All is More: Energy Flow Networks for Jet Quenching

TL;DR

This paper tackles jet-by-jet discrimination to study jet quenching in heavy-ion collisions by separating vacuum-like pp jets from PbPb jets that include medium effects. It proposes energy-flow-network (EFN) based classifiers trained on jet constituents and enhances them with physics-motivated observables through observable-enhanced EFNs (oEFNs), plus compact moment EFNs (MEFNs) that capture latent-space moments. In realistic UE+MR conditions, oEFNs achieve state-of-the-art per-jet discrimination with ROC AUC about $0.83$, outperforming linear and nonlinear observable baselines, while MEFNs provide comparable performance with much smaller latent representations. The results demonstrate that combining constituent-level information with high-level observables yields robust, interpretable models for jet-quenching studies and suggest practical paths for experimental deployment and future explorations of centrality, $p_T$, and model dependence.

Abstract

Jet quenching, the modification of jets by the quark-gluon plasma in heavy-ion collisions, provides a sensitive probe of the properties of the medium. A jet-by-jet discrimination study between proton-proton and lead-lead jets using energy flow networks and simple baselines, explicitly retaining medium response and underlying event contamination is presented. As references, linear discriminants and neural networks have been trained on high-level observables such as $N$-subjettiness and energy flow polynomials, including an extended energy flow polynomial set, in order to quantify the achievable performance without constituent-level learning. Energy flow networks are then trained on jet constituents and extended to observable-enhanced energy flow networks that concatenate standardized $N$-subjettiness and/or energy flow polynomials to the energy flow network latent space. In the realistic scenario, including both underlying event contamination and medium response, observable-enhanced energy flow networks set state-of-the-art performance with receiver operating characteristic area under the curve of $\simeq 0.83$, improving markedly over linear and non-linear baselines and previous work with different architectures. Finally, results from moment energy flow networks, an energy flow network variant that attains comparable area under the curve with a substantially more compact and interpretable latent space are shown. These results establish energy-flow-network-based approaches (especially when enhanced with physics-motivated observables) as practical and robust tools for jet-quenching studies.

Figures (18)

• Figure 1: Distributions of 4 $N$-subjettiness variables and their correlations (pairwise scatter with marginal histograms); (top) without UE contamination; (bottom) with UE contamination.
• Figure 2: Distributions of 4 EFPs and their correlations (pairwise scatter with marginal histograms); (top) without UE contamination; (bottom) with UE contamination.
• Figure 3: ROC curves for linear models across input sets and UE configurations.
• Figure 4: Classifier output score distributions for linear models across input sets and UE configurations.
• Figure 5: ROC curves for NN classifiers across input sets and UE configurations.