Validating a Machine Learning Approach to Identify Quenched Jets in Heavy-Ion Collisions

This paper validates a Long Short-Term Memory (LSTM) neural network approach that successfully identifies jet quenching in heavy-ion collisions by leveraging jet substructure and parton shower history, demonstrating robust performance even when accounting for detector effects and generalizing to untrained observables.

The Gist

Imagine a high-energy physics experiment as a massive, chaotic mosh pit. In this pit, particles smash into each other at nearly the speed of light. Sometimes, this collision creates a super-hot, super-dense soup of energy called the Quark-Gluon Plasma (QGP). Think of the QGP like a thick, sticky honey that fills the entire room.

When a high-speed particle (a "jet") tries to fly through this honey, it doesn't just glide through; it gets slowed down, scattered, and loses energy. This process is called jet quenching. Physicists want to study this to understand how the "honey" behaves, but there's a problem: the mosh pit is so crowded and noisy that it's hard to tell which jets actually got slowed down by the honey and which ones just happened to look slow because of the crowd or the cameras filming the event.

Here is how the authors of this paper solved that puzzle, explained simply:

1. The Problem: Too Much Noise

In a real experiment, you have two main issues:

• The Background Noise: The "honey" itself is made of billions of other tiny particles. It's like trying to hear a single person speak in a stadium full of cheering fans.
• The Camera Blur: The detectors (cameras) aren't perfect. They sometimes blur the picture or miss details, making it hard to see exactly what happened.

Scientists need a way to look at a single jet and say, "Yes, this specific jet definitely got slowed down by the honey," rather than just guessing based on averages.

2. The Solution: A "Jet Detective" AI

The team built a special type of Artificial Intelligence (AI) called an LSTM (Long Short-Term Memory) network. You can think of this AI as a super-detective that looks at the "footprints" a jet leaves behind.

• How it learns: They didn't just show the AI pictures of jets. They showed it the entire history of how the jet was built, step-by-step, like watching a movie of a tree growing branch by branch.
• The Training: They fed the AI millions of simulated collisions. Some jets flew through empty space (vacuum), and others flew through the "honey" (QGP). The AI learned to spot the tiny, subtle differences in the "branching patterns" that only happen when a jet hits the honey.
• The Trick: They taught the AI to ignore the "stadium noise" (background particles) and the "camera blur" (detector errors) so it could focus purely on the physics of the jet slowing down.

3. The Test: Did the AI Get it Right?

To prove their AI wasn't just memorizing the wrong things, they gave it a series of tests it had never seen before.

• The "Photon Anchor": In their simulations, they used a special setup where a jet is paired with a photon (a particle of light). The photon is like a perfectly accurate ruler that doesn't get slowed down by the honey. By comparing the jet to the photon, they knew exactly how much energy the jet should have lost.
• The Result: The AI's predictions matched the "ruler" perfectly. If the AI said a jet was heavily quenched, the photon confirmed it had lost a lot of energy. If the AI said it was barely touched, the photon confirmed it was fine.

4. The "Blind" Checks

To make sure the AI wasn't just guessing, they asked it to predict other things it hadn't been trained on, like:

• The Shape of the Jet: Does the jet spread out more like a spray? (Yes, quenched jets spread out more).
• The Fragments: Does the jet break into more tiny, soft pieces? (Yes, quenched jets do this).
• The Momentum: Is the jet's push off-balance compared to the photon? (Yes, it is).

The AI correctly identified that the "heavily quenched" jets were the ones that were wider, softer, and more unbalanced. This proved the AI was actually learning the physics of the "honey," not just random noise.

5. The Real-World Test

Finally, they ran the AI through a simulation of a real detector (like the CMS detector at CERN) to see if it would still work with "blurry" real-world data.

• The Verdict: Even with the camera blur and the noisy background, the AI still successfully identified which jets were quenched and how much energy they lost.

Summary

The paper demonstrates that they have built a smart, specialized AI that can look at a single particle spray in a chaotic, noisy environment and accurately tell you: "This jet hit the hot plasma and lost energy," while ignoring the background noise and camera glitches. This gives scientists a powerful new tool to study the "honey" of the early universe, one jet at a time.

Technical Summary

1. Problem Statement

Jet quenching—the energy loss of high-energy partons traversing the Quark-Gluon Plasma (QGP)—is a primary probe for studying the properties of the QGP created in heavy-ion collisions. However, isolating true quenching effects is challenging due to several confounding factors:

• Complex Interplay: Multiple physics processes (e.g., jet flavor, path length, medium density fluctuations) affect jet observables simultaneously.
• Experimental Bias: Selection criteria often bias samples toward jets that lose less energy.
• Background and Detector Effects: Uncorrelated thermal background particles, detector resolution smearing, and differences in quark/gluon fractions between proton-proton ($pp$) and heavy-ion ($AA$) collisions can mimic quenching signatures.
• The Core Challenge: Existing machine learning (ML) classifiers often struggle to distinguish between true medium-induced quenching and artifacts caused by background fluctuations or detector effects. There is a need for a method that can identify quenching on a jet-by-jet basis while accounting for realistic detector responses.

