⚛️ phenomenology

Dissecting flavor tagging with graph neural networks

This paper demonstrates that Graph Neural Networks like ParticleNet enhance quark/gluon jet tagging by jointly leveraging constituent multiplicity and substructure, allowing them to correctly identify a significant fraction of jets where multiplicity alone suggests the opposite class.

Original authors: Ankita Budhraja, Sunny Kumar Singh, Balbeer Singh

Published 2026-10-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ankita Budhraja, Sunny Kumar Singh, Balbeer Singh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the chaotic aftermath of a high-energy collision, such as those created at the Large Hadron Collider, energy does not simply vanish; it condenses into tight, focused sprays of particles called jets. These jets are born when fundamental particles like quarks or gluons, which are the building blocks of matter, are knocked loose and fly apart at nearly the speed of light. As they travel, they shed energy by creating more particles, a process that continues until the energy drops low enough for them to clump together into stable, colorless particles that our detectors can actually see. While quarks and gluons are the parents of these sprays, they are never seen directly. Instead, scientists must look at the children—the pattern, number, and arrangement of the particles in the jet—to guess who the parent was. This is a crucial task because many searches for new physics rely on spotting rare events hidden within a sea of ordinary background noise, and knowing whether a jet came from a quark or a gluon helps researchers separate the signal from the static.

For decades, physicists have used simple rules to make this distinction, such as counting how many particles are in a jet. Because gluons carry a stronger "color charge" than quarks, they tend to radiate more energy and create jets with more particles that are spread out more broadly. However, nature is rarely so simple, and these sprays often overlap in ways that make a simple count insufficient. In recent years, scientists have turned to machine learning, specifically a type of artificial intelligence called a graph neural network, to solve this problem. These networks treat a jet not as a list of numbers, but as a dynamic map where every particle is a point connected to its neighbors, allowing the computer to learn complex relationships between the particles that human-designed formulas might miss. A specific network known as ParticleNet has become a standard tool for this work, achieving impressive accuracy in telling quark jets from gluon jets.

A team of researchers recently set out to understand exactly how this powerful tool works. They wanted to know if the network was simply learning to count particles, or if it was truly understanding the deeper structure of the jet. To find out, they trained the network on millions of simulated jets from a process where a Z boson is produced alongside a jet, carefully matching the quark and gluon jets so that they had the same energy and direction. This ensured that the network could not rely on the overall energy of the jet to make its decision; it had to rely entirely on the internal arrangement of the particles. The researchers then tested the network across a wide range of energies, from 80 to 400 GeV, to see how its performance changed.

The study confirmed that the network is indeed highly skilled at its job, correctly identifying the type of jet about 80 to 90 percent of the time, with performance improving as the jet energy increases. However, the real discovery lay in dissecting why it was so good. The researchers found that the network's success is largely driven by the number of particles in the jet, a feature that has long been known to be the strongest single clue. The network learned to adjust its expectations based on the jet's energy, knowing that a high-energy jet naturally contains more particles than a low-energy one, regardless of whether it came from a quark or a gluon. This ability to combine the particle count with the jet's energy scale was a key part of its intelligence.

Yet, the network was not just a sophisticated counter. The researchers discovered that even when two jets had the exact same number of particles, the network could still tell them apart about 20 to 34 percent of the time. This meant the network was looking at something else: the shape of the spray. By artificially stretching and shrinking the jets in their simulations, the team showed that the network considers a broader, more spread-out jet to be more like a gluon, and a tighter, narrower jet to be more like a quark. This sensitivity to the width of the jet was most important when the particle count was ambiguous—when the number of particles was not enough to give a clear answer on its own. In these tricky cases, the network looked at how the particles were distributed in space to make the final call.

The researchers also found that the network's output could be understood as a sum of two parts: a prediction based on the particle count and a "residual" prediction based on the remaining details of the jet's structure. This decomposition revealed that the network was not just memorizing patterns but was effectively learning the probability of a jet being a gluon or a quark given its specific features. While the network is not perfect and cannot distinguish jets that are fundamentally identical in their properties, it successfully uses the subtle differences in how particles are arranged to improve upon traditional methods. The work provides a clear window into the "black box" of machine learning, showing that these advanced tools are not just finding hidden correlations but are learning the physical principles that govern how particles behave, offering a new way to interpret the complex data that comes from the world's most powerful particle colliders.

