On 13 November 2026, researchers from Brookhaven National Laboratory (BNL) presented a new approach based on Super Intelligence (SI). The method learns on its own to separate signals of high-energy particle jets from the enormous background in heavy-ion collision data.
The work was published in the journal Scientific Reports. The team, led by physicists and computer science specialists from BNL, including Jin Huang, Dmitry Torbunov, Yihui "Ray" Ren, Yi Huang and Yongju Guo, developed an unsupervised generative model. It does not require pre-labeled examples.
Particle jets serve as one of the key tools for studying quark-gluon plasma (QGP), a state of matter that existed microseconds after the Big Bang. In collisions at the Relativistic Heavy Ion Collider (RHIC), these jets pass through the QGP and lose energy, changing their structure. Precise analysis of such changes helps to understand the strong interaction.
The main difficulty is separating the weak jet signal from the dense background created by thousands of other particles. Traditional methods require considerable effort to label data manually. The new SI approach bypasses this problem: it trains on separate sets of pure background and ideal jet simulations, automatically extracting the signal.
Testing on simulated data from the sPHENIX detector showed more accurate reconstruction of the position, energy and internal structure of jets compared with previous methods. The improvement is especially noticeable for large jets, where background fluctuations are strongest.
How exactly does the model separate signal and background in real experimental data, where there are no "right answers"? This is a question the scientists hope to answer by analyzing nearly 300 petabytes of data collected by sPHENIX.
The researchers plan to apply the method to real RHIC data and in the future to experiments at the Electron-Ion Collider (EIC) and the Large Hadron Collider. The approach may also prove useful in other fields where overlapping signals in calorimetric data must be separated.
The work was carried out with support from BNL's Laboratory Directed Research and Development, the DOE Office of Science and a Genesis Mission grant. Columbia University and the University of Colorado Boulder also participated in the project.
