UNIVERSITY PARK, Pa. — In the "Princess and the Pea" fairy tale, a young woman must prove her status as royalty by detecting whether a pea was placed under dozens of mattresses during her sleep. What if she used deep learning to detect that pea? That’s the approach taken by Doug Cowen, professor of physics and astronomy & astrophysics, and his research group, who are trying to detect nearly imperceptible particles, called neutrinos, arriving at the Earth from distant galaxies. Thanks to Penn State’s growing partnership with technology leader IBM, the Cowen group is now able to scale up their research to detect more neutrinos, helping to further improve our understanding of the cosmos.
The Princess and the Neutrino
If sensing a pea under a pile of mattresses sounds impossible, consider the task of detecting neutrinos. They are smaller than atoms, they don’t have a charge, they travel near the speed of light, they are nearly massless and, most importantly, they only interact with two of the four fundamental forces in the universe. These characteristics lead to a behavior that make them even harder to observe than a single vegetable seed — they pass directly through matter without leaving much of a trace. In fact, billions zing through us at every moment, unseen and unheard. Thought to be one of the most prevalent subatomic particle in the universe, neutrinos are created during natural atomic decay processes, in supernovae, or from nuclear reactions in the cores of stars.
By better understanding how, when and why neutrinos from outside the Solar System arrive on Earth, scientists can paint a clearer picture of our universe’s history, and peek into the inner workings of processes such as nuclear fusion or supernovas. But before neutrinos can be studied, scientists need to develop a reliable way to detect them.
The first neutrino was detected in a laboratory in the 1950s, but it wasn’t until 2018 that a high-energy neutrino from outside our solar system was first detected on Earth. That breakthrough came seven years after several hundred scientists got a sensitive detector up and running below the ice in the South Pole — the IceCube collaboration, funded by the National Science Foundation, of which Cowen and his team are members.
Buried deep under ice in the South Pole, IceCube is a sophisticated observatory that filters out background noise to isolate only the neutrinos that make it through layers and layers of ice. There, ultrasensitive photodetectors sense when a particle traveling near the speed of light emits faint amounts of bluish light known as Cherenkov radiation.
But because neutrinos are so prevalent, and many that we detect are created in the Earth's atmosphere, part of the challenge for the Cowen team is further sifting out the astrophysical neutrinos from those originating nearby.
Neutrinos by any other flavor
A highly reliable way to know if a neutrino comes from a distant cosmological event is to identify the type or “flavor” of neutrino — electron neutrino, muon neutrino, or tau neutrino.
Based on our understanding of the universe, Cowen said, tau neutrinos have a very high likelihood of originating from outer space and not from the Earth’s atmosphere.
“We’ve seen dozens of electron flavor and dozens of muon flavor neutrinos, but there has basically been, in history of mankind, only one or two astrophysical tau neutrinos detected — we want to double or triple that,” said Cowen. “The sample of these events is very, very small, and at present IceCube is the only experiment that can see them.”
Just as it’s incredibly difficult to detect a neutrino in the first place, knowing what type of subatomic particle is detected is a nearly insurmountable challenge.
“There’s a difference in the signature we see from certain flavors, particularly electron neutrinos and tau neutrinos at high energies,” said Aaron Fienberg, postdoctoral scholar in Cowen’s research group. “It’s very subtle, so it’s difficult to manually construct an algorithm to distinguish between them.”
Enter deep learning.
The team takes IceCube observational data and represents that as a two-dimensional image, much like a graph with data plotted on X and Y coordinates. Using a type of deep learning known as a convolutional neural network, the team is training a computer model to correctly identify what type of neutrino was detected based on the subtle signatures each flavor leaves. They train the model with “events,” or signatures made by different flavors of neutrinos detected by IceCube, and the model teaches itself what features of the images are most salient for classification.
“You can stare at these different events and manually separate them from the backgrounds, or you can feed the whole thing to the neural network and let the network do that heavy lifting for you,” Cowen said.