Brookhaven AI Learns on Its Own to Pull Particle Jets Out of Early-Universe Noise
The unsupervised method beat conventional background subtraction in simulated sPHENIX data, and could sharpen measurements of quark-gluon plasma, the matter that filled the universe a microsecond after the Big Bang.

A team led by researchers at the U.S. Department of Energy's Brookhaven National Laboratory has built an artificial intelligence method that teaches itself to separate valuable particle-jet signals from the enormous background of debris produced when heavy ions collide at nearly the speed of light. The advance could let physicists measure the quark-gluon plasma, a hot, dense state of matter that existed a microsecond after the Big Bang, with greater precision.
At Brookhaven's Relativistic Heavy Ion Collider, or RHIC, physicists recreated early-universe conditions that produce the plasma, a "soup" of free quarks and gluons. RHIC, a DOE Office of Science user facility, finished operations in early 2026. Jets are narrow sprays of high-energy particles thrown out by energetic quarks and gluons, and they are among the best probes of the plasma. As a jet moves through it, its particles lose energy and its internal structure changes. Scientists want to study exactly those changes to learn how quarks and gluons behave and how the strong force, the strongest force in nature, binds them.
The obstacle is background. Each collision produces a dense spray of other particles, and separating the jet from that crowd has long limited the precision of jet measurements. "The detector sees a crowded 'image' in which the signal we care about is buried inside a dense spray of other particles," said Yeonju Go, a Brookhaven physicist and Stony Brook University assistant professor who works on RHIC data analysis.
The new approach, described in a paper in Scientific Reports, is a generative AI model that splits that image into two parts, the jet and the background. The team included researchers from Columbia University and the University of Colorado Boulder. It is the first use of an "unpaired, unsupervised" model on this problem, meaning it learned directly from simulated collision data without the one-to-one signal-and-background examples that do not exist in real experiments.
Brookhaven computer scientist Dmitrii Torbunov explained the difference. Traditional methods need humans to label huge training sets, much as a person would have to tag thousands of photos as cats or dogs. "The new framework doesn't need to be told what corresponds to a cat or a dog," he said. The model instead looks at a collection of pure background produced by the plasma and at ideal jet simulations, and works out how to extract the underlying jet while removing the background.
The researchers tested it on simulated data from the calorimeters of sPHENIX, a detector that ran during RHIC's final three runs. In those tests, the model reconstructed jets' positions, energies and internal structure more accurately than conventional subtraction methods, particularly for larger jets, where background fluctuations are strongest.
Torbunov said the approach suits science in general, because large backgrounds can make it hard or even impossible for a person to decide how much of each data point comes from signal. The team plans to apply the model to the real data sPHENIX collected, and possibly to future experiments at the Electron-Ion Collider, which is planned to succeed RHIC at Brookhaven.




