Scientists from the Brookhaven National Laboratory of the US Department of Energy, together with colleagues from Columbia University and the University of Colorado Boulder, have developed an artificial intelligence method that independently learns to separate jet signals - narrow streams of high-energy particles - from the dense background recorded in heavy-ion collisions. This is reported in the journal Scientific Reports.
Such collisions reproduce the conditions of the early Universe, when quark-gluon plasma (QGP) arose - a hot and dense state of matter in which quarks and gluons exist freely rather than being bound into protons and neutrons. QGP existed approximately one microsecond after the Big Bang. To study it, physicists use jets - streams of particles that, passing through the plasma, lose energy and change their internal structure. It is precisely these changes that can reveal the properties of the strong interaction - the most powerful force in nature, which holds quarks together.
The main difficulty is to isolate the jet signal against the background of a huge number of other particles. Traditional methods require manual labeling of data, which limits accuracy. The new model belongs to the class of "unpaired unsupervised learning": it does not need pre-labeled examples in which the background is known for each signal. Instead, the algorithm learns from simulated data containing both background and separate "clean" images of jets, and finds a way to separate them on its own.
"The detector sees a crowded picture in which the desired signal is hidden in a dense stream of other particles," explained Yen-Ju Guo, a physicist at the laboratory and a professor at Stony Brook University. "Our method learns to split the image into two parts - the jet signal and the background - without the need for paired examples, which simply do not exist in real data."
The researchers tested the model on simulated data from the sPHENIX detector, which operated at the RHIC collider. In these simulations, the algorithm reconstructed the position, energy and internal structure of jets more accurately than standard background subtraction methods. The advantage was especially noticeable for large jets, where background fluctuations are strongest.
The authors emphasize that so far this concerns only modeling. The next step is to apply the method to real data collected by sPHENIX (almost 300 petabytes, comparable to tens of billions of smartphone photographs), and in the future to future experiments at the Electron-Ion Collider (EIC) and the Large Hadron Collider. According to the scientists, the approach could also be useful in other fields where overlapping signals need to be separated.
"Improved background subtraction will make it possible to measure more accurately how quarks and gluons interact under extreme conditions, and to better understand the strong interaction and the behavior of matter in the early Universe," Guo noted.
Note that this concerns a statistical improvement in data processing, not proof of any new properties of the plasma. Nevertheless, if the method proves its effectiveness in real experiments, it could become an important tool for high-energy physics.