Self-Teaching AI Filters Jets From Heavy-Ion Collision Noise

The new framework skips that entirely. "The new framework doesn't need to be told what corresponds to a cat or a dog," Torbunov said. "Unsupervised learning methods can learn useful concepts and patterns directly from the data, without requiring humans to label every example."

The team calls it an "unpaired, unsupervised" generative model. They fed it two things: simulated detector data containing collision background, and separate, clean simulated jet images with no background. No one-to-one "answer key" was required. Brookhaven physicist Yeonju Go, also an assistant professor at Stony Brook University, put it this way: "The detector sees a crowded 'image' in which the signal we care about is buried inside a dense spray of other particles. Our AI method learns to split that image into two parts — the jet signal and the background — without requiring one-to-one 'answer key' examples that are not available in real experimental data."

What did the tests actually show?

The researchers tested the model on simulated data from calorimeters — detector components that measure particle energies — at sPHENIX, the detector that ran during RHIC's final three runs. In the simulations, the model reconstructed jets' positions, energies and internal structure more accurately than conventional background-subtraction methods, especially for larger jets, where background fluctuations hit hardest.

Read that carefully. Simulated data. The gains are documented in simulation, not yet in real collisions, and real sPHENIX data immerses jets in an extremely dense, fluctuating environment. Whether the accuracy holds there remains to be shown.

What does this change for physicists?

Potentially, a lot — if the model transfers. "Improving background subtraction can enable more precise measurements of how quarks and gluons interact under extreme conditions, helping scientists better understand the strong force and the behavior of matter in the early universe," Go said.

The next step is applying the model to real data. sPHENIX collected nearly 300 petabytes during its runs — a cache the lab compares to tens of billions of smartphone photos. "We definitely want to use the model on our experimental data, especially as sPHENIX has a very large data set," Go said. "This would be the first use of this kind of model in the nuclear and particle physics community, not just RHIC."

The researchers also see applications at the future Electron-Ion Collider, the Large Hadron Collider in Europe, and other experiments where separating signal from background is a fundamental challenge. Because the framework works directly with image-like detector data, they argue it could extend to any experiment disentangling overlapping signals. That's a claim about the future, and the team hasn't demonstrated it yet.

The work also reflects a broader push across DOE national laboratories to harness AI for scientific discovery. "This is not a one-time project," Go said of the physics-computer science collaboration, "and I'm very excited for the new opportunities that AI and machine learning in science will bring."