Boosting Sensitivity to $HH\to b\bar{b} γγ$ with Graph Neural Networks and XGBoost

This paper demonstrates that a Graph Neural Network (GNN) outperforms an XGBoost classifier in enhancing the sensitivity of $HH \to b\bar{b}\gamma\gamma$ searches at 13.6 TeV, significantly improving the expected upper limits on the double Higgs production cross-section and the Higgs boson self-coupling ($\kappa_\lambda$) compared to current ATLAS results.

The Gist

Imagine you are a detective trying to solve a mystery in a massive, crowded stadium. The mystery is: "Did two Higgs bosons (the 'God Particles') collide and create a specific pattern of energy?"

The problem is that these "double Higgs" events are incredibly rare. It’s like trying to find two specific people wearing bright yellow hats in a stadium of 100,000 people, all wearing different colors and moving around constantly. Most of what you see is just "noise"—background interference that looks almost like your target but isn't.

This paper describes a new way to use "Super-Detectives" (Artificial Intelligence) to find these rare patterns more accurately.

1. The Two Detectives: The "Checklist" vs. The "Artist"

The researchers compared two different types of AI "detectives" to see which one was better at spotting the yellow hats.

• The XGBoost Detective (The Checklist Expert):
  Think of this detective as someone with a very long, very strict checklist. They look at individual facts: "Is the person tall? Are they moving fast? Is their hat a certain shade of yellow?" They are very good at checking boxes, but they treat every fact as a separate item on a list. They don't really care how the facts relate to each other; they just want to know if the boxes are checked.

• The GNN Detective (The Geometric Artist):
  This is the new, advanced detective (the Graph Neural Network). Instead of just a checklist, this detective looks at the entire scene like a map or a web. They don't just see "a yellow hat" and "a fast runner"; they see "a yellow hat moving in a specific formation relative to another yellow hat, creating a certain shape in the crowd." They understand the geometry—the way the particles are positioned in space and how they are "connected" to one another.

2. The Result: The Artist Wins

The researchers found that the "Artist" (the GNN) was much better at its job.

Because the GNN understands the "shape" of the event (the way the particles dance together), it could tell the difference between a real signal and the background noise much more effectively than the "Checklist" detective.

The "Scorecard" of the victory:

• Better Accuracy: The GNN was significantly better at separating the "signal" (the real Higgs event) from the "background" (the noise).
• 28% Boost: By using the GNN, the scientists improved their sensitivity by 28%. In the world of particle physics, that is a massive leap—it’s like going from looking through a blurry window to looking through a high-definition telescope.
• Tighter Constraints: They used this to better understand the "Higgs self-coupling"—which is basically a way of measuring how the Higgs boson interacts with itself. The GNN gave them a much clearer, more precise answer.

3. Why does this matter?

The Higgs boson is the particle that gives everything in the universe its mass. Understanding how two Higgs bosons interact is like understanding the "glue" of the universe. If we find even a tiny deviation from what we expect, it could reveal "New Physics"—undiscovered laws of nature or even new particles we never knew existed.

In short: By teaching computers to see the "geometry" of a collision rather than just a list of numbers, scientists have given themselves a much sharper pair of eyes to peer into the deepest secrets of the universe.

Technical Summary

Technical Summary: Boosting Sensitivity to $HH \to b\bar{b}\gamma\gamma$ with Graph Neural Networks and XGBoost

1. Problem Statement

The primary objective of the study is to enhance the sensitivity of searches for double Higgs boson production ($HH$) at the Large Hadron Collider (LHC). Specifically, the research focuses on the $HH \to b\bar{b}\gamma\gamma$ decay channel at $\sqrt{s} = 13.6$ TeV.

While this "golden channel" offers a clean signature and excellent mass resolution, it suffers from a very small branching ratio ($\approx 0.26\%$) and significant backgrounds from single Higgs production and continuum QCD processes. The fundamental physics goal is to constrain the trilinear Higgs self-coupling ($\lambda_3$), a parameter that governs the shape of the Higgs potential and is essential for understanding electroweak symmetry breaking and vacuum stability. Current experimental constraints on $\lambda_3$ (parameterized as $\kappa_\lambda$) remain subject to large uncertainties.

2. Methodology

The authors implement and compare two distinct Machine Learning (ML) architectures to improve signal-versus-background discrimination:

• Tree-Based Classifier (XGBoost): A traditional gradient-boosted decision tree approach. It uses a flat vector of 32 high-level kinematic variables (including $p_T, \eta, \phi$, and invariant masses of various systems like $\gamma\gamma, b\bar{b},$ and $HH$).
• Graph Neural Network (GNN): A geometrically-aware model that treats each collision event as a bi-directional graph.
  • Nodes: Represent reconstructed objects (photons, $b$-jets, VBF jets) and composite systems ($\gamma\gamma, b\bar{b}, HH$).
  • Edges: Encode physical and topological relationships, such as the angular separation ($\Delta R$) between particles.
  • Architecture: Utilizes Edge-Conditioned Convolution (NNConv) layers to allow the network to learn from the spatial and hierarchical structure of the event. It includes a "virtual node" to integrate global VBF-related information.

Simulation & Analysis Workflow:

• Data Generation: Monte Carlo simulations using Powheg-Box (for signal) and MadGraph5 aMC@NLO (for background), processed through Pythia 8 and Delphes (for detector emulation).
• Pre-processing: Events were geometrically rotated in the transverse plane to align the leading photon with the beam axis, exploiting the detector's cylindrical symmetry.
• Statistical Inference: A binned likelihood fit was performed using the Pyhf framework to derive the 95% Confidence Level (CL) upper limits on the signal strength ($\mu_{HH}$) and the $\kappa_\lambda$ modifier.

3. Key Contributions

• Geometric Representation: The paper demonstrates that mapping collider events to graphs—rather than flat feature vectors—allows the model to capture complex spatial correlations and decay chain topologies that XGBoost misses.
• Hybrid Feature Engineering: The GNN successfully integrates both low-level particle kinematics and high-level composite system information within a single topological framework.
• Robustness Testing: The study provides a systematic evaluation of how different background uncertainty levels (from 0% to 40%) affect the sensitivity of both models.

4. Results

• Classification Performance: The GNN significantly outperformed XGBoost. The GNN achieved an AUC (Area Under the ROC Curve) of 87.70%, compared to 73.87% for XGBoost, representing an 18.7% improvement.
• Sensitivity to Cross-Section: The GNN improved the expected 95% CL upper limit on the $HH$ production cross-section by 28% (achieving $\approx 2.9 \times \text{SM}$ compared to XGBoost's $\approx 4.0 \times \text{SM}$).
• Discovery Significance: The expected discovery significance increased from 0.56 (XGBoost) to 0.88 (GNN).
• $\kappa_\lambda$ Constraints: The GNN yielded tighter confidence intervals for the Higgs self-coupling modifier. For the 95% CL, the GNN constrained $\kappa_\lambda$ to $[-0.5, 5.0]$, whereas XGBoost was broader at $[-0.9, 5.4]$.
• Systematic Resilience: The GNN showed greater stability and higher performance than XGBoost when subjected to increased background systematic uncertainties.

5. Significance

This research highlights a paradigm shift in High-Energy Physics (HEP) analysis: moving from feature-engineered "flat" models to geometry-aware deep learning. By explicitly encoding the topology of particle interactions, GNNs can extract more information from the same dataset. The results are highly competitive with current ATLAS experiment limits, suggesting that adopting graph-based architectures could be a critical factor in the future discovery of the Higgs self-coupling and other Beyond the Standard Model (BSM) physics at the High-Luminosity LHC.