Speaker
Description
The Monte Carlo event generator JEWEL (Jet Evolution With Energy Loss) provides a
theoretical framework to simulate parton energy loss in a dense QCD medium such as the
Quark–Gluon Plasma (QGP). Traditionally, comparisons between theory and experiment
rely on global observables, including the nuclear modification factor $R_{AA}$ and the
dijet momentum imbalance $x_J$. In this work, we take a different approach by adopting
a jet-by-jet analysis to identify quenched jets directly from the jet substructure with
supervised machine learning architectures.
For this purpose, we employ JEWEL with both the default configuration, which employs
Glauber initial conditions coupled to Bjorken hydrodynamic model, and a more realistic
medium description, combining T$
_
\mathrm{R}$ENTo initial conditions with the
(2+1)D hydrodynamic evolution of v-USPhydro.
Each jet is represented as a sequential trajectory along the Cambridge/Aachen
declustering tree, characterized by four substructure observables ($z$, $\Delta R$,
$k_\perp$, $m_{inv}$) at each splitting, which we analyze using two supervised machine
learning architectures: Long Short-Term Memory (LSTM) networks and Transformers.
Our results show that the machine learning models chosen in this work are excellent at
the binary classification problem of distinguishing quenched jets on a jet-by-jet basis,
with accuracies and AUC scores greater than 95\%. In addition, the trained models are
sensitive to medium properties that are hidden in traditional global observables,
demonstrating the potential of ML for probing the Quark-Gluon Plasma.
| Is this an experimental talk? | No |
|---|---|
| Is this on behalf of a collaboration? | No |
| Are you willing to present as a poster if it is not selected for oral presentation? | Yes |