Deep learning for jet modification in the presence of the QGP background

This paper demonstrates that while convolutional neural networks struggle to predict jet energy loss in the presence of a QGP background, dynamic graph convolutional neural networks applied to background-subtracted particle clouds maintain high accuracy by effectively exploiting the full jet structure under realistic experimental conditions.

The Gist

The Big Picture: The "Jet" in the "Soup"

Imagine you are trying to study a high-speed bullet (a Jet) flying through a dense, boiling fog (the Quark-Gluon Plasma or QGP).

In the world of particle physics, scientists smash heavy atoms together to create this "soup" of free-floating particles. When a high-energy particle (a jet) flies through this soup, it crashes into the particles in the fog, losing energy and changing shape. This is called Jet Quenching.

The Problem:
In a real experiment, the "fog" is incredibly thick and messy. It's like trying to take a clear photo of a bullet while standing in a blizzard. The snowflakes (background particles) get stuck on the camera lens and in the photo, making it hard to tell how much energy the bullet actually lost.

The Goal:
The researchers wanted to build a smart computer program (Artificial Intelligence) that could look at a single jet, figure out exactly how much energy it lost while passing through the soup, and do this for every single jet individually. This helps scientists understand the "soup" better without getting confused by the messy background.

---

The Two "Detectives": CNN vs. DGCNN

The paper tests two different types of AI "detectives" to solve this mystery.

1. The CNN (The "Polaroid Photographer")

• How it works: This AI takes the jet and turns it into a 2D picture (a grid of pixels), like a Polaroid photo. It looks at the brightness of the pixels to guess how much energy was lost.
• The Analogy: Imagine looking at a blurry photo of a firework in a storm.
  • In a clear sky (No background): The AI is amazing. It sees the firework clearly and guesses the energy loss perfectly.
  • In a blizzard (With background): The photo gets covered in snowflakes. The AI gets confused. Even if you try to wipe the snow off the photo (background subtraction), the image is still a bit blurry, and the AI's guesses start to get sloppy. It struggles because turning a 3D cloud of particles into a flat 2D picture loses some of the fine details.

2. The DGCNN (The "3D Sculptor")

• How it works: This AI doesn't use pictures. Instead, it treats the jet as a cloud of individual points (a point cloud), where every single particle is a distinct dot in 3D space. It builds a dynamic map connecting these dots, like a sculptor feeling the shape of a statue with their hands.
• The Analogy: Imagine holding the firework in your hands in the middle of a blizzard.
  • In a clear sky: It works great.
  • In a blizzard: Even when snowflakes are falling, this AI is smart enough to feel the difference between the "hot" firework particles and the "cold" snowflakes. It can ignore the snow and focus on the shape of the firework itself.
  • The Result: Even after the "snow" is removed, this 3D sculptor is much more accurate than the photographer. It keeps its cool and gives the right answer almost every time.

---

The Experiment: Cleaning the Mess

The researchers simulated a realistic scenario:

1. The Setup: They created 140,000 simulated jets.
2. The Mess: They threw in thousands of "background" particles to mimic the heavy-ion collision environment.
3. The Cleaning: They used a technique called Constituent Subtraction. Think of this as a sophisticated vacuum cleaner that tries to suck out the snowflakes from the photo without sucking up the firework.
   • Result: The vacuum worked well, but it accidentally removed a tiny bit of the firework too (over-subtraction).

The Verdict

• The Photographer (CNN): Did a good job when the sky was clear. But once the snow (background) arrived, its performance dropped. Even after cleaning the photo, it couldn't quite get back to its original perfection.
• The Sculptor (DGCNN): Was the clear winner. Even with the snow and the imperfect cleaning, it maintained high accuracy across the board.

Why Does This Matter?

In the past, scientists often had to look at average results from thousands of jets, which hides the details of individual events. This paper shows that by using the 3D Sculptor (DGCNN), we can finally look at one jet at a time and accurately measure its energy loss, even in the messiest, most realistic experimental conditions.

The Takeaway:
If you want to understand how a bullet slows down in a storm, don't just take a blurry photo of it. Instead, use a tool that can feel the shape of the bullet in 3D space, ignoring the snow around it. This new method allows physicists to probe the "soup" of the early universe with much higher precision.

Technical Summary

1. Problem Statement

In relativistic heavy-ion collisions, high-energy partons traverse the Quark-Gluon Plasma (QGP), losing energy via interactions known as jet quenching. A primary challenge in studying this phenomenon is selection bias: because the jet spectrum falls steeply with transverse momentum ($p_T$), jets that lose less energy are more likely to be selected in a fixed $p_T$ bin, skewing measurements.

To mitigate this, researchers aim to predict the fractional energy loss ($\chi = p_T^{\text{final}} / p_T^{\text{initial}}$) on a jet-by-jet basis. While machine learning (ML) has shown promise in predicting $\chi$ for "clean" jets (without background), real experimental conditions involve a massive, fluctuating thermal background of soft particles. This background distorts jet observables and complicates reconstruction. The core problem addressed is: Can deep learning models accurately predict per-jet energy loss in realistic heavy-ion environments where jets are embedded in a fluctuating QGP background, and which data representation (images vs. point clouds) is most robust?

