Quantifying vacuum-like jets in heavy-ion collisions: a Machine Learning study

TL;DR

The paper tackles the problem of quantifying jet modification in heavy-ion collisions by distinguishing vacuum-like jets from medium-modified ones on a jet-by-jet basis. It introduces a Transformer classifier that ingests low-level jet constituents to access information beyond standard high-level observables, demonstrating improved discrimination when medium response is present. The study finds near-equivalent performance of Transformer and BDT in vacuum-like scenarios, but the Transformer outperforms in MR-UE by capturing medium-induced correlations, enabling a robust upper bound on the fraction of vacuum-like jets (just under $0.3$ in SO and just under $0.2$ in MR-UE). This approach provides a data-driven way to quantify jet quenching components and highlights the value of low-level jet information for understanding QGP interactions. The methodology, including data simulation with Jewel 2.3 and a detailed uncertainty-aware bound, has implications for separating vacuum-like and medium-modified jet populations in experimental analyses.

Abstract

The modification of jets by interaction with the Quark Gluon Plasma has been extensively established through the comparison of observables computed for samples of jets produced in nucleus-nucleus collisions and proton-proton collisions. The presence of vacuum-like jets, jets that experienced little interaction with the Quark Gluon Plasma, in the nucleus-nucleus samples dilutes the overall observed modification hindering the detailed study of the underlying physical mechanisms. The ability to ascertain on a jet-by-jet basis the degree of modification of a jet would be an invaluable step in overcoming this limitation. We consider a Transformer classifier, trained on a low-level representation of jets given by the 4-momenta of all its constituents. We show that the Transformer is able to capture discriminating information not accessible to other architectures which use high-level physical observables as input. The Transformer allows us to identify, in the experimentally relevant case where both medium response and underlying event contamination are accounted for, a class of jets that have been unequivocally modified. Further, we perform a robust estimate of the upper bound for the fraction of jets in nucleus-nucleus collisions that are, for all purposes, indistinguishable from those produced in proton-proton collisions.

Figures (8)

• Figure 1: ROC curves and respective AUC for the BDT (Blue) and Transformer (Green) for SO scenario.
• Figure 2: BDT and Transformer predictions for the SO scenario. Left: Difference of the 2d histogram density between Vacuum and Medium samples, with Blue (Red) showing a greater density of Vacuum (Medium) density. Middle: The 2d histogram density for the Vacuum sample. Right: The 2d histogram density for the Medium sample.
• Figure 3: ROC curves and respective AUC for the BDT (Blue) and Transformer (Green) for MR-UE scenario.
• Figure 4: BDT and Transformer predictions for the MR-UE scenario. Left: Difference of the 2d histogram density between Vacuum and Medium samples, with Blue (Red) showing a greater density of Vacuum (Medium) density. Middle: The 2d histogram density for the Vacuum sample. Right: The 2d histogram density for the Medium sample.
• Figure 5: Left: the evolution of $f^{\text{max}}_{\mathcal{D},v}$ for the ten stringiest discriminants with the number of bins. Right: dependence of $f^{\text{max}}_{\mathcal{D},v}$ on the discriminant discriminating power, as measured by the ROC AUC, for 100 bins for all discriminants. Both plots for the SO scenario.