Machine Learning Enhanced Detection of Higgs Chain Decays in Vector Boson Fusion

This paper demonstrates that advanced Deep Learning methodologies applied to low-level calorimeter data can achieve a statistical significance of approximately 4.5σ in detecting heavy CP-even Higgs boson chain decays ($h_2 \to h_1h_1 \to b\bar b b\bar b$) produced via Vector Boson Fusion within the Next-to-Minimal Supersymmetric Standard Model at the LHC with 300 fb$^{-1}$ of integrated luminosity.

The Gist

Imagine the Large Hadron Collider (LHC) as a giant, high-speed particle smasher. Scientists use it to smash protons together to see what tiny pieces fly out. Usually, they are looking for the "Higgs boson," a particle discovered in 2012 that gives other particles their mass. But now, they want to see if there are heavier versions of this particle hiding in the debris.

This paper is about a specific, tricky search for a heavy, invisible "ghost" particle (let's call it H2) that might be created in a very specific way and then immediately splits into two smaller, familiar Higgs particles (let's call them H1).

Here is the story of how they tried to find it, explained simply:

1. The Setup: The "VBF" Factory

Usually, when the LHC smashes particles, the Higgs is made by smashing two heavy "gluon" particles together. But in this study, the scientists are looking for a different factory: Vector Boson Fusion (VBF).

Think of VBF like two fast-moving cars (quarks) zooming past each other on a highway. They don't crash directly; instead, they exchange a "ticket" (a force carrier) that creates a heavy particle (H2) in the middle of the road. The two cars keep going, but they get pushed slightly apart, leaving two "scrap" jets flying forward and backward. This is the signature of the VBF factory.

2. The Mystery: The "Chain Reaction"

Once this heavy H2 is created, it doesn't stay around. It instantly decays (breaks apart) into two lighter Higgs particles (H1).

• The Problem: These H1 particles are moving incredibly fast because H2 was so heavy.
• The Result: Because they are moving so fast, the two tiny particles inside each H1 (which are "bottom quarks") get squished together so tightly that they look like a single, messy spray of debris, rather than two separate items. In physics terms, they form a "fat jet."

So, the scientists are looking for a very specific scene:

1. Two "scrap" jets flying far apart (forward/backward).
2. Two "fat jets" in the middle, each containing a hidden spray of four bottom quarks.

3. The Challenge: Finding a Needle in a Haystack

The problem is that the LHC produces billions of "normal" crashes every second. Most of these crashes produce random sprays of bottom quarks that look exactly like the signal the scientists want. It's like trying to find a specific, rare type of snowflake in a blizzard where 99% of the flakes look identical.

The scientists tried a traditional method first:

• They set up simple rules (like "the debris must weigh this much" or "the jets must be this far apart").
• Result: It was a disaster. They only found a tiny hint of the signal (about 1.7 times the noise). In science, you need a "5-sigma" (5 times the noise) to claim a discovery. They were way off.

4. The Solution: The "AI Detective"

Since simple rules didn't work, the team turned to Machine Learning, specifically a type of Deep Learning called Convolutional Neural Networks (CNNs).

Think of the energy deposits in the detector as a digital photograph (a "jet image").

• The Old Way: Measuring the total weight and size of the photo.
• The AI Way: The AI looks at the texture and pattern of the photo. It learns to recognize the unique "fingerprint" of the heavy H2 breaking apart, even if the total weight looks similar to the background noise.

They trained the AI on millions of simulated crashes. The AI learned to spot the subtle differences between a "fake" spray of quarks and the "real" heavy H2 decay.

5. The Twist: Changing the Camera Lens

The scientists also tried two different ways to group the particles into "jets" (the photos):

1. Fixed Lens: Using a standard, unchanging size for the camera frame.
2. Variable Lens: Using a camera that automatically zooms in or out depending on how fast the particles are moving.

The Result:

• The AI using the Fixed Lens improved the signal to about 2.8 times the noise. Better, but still not a discovery.
• The AI using the Variable Lens (which adapts to the speed of the particles) was the winner. It boosted the signal to 4.5 times the noise.

The Bottom Line

While they didn't quite reach the "5-sigma" threshold for a confirmed discovery in this specific simulation, they proved that Machine Learning is a game-changer.

