As the SKACH Consortium meets for its Spring Meeting at the Zurich University of Applied Sciences, we deep dive into the work that ZHAW School of Engineering researchers are contributing globally to the Square Kilometer Array.
Currently under construction, by the end of the decade there will be hundreds of dishes in South Africa at the mid-frequency range and more than 130-thousand low-frequency antennas erected in Australia, making up the telescopes of the Square Kilometer Array Observatory.
This next-generation radio astronomy facility will look as far back as the Cosmic Dawn, when the very first stars and galaxies formed, tackling some of the most fundamental scientific questions of our time and, with an expected operational phase of at least 50 years, it will be one of the cornerstone physics machines of the 21st century.
Funded primarily by the State Secretariat for Education, Research and Innovation (SERI), Switzerland’s mission will contribute to providing data products for researchers so they can observe the universe. As one of the ten members of the SKA-Switzerland Consortium ZHAW is contributing its expertise in AI and machine learning and, even before the SKAO is functional and transmits data, there is a lot of work to be done.
ZHAW researcher Elena Gavagnin leads the project Mock-observation via generative Deep Learning,“The amount of data per second that will be generated is unprecedented. From an engineering point of view, one of the challenges is that this data needs to be processed to produce images.”
Since astrophysics cannot count on experiments to show how galaxies form, for example, scientists rely on simulations based on the laws of physics and implement astrophysical processes to study the observed phenomena and test hypotheses. However, the results of these simulations represent an idealised version of the astrophysical objects we have in the Universe. Telescopes provide observations far away from these “perfect” representations.
Elena Gavagnin, Frank-Peter Schilling and Philipp Denzel use generative deep learning techniques to study how the outcomes of ideal numerical simulations relate to more realistic observations and vice versa.
“Using machine learning, we can compare this ideal simulation with what the SKAO instruments will observe in reality. When we get the real observation of a galaxy, we can learn what the perfect galaxy behind would look like and infer properties from it, for example the dark matter distribution. To do this, we have a huge catalogue of simulated gas – the matter we see – and dark matter maps – the component of a galaxy that we cannot see. Our model is trained to reproduce the gas map when we give it the dark matter map and vice versa. This is like predicting the colour of an image from many black and white images,” Gavagnin explains.
Frank-Peter Schilling adds, “We are using state-of-the-art Diffusion Models, which excel at image translation tasks, in fact the same type of model which we also use to process medical images, a nice synergy.”
The ZHAW team is collaborating with the University of Zurich, which is in charge of providing the simulations, and the FHNW, which is developing a digital twin of the actual telescope.