Jet Quenching Identification via Supervised Learning in Simulated Heavy-Ion Collisions

This paper demonstrates that sequential machine learning architectures applied to jet declustering history trees outperform traditional static models and global observables in identifying jet quenching, revealing sensitivity to medium-specific features and offering a powerful new tool for analyzing parton-medium interactions in heavy-ion collisions.

The Gist

Imagine you are a detective trying to figure out if a car has been driving through a thick, sticky mud pit or just cruising on a clean highway.

In the world of particle physics, heavy-ion collisions (smashing two heavy atoms together) create a super-hot, super-dense soup of particles called Quark-Gluon Plasma (QGP). When a high-energy particle (a "jet") flies through this soup, it gets slowed down and battered, losing energy. This is called Jet Quenching.

The problem? Traditional tools for detecting this "mud" are like looking at a blurry photo of the car's final speed. They tell you the car was slower, but they can't tell you how the mud changed the car's engine, suspension, or tires along the way.

This paper is about using Artificial Intelligence (AI) to look at the car's entire journey, step-by-step, to see exactly how the mud affected it.

The Detective's Toolkit: Two Types of AI

The researchers tested two different ways of using AI to spot the "mud":

1. The "Snapshot" Detectives (Static Models):

   • How they work: These AI models (like Random Forests and MLPs) look at just one single moment of the jet's life—specifically, the very first time it splits into two pieces. It's like taking a single photo of the car after it exits the mud and trying to guess how it got there.
   • The Result: They are okay at guessing, but they get confused easily. If you train them on a specific type of mud (a simplified simulation) and then show them a different, more realistic type of mud, they get lost. They rely on "global" clues that aren't unique enough.

2. The "Storyteller" Detectives (Sequential Models):

   • How they work: These AI models (like LSTMs and Transformers) look at the entire history of the jet. They watch the jet split, then split again, then split again, step-by-step, like reading a story from beginning to end. They see the sequence of events.
   • The Result: These detectives are incredibly sharp. They achieve over 95% accuracy. By watching the whole "story" of how the jet breaks apart, they can spot subtle patterns that the snapshot detectives miss.

The "Mud" Experiment: Two Simulations

To test their AI, the researchers used two different computer simulations of the "mud" (the QGP):

• The "Simple Mud" (Jewel Default): A basic, simplified version of the plasma. It's like a cartoon version of mud.
• The "Realistic Mud" (v-USPhydro): A complex, physics-heavy simulation that includes viscosity, pressure, and fluctuations. It's like real, thick, sticky mud with unpredictable currents.

The Big Discovery: The "Transfer" Test

The most interesting part of the paper is what happened when they swapped the training data. This is called Cross-Domain Validation.

• Scenario A: Training on Real Mud, Testing on Simple Mud.

  • They taught the AI on the complex, realistic mud. Then they tested it on the simple mud.
  • Result: The AI was still a genius! It recognized the patterns from the complex mud and applied them to the simple one. It understood the universal rules of how mud affects a car.

• Scenario B: Training on Simple Mud, Testing on Real Mud.

  • They taught the AI only on the cartoon mud. Then they tested it on the complex, realistic mud.
  • Result: The AI failed miserably. It got confused because the simple mud didn't teach it about the complex, subtle ways the real mud behaves. It missed the "quenched" jets.

The Lesson: You can learn the general rules of physics from a complex, realistic simulation and apply them to simpler cases. But if you only learn from a simplified model, you won't be ready for the real, messy world.

The "Why" Behind the Success

The researchers used a special tool called SHAP (which is like an AI magnifying glass) to see what the AI was looking at.

• For the "Snapshot" detectives: They mostly looked at the size of the jet and the angle of the first split.
• For the "Storyteller" detectives: They realized that the first few steps of the jet's splitting history are the most important. The AI learned that the very first moments of the jet's journey through the mud hold the biggest clues about what happened.