Technical Summary: Dissecting Flavor Tagging with Graph Neural Networks

Problem Statement
Identifying whether a jet originates from a quark or a gluon is a persistent challenge in collider physics, crucial for background suppression in new-physics searches and for studying jet quenching in heavy-ion collisions. While standard jet substructure observables (e.g., constituent multiplicity, momentum dispersion, angularities) exploit differences in Quantum Chromodynamics (QCD) emission patterns—specifically the larger color charge of gluons ( $C_A/C_F = 9/4$ )—they often fail to capture complex, high-dimensional correlations. Machine Learning (ML), particularly Graph Neural Networks (GNNs) like ParticleNet, has demonstrated superior performance by treating jets as dynamic graphs of constituents. However, the specific information content learned by these networks remains partially opaque. Specifically, it is unclear how much of the discriminating power stems from constituent multiplicity versus the remaining substructure, and how this balance varies with jet transverse momentum ( $p_T$ ) and multiplicity ( $N$ ).

Methodology
The authors trained a ParticleNet classifier on quark and gluon jets generated from $Z$ +jet production in Pythia 8.318 at $\sqrt{s} = 14$ TeV. The dataset spans a wide transverse momentum range ( $80 \le p_T^{\text{jet}} \le 400$ GeV). To ensure the network learns substructure rather than production kinematics, the quark and gluon samples were strictly matched in $p_T$ and rapidity ( $|y|$ ) across 512 kinematic cells, resulting in identical marginal distributions for the two classes.

The analysis proceeds in three stages:

Baseline Construction: The authors constructed a "Multiplicity Likelihood Ratio" ( $L_N$ ) derived directly from the training data, representing the optimal classifier based solely on constituent multiplicity $N$ and $p_T$ . They also trained Boosted Decision Trees (BDTs) on sets of standard observables (multiplicity, momentum dispersion $p_D^T$ , angularities $\lambda_\beta$ , and jet charge $Q_\kappa$ ) to establish a traditional ML baseline.
Network Performance Evaluation: The ParticleNet's Area Under the Curve (AUC) was evaluated against the BDT and $L_N$ baselines across different $p_T$ bins.
Probabilistic Decomposition: The network output $M$ (log-likelihood ratio) was decomposed into two components: the multiplicity term $L_N(N, p_T)$ and a residual term $\Delta = M - L_N$ . This allows the isolation of information learned beyond multiplicity.
Controlled Deformation: To identify the physical nature of the residual term $\Delta$ , the authors performed controlled deformations on the jet constituents. Specifically, they rescaled the angular coordinates ( $\Delta\eta, \Delta\phi$ ) of all constituents by a factor $e^\delta$ while keeping multiplicity, momenta, and particle types fixed. This isolates the network's response to the radial spread (width) of the jet.

Key Results

Performance Gain: ParticleNet achieves an AUC increasing from 0.84 to 0.90 across the $p_T$ range. This represents a gain of 7–8% over the multiplicity likelihood ratio ( $L_N$ ), which is approximately three times larger than the gain achieved by a BDT trained on six standard observables (which improves $L_N$ by only 2–3%).
Multiplicty Dominance: Constituent multiplicity remains the single most powerful discriminator. At fixed $p_T$ , the multiplicity likelihood ratio accounts for the majority of the separation.
Probabilistic Interpretation: The network output is consistent with a probabilistic interpretation. At fixed multiplicity, the mean gluon probability assigned by the network matches the observed true gluon fraction. However, there is a significant spread in probabilities for jets of the same multiplicity, indicating the network utilizes additional structural information.
Resolving Ambiguity: The residual term $\Delta$ allows the network to correctly identify 20–34% of jets whose multiplicity strongly favors the opposite class (e.g., a quark jet with a high, gluon-like multiplicity). This success rate increases with $p_T$ .
Radial Spread Sensitivity: The controlled deformation study reveals that the residual term $\Delta$ is primarily driven by the radial spread of the jet. The network considers broader jets more gluon-like. This sensitivity is most pronounced at low and intermediate multiplicities ( $N \approx 18\text{--}27$ ), where multiplicity alone is ambiguous. At high multiplicities, where the class is already well-determined by $N$ , the response to radial spread diminishes.

Significance and Claims
The paper claims to provide a granular dissection of how GNNs perform flavor tagging. It demonstrates that while ParticleNet significantly outperforms traditional observables and multiplicity-only classifiers, this gain is not uniform; it is most critical for jets where multiplicity is ambiguous. The authors establish that the network effectively learns to combine the jet scale (reconstructed from constituents) with substructure details, specifically the radial energy flow.

Crucially, the authors interpret the network's output as the sum of a multiplicity log-likelihood ratio and a residual log-likelihood ratio. This decomposition suggests that the network's ability to "overrule" the multiplicity signal (correctly identifying jets with "wrong" multiplicity) is limited to a specific fraction of the dataset (20–34%) and is driven by the jet's width. The paper concludes that this framework offers a way to understand how taggers respond to complex substructure, potentially applicable to future studies of quenched jets, without claiming to have solved the fundamental overlap between quark and gluon distributions.

More like this