CSIRO’s ASKAP radio telescope is used by researchers to create enormous maps of the Universe by repeatedly scanning the sky. To generate these maps ASKAP collects data at the rate of 100 trillion bits per second – more data at a faster rate than Australia’s entire internet traffic. Supercomputers, like Setonix at the Pawsey Supercomputing Research Centre, then store, calibrate and transform this data.
By using machine learning and AI tools, astronomers can circumvent the need to store massive quantities of data for long periods, be alerted to interesting data, or collect similar objects that may be in very different parts of the Universe.
CSIRO researchers have been using machine learning and AI to identify unusual shapes and structures that might point to new physical phenomena, like the image of an ‘odd radio circle’, first discovered in ASKAP data for the Evolutionary Map of the Universe project (EMU), one of ASKAP’s nine survey science projects.
Machine learning and AI can also be used to classify different types of radio sources into categories, something that would take too long to do by eye.
One such tool is the EMU Search Engine (EMUSE), developed to identify galaxies with rare or complex shapes without manually browsing thousands of images. Currently, EMUSE contains images captured by ASKAP from across the southern sky, along with corresponding images from the Wide-field Infrared Survey Explorer (WISE) space telescope.
EMUSE allows researchers to find galaxies using either text- or image-based prompts. This was achieved by training a large vision-language model capable of interpreting both descriptive text and images, enabling the system to understand inputs of descriptive language and imagery in the search function.
When a screenshot of an FR-II radio galaxy was uploaded using the image upload option in the EMUSE app, it output a catalogue of similar FR-II radio galaxies from the EMU survey. The top 50 similar sources are plotted here.
The tool was also built to require significantly less data storage than typically needed of such a large archive, thanks to its prioritising of image features. The trained model picks out the most important features of any galactic image – like shapes, colours or patterns – and represents them as a list of numbers, reducing the survey data from hundreds of gigabytes to just tens of megabytes. Using AI in this way has enabled the team to accelerate the research and enhance opportunities to scale. This model and image features were then integrated into a publicly accessible online platform.
The platform can autonomously retrieve, classify, and catalogue radio sources — such as bent-tail galaxies, odd radio circles, or galaxies with intense star-forming regions — in a fraction of a second, a task that would otherwise take astronomers weeks or even months to complete manually. However, work continues to enable the tool to identify rare or previously unseen radio sources, which will improve as the AI models are refined and the EMU survey expands in the coming years.
Using EMUSE to explore the different galaxies in our Universe is a beautiful and awe-inspiring opportunity, but it also offers practical solutions to Earth-bound challenges. Similar machine learning and AI models can be applied to other fields where shapes and structures need analysis, such as medical imaging.
Read the full paper in Publications of the Astronomical Society of Australia.
We acknowledge the Wajarri Yamaji as the Traditional Owners and Native Title Holders of Inyarrimanha Ilgari Bundara, our Murchison Radio-astronomy Observatory, where the ASKAP radio telescope is located.