The Catch (Limitations)

The authors are honest about the limitations. Their "mud" simulation is still a bit too clean. In a real experiment, there is a lot of "background noise" (like other cars driving through the mud at the same time) that makes it harder to see the specific jet. Their AI is currently trained on a "clean" version of reality, so it might be too good at its job compared to what it would do in a real, messy lab.

The Bottom Line

This paper shows that Machine Learning is a powerful new microscope for particle physics.

• Old way: Look at the final result (the car's speed).
• New way: Watch the whole story of how the jet changed as it flew through the plasma.

By using AI to read the "story" of the jet's evolution, scientists can distinguish between different types of plasma with incredible precision. This helps us understand the fundamental nature of the universe's building blocks, proving that sometimes, to understand the big picture, you have to look at the small details in the right order.

Technical Summary

1. Problem Statement

Jet quenching—the modification of high-energy parton showers as they traverse the Quark-Gluon Plasma (QGP)—is a primary probe for understanding the properties of the QGP. Traditionally, this phenomenon is studied using global observables like the nuclear modification factor ($R_{AA}$), which averages over large ensembles of jets. However, these global metrics capture limited information about the complex, event-by-event dynamics of parton-medium interactions.

The authors address the challenge of identifying quenched jets on a jet-by-jet basis. They investigate whether Machine Learning (ML) can distinguish between jets that have interacted with a medium and those that have not (vacuum jets), and whether the specific implementation of the medium model affects the ML model's ability to generalize.

2. Methodology

Simulation Framework

The study utilizes the Jewel Monte Carlo event generator to simulate jet evolution in heavy-ion collisions. Two distinct medium descriptions are employed to test model sensitivity:

1. Jewel Default: A simplified medium consisting of a static cloud of thermal partons with longitudinal expansion (Bjorken expansion). It uses perturbative QCD for energy loss (elastic scattering and BDMPS-Z radiation).
2. Jewel + v-USPhydro: A more realistic implementation where Jewel is coupled with v-USPhydro, a viscous relativistic hydrodynamic model using Smoothed Particle Hydrodynamics (SPH). This includes event-by-event fluctuations, anisotropic expansion, and viscous effects, initialized by the TRENTo model.

Data Representation

Jets are reconstructed using the anti-$k_T$ algorithm ($R=0.4$) and reclustering via the Cambridge/Aachen algorithm. The Soft Drop grooming procedure is applied to remove soft contamination. The authors define two types of input features for ML:

• Static (Non-Sequential): Features extracted only from the first splitting satisfying the Soft Drop condition (a single snapshot of the jet substructure). Variables include momentum fraction ($z_g$), angular distance ($\Delta R$), perpendicular momentum ($k_\perp$), and invariant mass ($m_{inv}$).
• Sequential: The entire history of the declustering process is treated as a time series. Each step $t$ provides a vector $x_t = [z, \Delta R, k_\perp, m_{inv}]$, forming a sequence $x = [x_1, ..., x_N]$.

Machine Learning Models

Five supervised learning architectures are compared:

1. Random Forest (RF): An ensemble tree-based model (baseline for static data).
2. Multilayer Perceptron (MLP): A feed-forward neural network (baseline for static data).
3. Long Short-Term Memory (LSTM): A recurrent neural network (RNN) designed for sequential data.
4. LSTM + Attention: An LSTM augmented with an attention mechanism to weigh the importance of different time steps.
5. Transformer: A deep learning architecture using self-attention to process the entire sequence in parallel.

Evaluation Strategy

• In-Domain Validation: Models trained and tested on the same medium (e.g., Default $\to$ Default).
• Cross-Domain Validation: Models trained on one medium and tested on the other (e.g., Default $\to$ v-USPhydro and vice versa) to test generalization and physical robustness.
• Metrics: Accuracy, AUC, Precision, Recall, and F1-Score. Uncertainties are estimated via non-parametric bootstrap.
• Interpretability: SHAP (SHapley Additive exPlanations) values are used to determine feature importance and sequence attribution.

