Physics Professor Joins $15 million NSF Institute to Develop Accelerated AI-Driven Algorithms

Two people in protective clothing crouch near four large tanks
A look inside ProtoDune, a single-phase liquid argon time projection chamber, the type of neutrino detector that Scholberg’s research will focus on. Image: Maximilien Brice/CERN

Something hits a gigantic underground tank, flashes dimly for a few seconds and you have 15 minutes to figure out where it came from. The catch? It could have come from anywhere in the universe.

We’re talking about neutrinos, the lesser-known cousin of protons and electrons, and how scientists can use them to detect a supernova, the spectacularly explosive (and very rare) death of a star.

Thanks to a newly funded National Science Foundation Institute, Kate Scholberg, Arts & Sciences Distinguished Professor of Physics, is aiming to develop ultra-fast AI algorithms that could quickly pinpoint where neutrinos come from, and therefore help astronomers locate the supernova in time to observe it.

When stars die and collapse, neutrinos are expelled in very large numbers. Unlike electrons and protons, they have no electric charge and almost imperceptible mass, so they don’t interact with things the way other particles do.

Kate Scholberg headshot
Kate Scholberg, Arts & Sciences Distinguished Professor of Physics

Like infinitesimally small and invisible ghosts, they can go through matter, including stars and planets, aren’t slowed down by friction and have only a very faint relationship with gravity. This means that when they are projected from somewhere, anywhere, in the universe, they will travel for billions of miles very rapidly, following their initial trajectory unbothered by anything in their path. 

The only way to detect these neutrinos is to employ aptly named neutrino detectors. With a bit of luck, some of the particles scattered after a supernova will cross neutrino detectors here on Earth and scientists can begin their work.

Scholberg is most interested in a kind of neutrino detector called a “liquid argon time projection chamber,” which are in fact gigantic underground tanks filed with argon, an inert gas, in its liquid form. When argon atoms are hit by these neutrinos, it generates a very dim flash of light and a small electric discharge.

“They make distinctive little stubs and flashes in the argon, the size of a toy sparkler,” Scholberg said.

Neutrinos can come from other places beyond a supernova. They are also not the only particles hitting those argon gas tanks. Isolating neutrinos from noise, figuring out their source and determining if and where a star is dying requires very fast data processing, analysis and interpretation. Enter the artificial intelligence algorithms that Scholberg’s newly-funded institute will study.

A red, starry section of the universe
Remnants of supernova SN 1987A, taken by the NASA/ESA Hubble Space Telescope in 2017. SN 1987A is located in the center of the image, surrounded by a bright red ring. Image: Nasa/European Space Agency.

Called Accelerated AI Algorithms for Data-Driven Discovery (A3D3), it represents a collaboration between five institutions, including Duke. Led by Shih-Chieh Hsu, associate professor at the University of Washington, A3D3 aims at targeting problems in high-energy physics, multi-messenger astrophysics and systems neuroscience by developing AI tools able process very large datasets in real-time.

“It will be very interesting to figure out ways to collaborate and cross-pollinate across these different research nodes,” said Scholberg.

“I’ve worked with some of these collaborators before, but others I don't know at all,” she said. “It's fun and interesting to get a bigger picture and see some of the applications that these accelerated machine-learning techniques and technologies can have.”

A3D3 is one of five institutes recently funded by NSF’s Harnessing the Data Revolution initiative. With a total investment of $75 million, this initiative promotes collaborations between scientists and engineers to foster the development of data-intensive tools and methodologies.

Hilmar Lapp, Director of Informatics at the Duke Center for Genomic and Computational Biology (GCB), is part of another of these five institutes, called “Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning.” Read about it on the GCB’s website.