2. Methodology

A. Data Generation and Simulation

• Event Generator: The study uses JEWEL v2.2.0 to generate photon-jet events at $\sqrt{s_{NN}} = 5.02$ TeV.
  • Vacuum Jets ($pp$): Simulated with the "vacuum" option.
  • Medium Jets ($PbPb$): Simulated with the "simple" option and "recoil" enabled to include parton-medium interactions.
• Background Modeling: Both vacuum and medium samples are embedded in a realistic heavy-ion background generated by Angantyr (within Pythia 8.3) for 0–10% central collisions. This models uncorrelated thermal fluctuations without explicitly simulating hydrodynamic medium response (which is handled by JEWEL).
• Detector Simulation: Delphes-3.5.0 is used to simulate the CMS detector response, including calorimeter energy smearing, tracking efficiency, and particle-flow reconstruction.
• Preprocessing:
  • Recoil Subtraction: A specific subtraction removes thermal momentum components from recoil partons to prevent unphysical features from entering the ML input.
  • Constituent Subtraction: Applied to remove soft background particles from jets.
  • Jet Reconstruction: Jets are reconstructed using the anti-$k_T$ algorithm ($R=0.4$) and reclustered using the Cambridge/Aachen (C/A) algorithm to extract substructure.

B. Machine Learning Architecture

• Model Type: A Long Short-Term Memory (LSTM) neural network, a type of Recurrent Neural Network (RNN) suited for sequential data.
• Input Features: The model processes the jet substructure derived from the parton shower history. Jets are declustered angularly, and each splitting step $t$ is represented by a feature vector $x_t$:
  • Momentum fraction ($z$)
  • Angular distance ($\Delta R$)
  • Perpendicular momentum ($k_\perp$)
  • Invariant mass ($m_{inv}$)
• Training Strategy:
  • Supervised Learning: Vacuum jets are labeled 0; medium jets are labeled 1.
  • Calibration: The photon energy in photon-jet events serves as an independent calibration for the jet energy loss, ensuring the model learns energy loss rather than fragmentation bias.
  • Loss Function: Weighted Mean Squared Error (MSE).
  • Hyperparameter Optimization: Performed using the Hyperopt package to optimize LSTM layers, FC dimensions, learning rates, and batch sizes.

3. Key Contributions

1. Validation of True Quenching Features: The study demonstrates that an LSTM trained on jet substructure and shower history can distinguish true QGP-induced quenching from background fluctuations and detector smearing.
2. Realistic Detector Integration: Unlike many simulation-only studies, this work incorporates full detector response (Delphes/CMS emulation) and uncorrelated thermal backgrounds, proving the method's robustness in a realistic experimental environment.
3. Cross-Validation with Unseen Observables: The authors validate the model not just by classification accuracy, but by applying the trained classifier to observables not used in training (photon-jet imbalance, fragmentation functions, and jet shapes).
4. Jet-by-Jet Classification: The approach moves beyond ensemble averages, providing a "quenching level" (0–100%) for individual jets, allowing for granular analysis of medium modifications.

4. Results

A. Training Performance

• The LSTM achieves an Area Under the Curve (AUC) of 0.777 at the Generator (GEN) level and 0.732 at the Reconstruction (RECO) level.
• The model successfully separates quenched jets into two distinct subsets based on output probability:
  • Top 40% (High Quenching): Shows significant substructure modifications (enhanced wide-angle, soft splittings).
  • Bottom 60% (Low Quenching): Exhibits patterns similar to vacuum jets.
• Consistency between GEN and RECO levels confirms the model learns physical quenching features rather than detector artifacts.

B. Cross-Validation with External Observables

The model's "quenching level" predictions were tested against three independent observables:

1. Photon-Jet Momentum Imbalance ($x_J$):

   • Jets classified as highly quenched (Q 0–20%) show the largest transverse momentum imbalance relative to the photon.
   • As the quenching level decreases (Q 80–100%), the imbalance shifts toward zero (vacuum-like).
   • This ordering holds true even with detector effects included.

2. Jet Fragmentation Function ($\xi = \ln(1/z)$):

   • Highly quenched jets show a significant enhancement in the large $\xi$ region (soft particles) and depletion in the intermediate region compared to vacuum jets.
   • This confirms the model correctly identifies the redistribution of energy to softer particles.

3. Jet Momentum Profile (Integrated Jet Shape):

   • Strongly quenched jets exhibit enhanced energy at larger angular distances ($\Delta R$) from the jet axis and depletion at smaller angles.
   • This pattern matches the physical expectation of energy broadening due to medium interactions.

5. Significance and Conclusion

This paper validates that sequential jet substructure analysis using LSTMs is a powerful tool for identifying jet quenching in heavy-ion collisions.

• Robustness: The method effectively disentangles true medium effects from thermal background and detector resolution issues.
• Experimental Applicability: By simulating the CMS detector environment, the study provides a roadmap for applying this ML approach to real experimental data.
• Physics Insight: The ability to assign a quenching level to individual jets allows physicists to study the distribution of energy loss mechanisms and disentangle competing theories of jet-medium interactions, moving beyond average suppression measurements to a more differential understanding of the QGP.

The authors conclude that this approach offers significant potential for future experimental analyses, enabling a jet-by-jet dissection of the QGP properties.