Unravelling the secrets of the universe: Deep learning in astronomy

The field of astrophysics will soon be radically transformed with the opening of a next-generation space telescope at the Vera C. Rubin Observatory in Chile, which is expected in 2025 or 2026. Thanks to the Legacy Survey of Space and Time (LSST), which the telescope will deploy, we will be able to map the entire sky over the Southern Hemisphere every three days. 

This has researcher Laurence Perreault-Levasseur excited, not only about the advances and discoveries the telescope will bring, but also because it will open up a whole new field of research combining deep learning with the study of the universe. 

Demystifying an astrophysical effect 

Perreault-Levasseur works at Montréal’s Mila Artificial Intelligence Institute and is the Canada Research Chair in Computational Cosmology and Artificial Intelligence, and an Assistant Professor in the Physics Department at Université de Montréal . She’s studying an exciting phenomenon in astrophysics known as strong gravitational lensing.

To date, astrophysicists have found and studied about 1,000 strong gravitational lenses using traditional methods of observation. The LSST will provide newer and more frequent maps, which could bring that count up to over 170,000, with each lens generating thousands of data points. Traditional analysis methods will prove insufficient to tackle all the data points, which could quickly number in the billions, but Perrault-Levasseur is confident that she has the right tools to keep pace.

Gravitational lensing is a phenomenon that occurs when two massive celestial objects — such as galaxies, galaxy clusters, or star clusters — are observed one in front of the other. The closer object’s mass creates a gravitational lens, which causes the light from the farther object to bend around the closer object. So, when observed through a telescope, the light comes at the observer from several different angles. In strong gravitational lensing, this results in multiple images of the same object — like seeing double. The images can also appear as arcs or rings around the closer object. To demystify this phenomenon, try holding the foot of a wine glass in front of a candle flame or looking at a fish through the corner of a glass tank.

Essentially, the closer object is acting as a giant lens. Strong gravitational lensing can help astrophysicists study the farther object — the candle flame — because its image is being magnified, or the closer object — the wine glass — because of the distortions it causes. 

Perreault-Levasseur is interested in the latter, as her work involves mapping the distribution of matter within gravitational lenses. In the wine-glass metaphor, she’s trying to determine whether the glass is smooth or rough, like frosted glass.

Strong gravitational lensing can also advance the study of the universe, known as cosmology, by telling us more about the distribution of dark matter. Although this mysterious substance accounts for 80 percent of all matter in the universe, it remains vastly unknown because we can’t observe it directly. Learning about its properties or characteristics is a major priority in this field of study.

“We’re very hopeful that we’ll be able to learn more and solve some of the mysteries of the creation of the universe — we might be able to make better cosmological simulations of our universe and the formation of galaxies,” says Perreault-Levasseur enthusiastically. So strong gravitational lensing may just be the gateway to unravelling some of the greatest mysteries of the universe!

Where machine learning meets human innovation

Before data from the Vera C. Rubin Observatory comes pouring in, Perreault-Levasseur and her team of around twenty researchers, mostly PhD and postdoctoral students, are hard at work creating and testing new models to analyze existing data on the thousand or so strong gravitational lenses known to science.

They’re developing deep learning algorithms in astronomy, which will allow them not only to manage but to analyze massive amounts of data. When the incoming data from the LSST arrives, the team wants to be ready for it. Their innovative method will need to be replicable and adaptable to each of the roughly 170,000 strong gravitational lenses expected to be found.

The team will use CFI-funded graphics processing units (GPUs), the fundamental infrastructure needed to supply the phenomenal computing power that makes deep learning possible.

And phenomenal power means phenomenal speed. Compared to traditional methods, cosmological data analysis for strong gravitational lenses will be more precise and much faster — 10 million times faster. Perreault-Levasseur estimates that the AI-based analysis methods being developed will bring the time needed to process the data for all 170,000 strong gravitational lenses down from 1,400 years (three days for each one) to 30 minutes total.

Cosmological research may seem beyond the grasp of many us mere Earthlings, but its impact couldn’t be closer to home. The beauty of research like Perreault-Levasseur’s is that the analysis methods and machine learning models she’s developing have potential applications in other fields, such as medical imaging and climate modelling. “The structures of the problems are very, very similar,” she says. “If we invest in solving problems in astrophysics, it could have a huge impact on completely different fields that affect us very directly, such as healthcare.”

Thanks to CFI funding, we’ve been able to acquire essential computing resources to develop the next generation of cosmological analysis algorithms. We are in an era of precision science, and we need the latest GPUs to be able to analyze our images with the greatest possible precision. It’s crucial that we have the right tools for the job, and this computer infrastructure is exactly that.

- Laurence Perreault-Levasseur, Université de Montréal and Mila