3. Key Contributions

• Architectural Hierarchy: The paper establishes a clear hierarchy where sequential models (LSTM, Transformer) significantly outperform static models (RF, MLP) in jet quenching classification. This demonstrates that treating jet modification as a temporal process is more effective than analyzing static snapshots.
• Medium Sensitivity: The study reveals that ML models are highly sensitive to the specific implementation of the medium. While traditional observables ($R_{AA}$) might yield similar results for different medium models, ML models can distinguish the underlying physics, leading to significant performance drops in cross-domain validation.
• Asymmetric Generalization: A critical finding is the asymmetry in cross-domain performance. Models trained on the realistic v-USPhydro medium generalize well to the simplified Default medium (maintaining high accuracy). Conversely, models trained on Default fail to generalize to v-USPhydro, suffering significant performance degradation. This suggests that the realistic medium encodes a richer set of physical patterns that encompass the simpler case, whereas the simplified model lacks the complexity to capture the full phenomenology.
• Sequence Importance: SHAP analysis of sequential models reveals that the earliest steps of the Soft Drop declustering sequence (typically steps $t=2$ to $5$) carry the most discriminative information, rather than the entire history.

4. Results

Static Models (RF and MLP)

• Performance: Moderate. AUC scores range from 0.80 to 0.88.
• Limitations: High false positive rates (Precision $\approx$ 0.5–0.6) and poor cross-domain generalization.
• Feature Importance: SHAP analysis identifies groomed angular separation ($R_g$) and groomed mass ($m_g$) as the most critical features. Higher $R_g$ correlates with quenching, while higher $m_g$ correlates with unquenched jets.

Sequential Models (LSTM, LSTM+Att, Transformer)

• Performance: Exceptional. In-domain accuracy exceeds 95%, with AUC scores often >0.98.
• Cross-Domain Results:
  • v-USPhydro $\to$ Default: High performance maintained (Accuracy $\approx$ 96%, AUC $\approx$ 0.99).
  • Default $\to$ v-USPhydro: Significant performance drop (Accuracy drops to $\approx$ 90-92%, Recall drops significantly).
• Observation: The addition of an Attention mechanism or the use of Transformers did not significantly improve performance over standard LSTMs, likely because the jet substructure sequences are relatively short and lack long-range dependencies that require complex attention mechanisms.

Interpretability

• Static: Confirms that medium-induced broadening (larger angular separation) is a key signature.
• Sequential: The model relies heavily on the first few hard branchings. In the v-USPhydro case, the model's sensitivity shifts slightly to later steps ($t=4, 5$) compared to the Default case, indicating that the realistic medium imprints subtle, delayed modifications that the model learns to utilize.

5. Significance and Limitations

Significance:

• The work validates sequential machine learning as a superior tool for jet quenching analysis compared to traditional global observables or static feature analysis.
• It highlights that realistic hydrodynamic modeling is crucial for training robust ML models. Training on oversimplified models may lead to classifiers that fail in realistic experimental conditions.
• It provides a framework for "jet-by-jet" analysis, potentially uncovering subtle medium effects hidden in ensemble averages.

Limitations:

• Lack of Thermal Background: The simulations do not include a fully simulated thermal heavy-ion background (soft particles). In real experiments, background fluctuations and subtraction ambiguities degrade jet substructure observables. The current results represent an idealized "upper bound" on performance.
• Selection Bias: The study acknowledges that jets with a specific reconstructed $p_T$ may originate from different physical histories (hard partons with high energy loss vs. moderate partons with low loss), complicating the binary classification task.
• Model Specificity: The results are specific to the Jewel framework and the specific medium implementations used; generalization to other MC generators or experimental data requires further study.

Conclusion:
The paper concludes that while static observables provide a baseline, sequential deep learning architectures offer a powerful, physics-aware approach to jet quenching identification. However, the success of these models is contingent upon training with realistic medium descriptions that capture the full complexity of the QGP, as evidenced by the failure of models trained on simplified scenarios to generalize to realistic ones.