Identifying quenched jets in heavy ion collisions with machine learning

Aug 4, 2026

Upload by Hazel

Measurements of jet substructure in ultra-relativistic heavy ion collisions indicate that jet showering is modified by the interaction with the quark-gluon plasma. Hard substructure changes can be investigated using data-driven techniques, and this work develops a machine learning method to identify quenched jets. Jet showering is simulated with Jewel (quenching) and Pythia 8 (no quenching), while sequential angular-ordered substructure variables train an LSTM-based neural network. The method detects quenching despite the large uncorrelated soft-particle background.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

AccessFree to download

Introduction

Data sample

Supervised machine learning

Feature engineering

Training of LSTM-based neural network

Hyper-tuning

Robustness

Results

Conclusion

Training with Jewel Vacuum Events

Published for SISSA by Springer Received: July 14, 2022 Revised: March 9, 2023 Accepted: March 30, 2023 Published: April 28, 2023 Identifying quenched jets in heavy ion collisions with machine learning Lihan Liu,a Julia Velkovska,a Yilun Wua and Marta Verweijb a Department of Physics and Astronomy, Vanderbilt University, PMB 401807, 2301 Vanderbilt Place, Nashville, TN 37235, U. S. A. b Department of Physics, Utrecht University, Heidelberglaan 8 3584 CS Utrecht, The Netherlands E-mail: [lihan.liu@vanderbilt.edu](lihan.liu@vanderbilt.edu), [julia.velkovska@vanderbilt.edu](julia.velkovska@vanderbilt.edu), [yilun.wu@vanderbilt.edu](yilun.wu@vanderbilt.edu), [m.verweij@uu.nl](m.verweij@uu.nl) Abstract: Measurements of jet substructure in ultra-relativistic heavy ion collisions suggest that the jet showering process is modi􀀜ed by the interaction with the quark-gluon plasma. Modi􀀜cations of the hard substructure of jets can be explored with modern datadriven techniques. In this study, a machine learning approach to the identi􀀜cation of quenched jets is designed. Jet showering processes are simulated with a jet quenching model Jewel and a non-quenching model Pythia 8 . Sequential substructure variables are extracted from the jet clustering history following an angular-ordered sequence and are used in the training of a neural network built on top of a long short-term memory network. We show that this approach successfully identi􀀜es the quenching e􀀛ect in the presence of the large uncorrelated background of soft particles created in heavy-ion collisions. Keywords: Jets and Jet Substructure, Quark-Gluon Plasma ArXiv ePrint: 2206.01628 Open Access, 􀀍c The Authors. Article funded by SCOAP3 . [https://doi.org/10.1007/JHEP04](https://doi.org/10.1007/JHEP04) (2023)140 J HEP04(2023)140 Contents 1 Introduction 1 2 Data sample 3 3 Supervised machine learning 7 3.1 Feature engineering 7 3.2 Training of LSTM-based neural network 8 3.3 Hyper-tuning 10 3.4 Robustness 11 4 Results 13 5 Conclusion 13 A Training with Jewel Vacuum Events 16 1 Introduction Highly energetic partons (quarks and gluons) are expected to lose energy while traversing the extremely hot and dense quark-gluon plasma (QGP) created in ultra-relativistic heavy ion collisions [1] . This phenomenon, known as 􀀐jet quenching􀀑, was discovered at the Relativistic Heavy Ion Collider (RHIC) through the observation of suppression of high transverse momentum (pT ) hadrons [2􀀕5] and back-to-back dihadron correlations [6] . These observations were con􀀜rmed at the Large Hadron Collider (LHC) and extended into a larger kinematic range [7􀀕11] . In addition to measurements on high pT hadrons, which are the leading fragments of jets, the LHC opened new opportunities for direct observations of jet quenching and jet modi􀀜cations in the QGP. Signi􀀜cant dijet transverse momentum asymmetry [12􀀕15] and suppressed jet production [16􀀕18] were observed in lead-lead (PbPb) collisions. Jet-hadron correlations gave insights into how the energy lost by the jets is distributed in the medium produced in the collisions. Measurements of the jet shapes (transverse momentum radial pro􀀜le) [19􀀕21] and the jet fragmentation functions [22, 23] were performed to further investigate the mechanism of the jet shower modi􀀜cations in the QGP. To quantify the medium modi􀀜cations, measurements in proton-proton (pp) collisions [24􀀕32] are used as a reference for jet modi􀀜cations in the medium. The jet splitting function has been measured in both pp and PbPb collisions [33􀀕35], through the usage of a jet grooming algorithm that is able to split (􀀐decluster􀀑) a single jet into two subjets and locate the hard splitting by removing softer wide-angle radiation contributions. The hard splitting, characterized by J HEP04(2023)140 two well-separated subjets (􀀐two-prong􀀑 structure), provides access to the early stages in the parton shower evolution. In recent years many m

FAQs

What physics problem does this study address?

It aims to identify quenched jets and quantify modifications to jet showering induced by the quark-gluon plasma in ultra-relativistic heavy ion collisions.

How is jet quenching modeled for training and evaluation?

Jets are simulated using Jewel with a quenching model and Pythia 8 without quenching, providing paired examples for supervised learning.

What machine learning approach is used to classify quenched jets?

Sequential jet substructure variables extracted from an angular-ordered clustering history are used to train a neural network built on a long short-term memory (LSTM) architecture.

Does the method remain effective in realistic heavy-ion backgrounds?

Yes. The approach is shown to successfully identify the quenching effect even in the presence of a large uncorrelated background of soft particles.

Aug 4, 2026

Upload by Hazel