Stable and Interpretable Jet Physics with IRC-Safe Equivariant Feature Extraction

This paper demonstrates that incorporating infrared and collinear safety along with E(2) and O(2) equivariance into graph neural networks for quark-gluon discrimination enhances model stability and interpretability by directly aligning learned latent representations with established QCD observables like Energy Flow Polynomials.

The Gist

Imagine you are a detective trying to solve a crime. The crime scene is a high-energy particle collision, and the "suspects" are tiny particles called quarks and gluons. When these particles fly out of the collision, they don't travel alone; they leave behind a messy trail of other particles, forming a spray called a jet.

Your job is to look at the shape and texture of these jets and figure out: "Was this made by a quark or a gluon?"

For a long time, scientists used complex math formulas (hand-crafted observables) to solve this. But recently, they started using Artificial Intelligence (AI), specifically a type called Deep Learning, which is like a super-smart detective that can find patterns humans miss. The problem? These AI detectives are "black boxes." They get the answer right, but they don't tell you how they got there. They might be cheating by looking at tiny, irrelevant details (like a speck of dust on the camera lens) rather than the actual crime scene.

This paper is about teaching these AI detectives to be honest, stable, and explainable by giving them a set of "rules of the road" based on the laws of physics.

The Problem: The "Black Box" Detective

Imagine you have two detectives:

1. Detective Random: This detective looks at the whole crime scene. If a tiny, harmless speck of dust (a "soft particle") lands on the scene, Detective Random might get confused and change their mind about who the culprit is. They are unstable.
2. Detective Physics: This detective follows strict rules. They know that if a tiny speck of dust lands on the scene, it shouldn't change the identity of the criminal. They are stable.

The paper asks: Can we build an AI that is just as good at solving the case as the "Random" one, but as stable and honest as the "Physics" one?

The Solution: Teaching the AI "Physics Rules"

The authors built special AI detectives using Graph Neural Networks. Think of a jet not as a cloud of smoke, but as a social network of particles. Each particle is a person, and they are connected to their neighbors.

They taught these networks two main "inductive biases" (which is just a fancy way of saying "strong rules"):

1. IRC Safety (The "Ignore the Dust" Rule):

   • The Concept: In physics, adding a tiny, low-energy particle (soft) or splitting a particle into two that fly in the exact same direction (collinear) shouldn't change the identity of the jet.
   • The Analogy: Imagine you are identifying a person by their face. If a fly lands on their nose, or if they split their hair perfectly down the middle, it shouldn't change who they are.
   • The AI: The authors designed the AI so that if a tiny particle is added or a particle splits, the AI's answer doesn't change. This prevents the AI from "cheating" by focusing on tiny, unstable details.

2. Equivariance (The "Rotate and Slide" Rule):

   • The Concept: The laws of physics look the same whether you rotate your head or move to a different spot in the room.
   • The Analogy: Imagine looking at a snowflake. If you rotate the snowflake, it's still the same snowflake. If you slide it across the table, it's still the same snowflake.
   • The AI: The authors built the AI so that if the whole jet rotates or shifts slightly, the AI's internal understanding shifts in a predictable way, but the final conclusion (Quark vs. Gluon) stays the same.

The Experiment: Putting the Detectives to the Test

The team trained four types of detectives:

• The "Physics" Detectives: Followed the IRC Safety and Rotation rules.
• The "Random" Detective: No special rules, just raw data.

They tested them on a massive dataset of simulated particle collisions.

The Results:

1. Performance: Surprisingly, the "Physics" detectives were just as good at solving the case as the "Random" one. They got the same score (about 90% accuracy).

   • Takeaway: You don't have to sacrifice accuracy to be honest.

2. Stability (The "Dust" Test):

   • They added a tiny bit of "noise" (a soft particle) to the jets.
   • The "Random" detective got confused. Their answers jumped around wildly depending on where the dust landed.
   • The "Physics" detectives remained calm. Their answers stayed steady.
   • Takeaway: The Physics detectives are more reliable in the real world, where noise is inevitable.

3. Interpretability (The "Why" Test):

   • This is the coolest part. The authors looked inside the "brain" of the AI to see what it was thinking.
   • They found that the "Physics" detectives were thinking in terms of Energy Flow Polynomials (EFPs). These are known, standard physics formulas that scientists have used for decades.
   • The "Random" detective was thinking in a jumbled mess of patterns that no human could easily understand.
   • Takeaway: By forcing the AI to follow physics rules, we can actually read its mind and see that it's using real, understandable physics concepts.

The Big Picture

This paper proves that you can build AI for particle physics that is:

• Smart: It solves the problem as well as any other AI.
• Robust: It doesn't get confused by tiny errors or noise.
• Transparent: We can look inside and say, "Ah, I see! It's using this specific physics rule to make that decision."

It's like upgrading from a detective who guesses based on a hunch to a detective who follows a strict, logical codebook that everyone can read and trust. This makes the AI a much better tool for scientists trying to understand the fundamental building blocks of our universe.

Technical Summary

1. Problem Statement

Deep learning has revolutionized jet physics, particularly in quark-gluon discrimination, achieving high performance by extracting complex features from low-level particle data. However, a critical bottleneck remains: interpretability. Standard deep learning models often function as "black boxes," learning features that may rely on detector artifacts or non-physical correlations rather than fundamental Quantum Chromodynamics (QCD) principles.

The paper addresses two specific challenges:

1. Lack of Physical Grounding: Learned representations are not guaranteed to respect fundamental symmetries of the strong force, specifically Infrared and Collinear (IRC) safety (robustness against soft radiation and collinear splitting) and geometric symmetries in the rapidity-azimuth plane.
2. Instability: Models lacking these constraints may produce unstable predictions when jets are subjected to soft emissions or recoil effects, limiting their reliability in experimental settings.

