Machine learning gives a better glimpse of galaxies

Telescopes are limited by the size of the mirror or lens they use. Using a form of artificial intelligence, researchers at the Swiss Federal Institute of Technology in Zurich (ETH Zurich) may have found a way to push past that limit, offering scientists the prospect of the sharper images in optical astronomy.

The diameter of a telescope’s lens or mirror, known as the aperture, can limit a telescope. The bigger the mirror or lens, the more light it gathers, allowing astronomers to detect fainter objects, and to observe them more clearly.

The study uses machine learning technology by teaching a neural network, a computational approach that simulates the neurons in a brain, what galaxies look like. It then asks it to automatically recover a blurred image and turn it into a sharp one. The neural net needs examples, such as a blurred and a sharp image of the same galaxy, to learn the technique.

The system uses two neural nets competing with each other, an approach called a generative adversarial network (GAN). The whole teaching programme took just a few hours on a high performance computer.

The trained neural nets were able to recognise and reconstruct features that the telescope could not resolve, such as star-forming regions, bars and dust lanes in galaxies. The scientists checked it against the original high-resolution image to test its performance, finding it better able to recover features than anything used to date, including the 'deconvolution' approach used to improve the images made in the early years of the Hubble Space Telescope.

The team now plans to train a GAN using simulated galaxies as opposed to real ones to further experiment with the project.

“The main challenge was the selection of a suitable selection of galaxies spanning the range of properties and features expected of galaxies so that the GAN could learn what galaxies look like, so that it can then infer what’s just below the resolution limit,” says Kevin Schawinski, lead author of the study.

The researchers are keen for others in the field to get involved in the technology, which is why they have made their code available at space.ml.

"There is no reason why we can't then apply this technique to the deepest images from Hubble, and the coming James Webb Space Telescope, to learn more about the earliest structures in the Universe," Schawinski adds.

The results of the project points to a more data-driven future for astrophysics in which information is learned automatically from data, instead of manually crafted physics models.

The study appears in the journal Monthly Notices of the Royal Astronomical Society.