Synthesis imaging by Radio Interferometry (RI) is a powerful technique in astronomy that leverages antenna arrays to observe the sky with otherwise inaccessible resolution and sensitivity. It has revolutionised astrophysics and cosmology since the 1940s, with the 1974 Nobel Prize for Physics awarded to Ryle and Hewish for its development. A new generation of RI telescopes is currently emerging. Among these, the Square Kilometre Array (SKA) will soon be to radio astronomy what CERN is to particle physics, with science goals ranging from studying cosmic magnetism, dark matter, and dark energy, to understanding the structure and evolution of stars and galaxies.
The mathematical inverse problem for the formation of images from the data acquired by RI telescopes is extremely challenging, as observational constraints imply data incompleteness, calibration challenges, measurement noise, and other sources of uncertainty. Bespoke computational imaging algorithms are required, capable of “regularising” the problem by injecting powerful image models into the data. Current RI imaging methodology, such as the famous CLEAN algorithm that has served the field for decades, fails to deliver the unprecedented regime of joint precision, robustness, scalability, and efficiency that modern telescopes demand. An effort commensurate to the billions of pounds invested in the development and operation of such instruments is warranted for the emergence of an imaging algorithm realising their full potential.
This project brought together the multi-disciplinary expertise of the Laboratories of Profs Thiran and Kneib at EPFL and Prof. Wiaux in Edinburgh to develop an AI version CLEAN specialized to the formation of images on extremely wide fields of view on the celestial sphere. The work, submitted for publication in one of the premier journals of astronomy, demonstrates that the new method delivers much higher precision and speed than CLEAN.