2. Methodology

A. Simulation Framework

The authors constructed a realistic simulation pipeline using three main components:

1. Initial State (PYTHIA): Generated high-$p_T$ partons and their vacuum showers for proton-proton ($p+p$) collisions at $\sqrt{s_{NN}} = 5.02$ TeV. These serve as the baseline ($p_T^{\text{initial}}$).
2. Jet-Medium Interaction (LBT Model): The Linear Boltzmann Transport (LBT) model simulated the evolution of these partons through the QGP. It accounts for elastic and inelastic scatterings, recoil partons, and medium response (negative partons) to ensure energy-momentum conservation.
3. Background Modeling (Thermal Toy Model): To mimic experimental conditions, a thermal model generated soft $\pi^\pm$ particles following a Boltzmann distribution, creating a dense background environment typical of 0-10% centrality Pb+Pb collisions.
   • Jet Categories:
     • LBT-only: Jets from LBT without background (Idealized).
     • LBT + Background: Jets embedded in the thermal background.
     • LBT + Background Subtracted: Jets where background was removed using Constituent Subtraction.

B. Background Removal: Constituent Subtraction (CS)

To recover jet kinematics from the background, the authors applied Constituent Subtraction. Unlike global subtraction, CS operates at the particle level:

• It estimates the average background momentum density ($\rho$) using $k_T$ clusters.
• It pairs real particles with "ghost" particles representing the background.
• It iteratively subtracts momentum from real particles based on their distance to ghosts, preserving the hard jet core while removing soft contamination.

C. Machine Learning Architectures

Two distinct architectures were trained to predict $\chi$:

1. Convolutional Neural Networks (CNNs):
   • Input: "Jet Images" ($33 \times 33$ matrices) representing the total $p_T$ deposited per pixel in the $\eta-\phi$ plane.
   • Preprocessing: Images were aligned (hardest subjet at center, second at $-\pi/2$) and parity-flipped for consistency.
   • Architecture: Three convolutional layers followed by fully connected layers.
2. Dynamic Graph Convolutional Neural Networks (DGCNNs):
   • Input: "Particle Clouds" (unordered sets of particles with 4-momentum and coordinates).
   • Architecture: Based on ParticleNet. It uses EdgeConv modules to dynamically update neighborhood graphs (using $k$-nearest neighbors) between layers. This allows the network to learn local geometric relationships and correlations between particles directly, without discretization.

3. Key Contributions

• Realistic Background Integration: Unlike previous studies that used background-free jets, this work explicitly embeds jets in a fluctuating thermal background and applies Constituent Subtraction, bridging the gap between theoretical ML models and experimental reality.
• Comparative Analysis of Representations: The paper provides a rigorous comparison between image-based (CNN) and point-cloud-based (DGCNN) representations in the context of heavy-ion jet quenching.
• Quantification of Background Impact: It demonstrates how background contamination degrades image-based features and how different architectures recover (or fail to recover) predictive power after subtraction.

4. Results

Performance on Background-Free Jets

• CNNs: Achieved high accuracy in predicting $\chi$ for "LBT-only" jets. The jet images clearly showed correlations between energy loss and the redistribution of soft particles to large angles.
• DGCNNs: Also performed well, serving as a strong baseline.

Performance with Background (Embedded)

• CNNs: Performance degraded significantly. The thermal background particles clustered into the jet cone, obscuring the subtle structural changes caused by quenching. The correlation between $\chi$ and jet image pixels weakened, leading to large prediction errors, especially for highly quenched jets ($\chi < 0.7$).

Performance with Background Subtraction

• CNNs: After applying Constituent Subtraction, CNN performance improved but remained below the background-free baseline. The authors attribute this to "over-subtraction" of soft particles, which removes genuine jet substructure information along with the background.
• DGCNNs: Applied to background-subtracted particle clouds, the DGCNN maintained high accuracy across the entire $\chi$ range (0.4–1.0).
  • Crucially, the DGCNN trained on subtracted background jets outperformed the CNN trained on idealized, background-free jets.
  • The root-mean-square deviation (RMSE) for DGCNN was consistently lower than CNN across all $\chi$ values.

Why DGCNNs Excel

The paper argues that DGCNNs are superior because:

1. No Discretization: They preserve full per-particle kinematic information, whereas CNNs lose resolution by binning particles into pixels.
2. Geometric Fidelity: EdgeConv operations directly compute relationships between neighboring particles, capturing fine-grained substructure (e.g., multiplicity and angular distributions) that are strongly correlated with medium-induced modifications.
3. Robustness: They are less sensitive to the distortions introduced by background subtraction algorithms.

5. Significance and Future Directions

• Mitigating Selection Bias: This work validates that ML can effectively predict per-jet energy loss even in noisy experimental environments, offering a path to correct selection biases in jet quenching measurements.
• Paradigm Shift: The results suggest a shift from image-based representations to point-cloud/graph-based representations for heavy-ion jet physics, as the latter better utilizes the full jet structure.
• Future Work: The authors outline five key directions:
  1. Incorporating hadronization (currently using partons).
  2. Using more realistic background models (hydrodynamic simulations + underlying events).
  3. Developing better background subtraction techniques that preserve substructure.
  4. Testing generalizability across different jet quenching models.
  5. Exploring groomed-jet definitions for energy loss to reduce sensitivity to soft background.

In conclusion, the study demonstrates that Dynamic Graph Convolutional Neural Networks (DGCNNs) operating on particle clouds are the superior tool for extracting precise jet energy loss information in the complex, fluctuating environment of the QGP, outperforming traditional image-based CNNs even when background subtraction is applied.