Jet momentum reconstruction in the QGP background with machine learning

TL;DR

This work tackles jet momentum reconstruction in heavy-ion collisions by leveraging a Dense Neural Network trained on jet samples that include medium-induced quenching (LBT) and a hybrid PYTHIA+LBT dataset. The study demonstrates that models trained only on vacuum jets fail to recognize medium-recoil components, leading to oversubtraction of the QGP background, an error that persists even after unfolding. Training on quenched jets or on a combined dataset significantly reduces bias and yields background-subtracted jet momenta that outperform traditional Area-based and Constituent Subtraction methods; unfolding then refines the jet $R_{AA}$ estimates toward the LBT baseline. These results underscore the importance of quenching-aware training data and suggest that physics-informed, multi-model training can enhance the reliability of jet observables in heavy-ion collisions, with potential extensions to include more observables and priors to further improve robustness and interpretability.

Abstract

We apply a Dense Neural Network (DNN) approach to reconstruct jet momentum within a quark-gluon plasma (QGP) background, using simulated data from PYTHIA and Linear Boltzmann Transport (LBT) Models for comparative analysis. We find that medium response particles from the LBT simulation, scattered out of the QGP background but belonging to medium-modified jets, lead to oversubtraction of the background if the DNN model is trained on vacuum jets from PYTHIA simulation. By training the DNN model on quenched jets generated using LBT or the combination of jet samples from PYTHIA and LBT, we significantly reduce this prediction bias and achieve more accurate background subtraction compared to conventional Area-based and Constituent Subtraction methods widely adopted in experimental measurements. We further study the performance of these machine learning models on evaluating the nuclear modification factor of jets, and find that while the unfolding procedure is necessary for correcting residuals in reconstructed jet momenta, models trained on samples incorporating quenched jets still achieve superior accuracy than those trained on vacuum jets even after unfolding.

Figures (8)

• Figure 1: (Color online) Residual distributions of the jet $p_\mathrm{T}$ predicted by the ML model trained on the PYTHIA data, compared between testing on datasets of PYTHIA, LBT, and LBT without recoil (and negative) particles, respectively.
• Figure 2: (Color online) Residual distributions of the jet $p_\mathrm{T}$ predicted by different estimators, ML models trained on the PTYHIA, LBT, and combined PYTHIA+LBT data, and the conventional Area-based and Constituent Subtraction methods. The testing dataset is generated by the LBT model.
• Figure 3: (Color online) Residual distributions of the jet $p_\mathrm{T}$ predicted by different ML models trained on the PTYHIA, LBT, and combined PYTHIA+LBT data, respectively. The testing dataset is generated by the PYTHIA model.
• Figure 4: (Color online) The nuclear modification factor of jets obtained from the ML model trained on the LBT dataset, compared between with and without the unfolding correction, and a baseline of the LBT jets reconstructed inside a QGP background (LBT jet within bkg).
• Figure 5: (Color online) The nuclear modification factor of jets obtained from ML models trained on the PYTHIA, LBT, and combined PYTHIA+LBT datasets with the unfolding correction, compared to a baseline of the LBT jets reconstructed in the presence of a QGP background (LBT jet within bkg).