The authors aim to bridge the gap between high-performance machine learning and established analytic QCD observables by embedding strong physics-motivated inductive biases directly into Graph Neural Network (GNN) architectures.

2. Methodology

The authors propose a systematic comparison of four GNN architectures designed for quark-gluon discrimination, varying by the inductive biases they enforce:

A. Inductive Biases

1. IRC Safety: Ensures the network output is invariant to the addition of arbitrarily soft particles or collinear splittings. This is achieved via energy-weighted message passing and fixed-radius graph construction in the $(y, \phi)$ plane, ensuring the graph topology remains stable in soft/collinear limits.
2. Equivariance:
   • $E(2)$ Equivariance: Enforces invariance under both rotations and translations in the rapidity-azimuth plane (Euclidean group).
   • $O(2)$ Equivariance: Enforces invariance under rotations only (Orthogonal group).
   • These are implemented using specific message-passing updates where vector features transform covariantly and scalar features remain invariant.

B. Network Architectures Compared

The study compares four specific models trained on the same dataset:

1. $E(2)$-EMPN: IRC-safe and $E(2)$-equivariant.
2. $O(2)$-EMPN: IRC-safe and $O(2)$-equivariant.
3. EMPN: IRC-safe but non-equivariant (standard energy-weighted message passing).
4. MPNN: IRC-unsafe and non-equivariant (baseline model using raw particle features without symmetry constraints).

C. Dataset and Training

• Data: 2 million simulated jets ($p_T \in [500, 550]$ GeV) from Pythia 8, split into quark and gluon samples.
• Graph Construction: Jets are represented as graphs where nodes are massless constituents. Edges connect particles within a fixed radius $\Delta R < 0.5$.
• Training: All models use 3 message-passing layers, followed by a graph readout and an MLP classifier. Training is repeated 10 times with random initialization to assess stability.

D. Analysis Techniques

• Principal Component Analysis (PCA): Applied to the latent graph-level embeddings to understand how information is distributed across dimensions.
• Energy Flow Polynomial (EFP) Regression: The leading Principal Components (PCs) are regressed against a basis of EFPs (a known over-complete basis for IRC-safe observables) to quantify the correspondence between learned features and analytic physics.
• Perturbation Tests: Models are tested on modified jets with additional soft particles (inside the cone) or recoil effects (outside the cone) to measure robustness.

3. Key Contributions

1. Architecture Design: The development of message-passing GNNs that strictly enforce IRC safety and $E(2)/O(2)$ equivariance, ensuring the learned features respect fundamental QCD properties.
2. Demonstration of Stability: Showing that while performance (AUC) is comparable across all models, physics-constrained models are significantly more stable under soft emission perturbations.
3. Interpretability via EFPs: Establishing a direct mathematical link between the latent space of the neural networks and known analytic observables. The constrained models' latent directions are shown to be linear combinations of EFPs, whereas unconstrained models are not.
4. Latent Space Structure: Revealing that equivariant models concentrate variance into fewer, more interpretable principal components aligned with physics, whereas unconstrained models distribute variance more broadly into less interpretable directions.

4. Results

Classification Performance

• All four models achieved competitive performance with Mean AUC $\approx 0.89 - 0.90$.
• Key Insight: Enforcing IRC safety and equivariance does not degrade classification accuracy. This confirms that the optimal quark-gluon likelihood ratio is an IRC-safe observable, and physical constraints do not prevent the network from finding the optimal decision boundary.

Latent Space Analysis (PCA)

• Variance Concentration: Equivariant models ($E(2)$ and $O(2)$) concentrated $\sim 77\%$ of the variance in the first 3 Principal Components (PCs), whereas non-equivariant models required more components to explain the same variance ($\sim 71\%$).
• Interpretability: The leading PCs of the constrained models showed a very high coefficient of determination ($R^2 > 0.9$) when regressed against EFPs.
  • $E(2)$-EMPN PC1: $R^2 = 0.97$.
  • MPNN (Unconstrained) PC1: $R^2 = 0.92$, but PC2 and PC3 dropped significantly ($R^2 \approx 0.65 - 0.73$), indicating that subleading features in the unconstrained model are not well-described by IRC-safe observables.

Robustness to Soft Emissions

• Perturbation Tests: When jets were modified with soft particles (recoil or soft addition), the IRC-unsafe MPNN showed a sharp decline in performance (AUC dropping to $\sim 0.5$) and high variability across training runs.
• Stability Hierarchy: The stability order was $E(2)\text{-EMPN} > O(2)\text{-EMPN} > \text{EMPN} > \text{MPNN}$.
• The $E(2)$-equivariant model remained essentially unchanged even with significant soft perturbations, demonstrating that enforcing translational and rotational symmetry makes the model robust to kinematic shifts that are physically irrelevant.

5. Significance

This work provides a principled framework for interpretable deep learning in collider physics.

• Trustworthiness: By embedding physical symmetries, the models are guaranteed to respect the fundamental properties of QCD, making their predictions more trustworthy for experimental analysis.
• Bridging Theory and ML: The study successfully maps the "black box" latent space of GNNs to the "white box" world of analytic observables (EFPs). This allows physicists to understand what the network is learning (e.g., long-chain angular correlations) rather than just that it is learning.
• Robustness: The results suggest that for collider physics applications, where data is subject to detector effects and soft radiation, symmetry-aware and safety-aware architectures are superior to unconstrained baselines, offering better generalization and stability without sacrificing performance.

In conclusion, the paper argues that incorporating strong inductive biases is not merely a theoretical preference but a practical necessity for building robust, reliable, and interpretable machine learning tools in high-energy physics.