Could you briefly introduce yourself and your background?

I am Emma Tolley, assistant professor at EPFL in Switzerland. My background is in particle physics. I did my PhD on the ATLAS detector, looking for evidence of dark matter production at the Large Hadron Collider. During this work, I became very interested in cosmological limits and interpretations of dark matter, and I wanted to continue working on this question from an astrophysics perspective.

I joined EPFL in 2020 as a data scientist, working on high-performance computing challenges related to the Square Kilometre Array Observatory. Since then, I have worked on several projects and progressed to my current position as assistant professor.

Would you describe your work as being at the interface between physics, astronomy and HPC?

Yes. I consider myself a physicist, but I really like technical challenges with HPC, trying to use accelerators and dealing the data processing challenge. My group is maybe like 50% astronomers and astrophysicists, and 50% HPC experts and engineers. This combination is essential because, in radio astronomy, the computing challenge is directly linked to the scientific challenge. To produce very high-resolution images, we cannot simply build a 100-kilometre radio dish. Instead, instruments such as the SKA rely on many smaller antennas. The trade-off is that the imaging process becomes extremely computationally intensive. HPC and software development are therefore not separate from the science; they are part of what makes the science possible.

« The main topic we are working on in ODISSEE is data compression. »

Why is this particularly important for dark matter research?

Dark matter is one of like the major open questions in physics. We think that it makes up 85% of the matter content of the universe, but we have no idea what it is. With current instruments, though, we can see it out there. We see it interacting gravitationally through some of these gravitational lensing measurements that we’ve done.

In particle physics, a big part of what we were trying to do is really understand the interactions, so the forces and matter content of sort of everything in the universe. So dark matter was a huge missing piece of that. We don’t know if it’s like a microscopic black hole, or if it’s actually some kind of standing energy wave, or if it’s something like a sort of like dark proton.

Radio astronomy is particularly important because it can provide new observational constraints. Given that we still do not know what dark matter is, these constraints are crucial for narrowing down possible explanations. For example, one of my main interests is the Epoch of Reionization, the period when the first stars and galaxies formed. This is one of the major science goals of the SKA. It requires very long observations, around 1,000 hours, and therefore produces extremely large volumes of data. So efficient data compression could become essential to make these analyses feasible.

How is EPFL addressing this data bottleneck in ODISSEE?

We are exploring two main approaches. The first is to develop GPU implementations of standard data compression algorithms used in radio astronomy. The idea is to read compressed data onto the GPU and decompress it directly there. This helps reduce the bandwidth limitations that usually occur when moving data between CPU and GPU.

The second approach is to explore AI-based data compression methods. Neural representations have already shown very impressive compression rates in other fields, such as climate and weather simulations. In radio astronomy, the data is very noisy. What matters most is preserving the underlying signal, while much of the thermal noise is less useful. If we can compress data while preserving the scientifically relevant information, we can produce datasets that are much easier to store, move and process.

How do you ensure that compression does not damage the science?

This is a good question. We test the methods on both real and simulated data. With real observations, we check whether the final image and the science target are preserved after compression. For instance, we are testing datasets containing small transient radio sources that appear and disappear. These signals require integration over time to become visible, so they are a good test of whether compression removes important information.

We also use simulated radio data for cosmological-style sources, where the objective is to preserve the statistical properties of the sky. We need to ensure that compression does not introduce artificial structures or artefacts, as can happen with image compression formats.

« Solutions such as the compression methods explored in ODISSEE will be needed to carry out some of the most ambitious experiments proposed. »