Skip to main content

Identifying neutrinos with machine learning

Research and discovery High performance systems Dec 17, 2019

You might remember learning about protons, neutrons, and electrons in Chemistry. There are other subatomic particles that might be less familiar, but that doesn’t mean we’re not surrounded by them!

After photons, neutrinos are the second most abundant particles in the universe. There are billions of them streaming through your skin as you read this, and according to Mark Messier, Professor of Physics at Indiana University, they’re important to how the universe evolves. Messier and his graduate students Micah Groh and Ryan Murphy study neutrinos to determine their patterns and types.

Mark Messier, IU Professor of Physics

To back up slightly, there are three types (or flavors) of neutrinos: electron, muon, and tau. However, as they travel from one place to another, they can change their flavor. Messier, Groh, and Murphy work on an experiment called NOvA that measures the rates of this shape-shifting. By looking at the patterns in these rates, they hope to determine if matter and antimatter neutrinos look the same, or if they differ from one another.

In order to detect these patterns, the team records the tracks that appear in their detector on the rare occasion when a neutrino does interact. According to Murphy, they then try “to identify each of those tracks, measure how much energy they have, and then put it all together to determine the type of neutrino that produced the interaction and how much energy it had when it entered the detector.”

Ryan Murphy, IU graduate student

Putting all of this together, Messier explains, is “a bit like putting a crime scene together after the crook has escaped out the window,” and machine learning acts as a forensic scientist, stringing clues together to reveal the neutrino’s identity. Millions of examples of neutrino interactions are fed through an algorithm so that it can learn each flavor’s identifying characteristics. “The Carbonate computer cluster here at IU has top-of-the-line GPU processors which are perfect for processing these examples quickly. Having this resource readily available at IU allowed us to try lots of ideas out so we could find the best ones,” explains Groh. Scott Michael, Manager of Research Applications and Deep Learning for UITS Research Technologies notes, “It’s these types of projects that caused us to deploy the CarbonateDL resource to meet the hardware need of modern high powered GPUs to run deep learning frameworks like PyTorch and Tensorflow.”

The distinction between matter and antimatter is largely arbitrary, akin to how 2 x 2 or -2 x -2 both equal 4. But something in the early history of the Universe tilted the balance toward matter over antimatter. Messier, Groh, and Murphy suspect that the difference might be linked to differences between neutrinos and antineutrinos.

These experiments require immense amounts of compute power and data storage. Neutrino events are rare, so the software used by Messier, Murphy, and Groh tries to make use of every bit of information the detector records. Groh notes that the next generation of neutrino detector will record interactions with millimeter precision, producing 20 petabytes of data per year. Current computing resources will allow scientists to ask ever more complex and detailed questions of this data.

More stories