A particle accelerator is like a bizarre ice cream shop where new flavors are created by smashing together existing ones. You take a sample of an ice cream you haven’t tried before, and before you can taste it, the swirl melts into the classic familiar flavors in less than a trillionth of a second. Physicists never directly observe the new flavor. Instead, they study what it melts into. From those traces, they reconstruct what must have existed for a fleeting instant.
The Standard Model is the recipe book—a theory that predicts which flavors should exist and how they should melt. So far, it has worked remarkably well. But some “melting patterns” hint that there may be ingredients we haven’t accounted for. Today, particle physicists test the Standard Model by reconstructing the fleeting lives of our universe’s fundamental building blocks, looking for physical behavior it cannot explain. In a new study published in Nature Communications, scientists at the particle accelerator at the European Organization for Nuclear Research (CERN) present a novel method for classifying jets of fundamental particles, speeding up the search for physics beyond the Standard Model.
The particles that we can see and interact with are called fermions, and they’re divided into two types: quarks and leptons. Each comes in six distinct “flavors,” a term coined in 1971 by physicist Murray Gell-Mann (YC ’48) and his student Harald Fritzsch while eating ice cream. Quarks cluster together to form protons and neutrons, and come in paired flavors: up and down, strange and charm, and top and bottom.
However, sampling these flavors takes a lot more than just scooping ice cream. Beneath the French-Swiss border, CERN operates the Large Hadron Collider, a twenty-seven-kilometer circular tunnel where two counter-rotating beams of protons race around at a speed only a millionth of a percent slower than the speed of light. At four points along the ring, the beams are steered into head-on collisions inside massive detectors. When these proton beams collide, new particles erupt from a point called the primary vertex. Quarks fragment into sprays known as jets, while other short-lived particles decay almost immediately. Some decays occur slightly away from the original point and sprout into secondary vertices like branching twigs on a tree.
ATLAS, one such detector, captures these events like a hundred- million-pixel camera. Devices called calorimeters absorb most light particles and measure their energy, while heavier particles pass through and leave traces in outer chambers. ATLAS records up to sixty million megabytes of data per second, yet only about one in ten thousand unique collision events survive a real-time trigger that discards the rest as noise. From those violent collisions, physicists infer a jet’s underlying quark flavor by examining displaced tracks and secondary vertices through a process called flavor tagging. Successful taggers must correctly identify real bottom quark (or “b-quark”) jets while rejecting impostors. In our ice cream shop, this means differentiating chocolate from vanilla, even when toppings might get in the way. When b-tagging was first used in a hadron collider to discover the top quark in 1995, Eugene Higgins Professor of Physics Paul Tipton was part of the Fermilab team behind the effort. “Our overall efficiency of tagging b-quarks was something like thirty percent,” Tipton said. “The whole thing was in its infancy back then.”
Now, the flavor-tagging process is being revolutionized by machine learning. In their new study, the ATLAS scientists, including contributors from Yale, present a new jet classification model called Graph Neural Network 2 (GN2). GN2 treats each jet as a network of connected tracks and vertices, learning the geometric aftertastes of heavy-flavor decays rather than evaluating tracks independently. Trained with sophisticated physical simulations and validated against real collision data, the GN2 model achieves seventy percent efficiency, much higher than previous iterations.
Despite its complexity, the algorithm has earned the confidence of more than three thousand physicists who comprise the ATLAS collaboration. Elise Le Boulicaut Ennis, a member of a Yale group led by Tipton and Sarah Demers, chair of the Department of Physics, said, “Trust is established by running validations and checks.” That trust becomes essential as ATLAS’s objectives become more complex. Physicists hope to observe Higgs bosons and measure how the Higgs couples to itself, a process whose signatures often include multiple b-quarks in a single event. “Something like these new taggers is really going to pay off,” Tipton said.
As the Large Hadron Collider undergoes upgrades and renovations, deeper integrations of machine learning are under discussion. “It’s a big conversation that’s happening right now,” Ennis said. “It may be that we’re going to head in that direction.” Yale continues to contribute both software and hardware to the ATLAS collaboration, analyzing particle collisions and upgrading the detector technology. As our soft-serve machines undergo maintenance, new flavors may emerge, forcing physicists to rewrite the menu.