a puzzle with most of the pieces removed. The model is shown only part of a radio image and asked to reconstruct what is missing. At first, it performs badly, but over repeated training it learns to make more accurate predictions based on limited visual context. DeMarco compared this to a person being shown only the eye of a dog and still being able to make an educated guess about what the full image might show. A second training method involves showing the AI different versions of the same radio galaxy. The image may be rotated, its brightness may be changed, or it may be slightly altered, but the model is taught that these are still versions of the same object. It is also shown other objects, some similar and some very different, so that it can learn when variation still belongs to the same source and when it points to something else. DeMarco described this as a form of "identity under pressure", where the AI learns to group or separate objects based on how much variation it sees. Training the model has not been a quick process. DeMarco said the team used around 600,000 raw images, generating many different variations during training. The work required GPUs, high-performance computing resources and repeated testing of different versions of the model. "This took a good year of research, running, number crunching, all of this, to get the final model, the most optimised one," he said. The project's importance lies not only in preparing for the future SKA era, but also in the possibility of revisiting archival data. Radio astronomy archives contain large amounts of data that have not all been inspected in detail by astronomers. DeMarco said Strada could eventually help identify unusual objects or patterns hidden in older data, including material that may have been stored for years without being fully examined. Most of the discoveries expected from tools such as Strada, he said, may not come only from future observations, but also from data already collected and waiting to be properly analysed. The system is also intended to fit into astronomy pipelines, where it can quickly ingest large numbers of images and help decide which are worth saving or flagging for further study. "It just saves what's interesting, what's important," he said, while stressing that the scientific interpretation still belongs to astronomers. The AI may provide an initial classification or confidence level, but the final question of what an object means, and why it matters, remains a job for astrophysicists. The project is open source and hosted on Hugging Face, a platform widely used to share AI models. According to DeMarco, Strada has already been downloaded and is being used, while the team is awaiting journal review for a publication in Astronomy and Computing. The work is also being presented internationally. The team has prepared three separate papers explaining different aspects of the model, including its architecture and the data used to train it, for the Machine Learning for Astrophysics conference later this year. DeMarco has also been invited to speak at the Spanish Advanced School of AI in Granada, where researchers will discuss the use of AI in astronomy. The team is now working on the next step: an enhanced version of Strada that will not only analyse radio astronomy images, but also generate them. The idea is to use AI to create synthetic examples of galaxy types that are underrepresented in existing archives, helping the model learn from categories where real data is limited. If successful, Strada V2 could become both an analysis and generative model, capable of understanding radio images and producing new training examples based on scientific prompts. DeMarco said AI is likely to become a standard tool in astronomy as next-generation observatories begin producing even larger quantities of data. "It has to be there," he said, adding that the design of future radio astronomy work already assumes that intelligent computation will play a role at some point in the process. Asked what he hopes Strada will achieve over the next five years, DeMarco said AI could help scientists notice something in the universe that humans might otherwise miss, not because humans lack the intelligence, but because they cannot realistically inspect everything. "There's nothing in theory that AI can do that a human cannot do," he said. "It is all about practicality. How fast can you do it? When can you do it?" Without AI, he said, some discoveries could take much longer - or never happen at all. "If you don't see all the data, then you might have missed out on something big," he said.