Probing the Structure of a Perfect Liquid Using Jets and Machine Learning
Aaron Angerami | 20-ERD-041
Relativistic heavy-ion collisions produce a hot and dense medium of deconfined quarks and gluons known as the quark-gluon plasma (QGP). These collisions provide a unique opportunity to study the dynamics of the strong nuclear force. Energetic jets of particles are produced at the earliest stages of the collision. These jets traverse the QGP and become attenuated through the phenomenon of jet quenching, which provides a means to study the microscopic structure of the QGP. Although jet energy loss is an established feature of heavy-ion collisions, how the energy loss of a given jet depends on its properties and internal structure is not well understood. Measurements of such dependence are limited by the present experimental capabilities in reconstructing the energies and angles of individual particles within the jet.
This project contains two components. The first data obtained by the ATLAS detector at the Large Hadron Collider to measure the energy loss as a function of the opening angle of the jet's particle spray. The results show that the number of jets of a given energy is reduced in lead–lead collisions compared to proton–proton collisions, and that this suppression increases as the opening angle increases. This suggests that wider jets lose more energy than narrow ones, which may be attributed to more of the particles within the jet being individually resolved in interactions with the medium. This mechanism has long been suspected but never explicitly observed, and it may be exploited to further study the dynamics of the plasma on short length scales.
The second component seeks to improve experimental capabilities in jet substructure measurements by applying machine learning (ML) techniques to reconstruct particles from energy deposits in the ATLAS calorimeter. Deep neural networks, which are well suited to problems with high-dimensional inputs, were trained using the full three-dimensional energy deposition information. Models using a convolutional architecture were found to dramatically improve the performance of particle identification (classification) and energy determination (regression) over current methods that only use a subset of the energy deposition information. Models using graph neural networks were also developed and showed even better performance. These models also naturally accommodate the non-uniform segmentation of the ATLAS calorimeter and may be straightforwardly extended to include complementary information from ATLAS' charged-particle tracking system. Although these improvements were not ultimately needed to perform the jet quenching measurement, they will be leveraged in subsequent analyses studying jet substructure in greater detail.
This work addresses a major scientific question highlighted in the 2015 Long Range Plan for Nuclear Science. This document is the primary strategic planning document for DOE-SC Nuclear Physics. It furthers the objectives set forth in the Laboratory's Investment strategy for science and technology of "probing the forms of quark and gluonic matter" (pg. 26) advancing the core competency areas of Nuclear, Chemical, and Isotopic Science and Technology. The ML component of the project has established LLNL as a leader in applying ML methods to experimental nuclear and particle physics. This led directly to a successful DOE-SC NP proposal with LBNL and UC Riverside: AI-Driven Detector Design for the EIC.
Publications, Presentations, and Patents
ATLAS Collaboration, 2023. "Measurement of Substructure-Dependent Jet Suppression in Pb+Pb Collisions at 5.02 TeV with the ATLAS Detector." Physical Review C. 25 January 2023; https://doi.org/10.48550/arXiv.2211.11470.
ATLAS Collaboration. 2020. "Deep Learning for Pion Identification and Energy Calibration with the ATLAS Detector." 12th International Workshop on Boosted Object Phenomenology, Reconstruction, and Searches in HEP (BOOST 2020). Virtual. 20-24 July 2020. https://cds.cern.ch/record/2724632, 2020.
ATLAS Collaboration. 2022. "Point Cloud Deep Learning Methods for Pion Reconstruction in the ATLAS Experiment", ATL-PHYS-PUB-2022-040 (2022). https://cds.cern.ch/record/2825379.
Dhanush Hangal. "Jet substructure measurements in heavy-ion collisions with ATLAS." Invited talk at the 14th International Workshop on Boosted Object Phenomenology, Reconstruction, Measurements and Searches in HEP (BOOST 2022), Hamburg, Germany. Aug 2022.
Hangal, D. 2022. "Jet Substructure in Heavy-Ion Collisions." 14th Conference on the Intersections of Particle and Nuclear Physics (CIPANP 2022), Lake Buena Vista, Florida. August 2022.
Karande, P. 2022. "Pion Reconstruction in the ATLAS Detector Using Graph Neural Networks." Machine Learning: Science and Technology, 2021001, ATL-PHYS-PUB-2020-018.