• Without AI: The signal was invisible (1.7σ).
• With AI: The signal became loud and clear (4.5σ).

The paper concludes that if the real LHC data looks like their simulation, using these advanced AI tools to look at the "texture" of particle sprays could finally allow scientists to find these heavy, chain-decaying Higgs particles. It suggests that the "Variable Lens" approach is the best way to see through the noise of the universe.

Technical Summary

Technical Summary: Machine Learning Enhanced Detection of Higgs Chain Decays in Vector Boson Fusion

Problem Statement
The discovery of the 125 GeV Higgs boson confirmed the Standard Model (SM) but left open questions regarding Dark Matter, matter-antimatter asymmetry, and the hierarchy problem, necessitating physics Beyond the Standard Model (BSM). Supersymmetry (SUSY), particularly the Next-to-Minimal Supersymmetric Standard Model (NMSSM), offers a natural framework to address these issues by extending the Higgs sector. A specific phenomenological interest lies in the production of a heavy CP-even Higgs boson ($h_2$) via Vector Boson Fusion (VBF), which subsequently decays into a pair of SM-like Higgs bosons ($h_1$), each decaying into a pair of $b$-quarks ($h_2 \to h_1h_1 \to b\bar{b}b\bar{b}$).

While the gluon-gluon fusion (ggF) channel is typically dominant for Higgs production, specific regions of the NMSSM parameter space (such as Benchmark Point C from Ref. [14]) suppress the $h_2$ coupling to top quarks, making VBF the primary production mechanism. However, detecting this specific chain decay is challenging. The signal involves a heavy resonance ($m_{h_2} \approx 700$ GeV) decaying into boosted lighter Higgs bosons ($m_{h_1} \approx 125$ GeV), resulting in a final state with two forward/backward light quark jets and four central $b$-quarks. Conventional cut-and-count analyses struggle to distinguish this signal from the overwhelming QCD multijet background (specifically $gg \to b\bar{b}b\bar{b}$), yielding a statistical significance far below the discovery threshold.

Methodology
The authors perform a collider simulation study at $\sqrt{s} = 14$ TeV using the NMSSM Benchmark Point C ($m_{h_2} = 700$ GeV, $m_{h_1} = 126.2$ GeV). The analysis pipeline includes:

1. Event Generation and Simulation: Signal events ($qq \to qqh_2 \to qqh_1h_1 \to qqb\bar{b}b\bar{b}$) and background events ($gg \to b\bar{b}b\bar{b}$) are generated using MadGraph5_aMC@NLO and PYTHIA8, followed by fast detector simulation with Delphes (CMS card).
2. Object Reconstruction:
   • Small-R Jets: Anti-$k_T$ algorithm ($R=0.4$) identifies forward/backward jets and $b$-tagged subjets.
   • Large-R (Fat) Jets: To capture the boosted $h_1 \to b\bar{b}$ decays, large-R jets are reconstructed using two algorithms: a fixed-radius ($R=1.2$) and a variable-radius algorithm (scaling $R$ from 0.4 to 2.0 based on constituent momentum).
   • Selection Criteria: Events require at least two forward/backward jets satisfying VBF topology ($|\Delta\eta_{jj}| \ge 4.5$, $M_{jj} \ge 700$ GeV) and at least two large-R jets in the central region ($|\eta| < 2$, $p_T > 200$ GeV), each containing at least two $b$-tagged subjets.
3. Machine Learning Architecture:
   • Jet Images: The energy distribution of constituents within the large-R jets is converted into 2D images in the $(\eta, \phi)$ plane.
   • Deep Learning Model: A Multi-layer Convolutional Neural Network (CNN) architecture, inspired by GoogLeNet, processes these jet images. The model utilizes two parallel image branches (for leading and sub-leading large-R jets) to extract latent feature vectors.
   • Hybrid Approach: To further enhance discrimination, a separate branch processes high-level kinematic variables (e.g., invariant masses of jet pairs, energy fractions, angular separations, and subjet symmetry measures). The outputs of the image branches and the kinematic branch are concatenated for final classification.

Key Contributions

• VBF Sensitivity in NMSSM: The paper demonstrates that the $h_2 \to h_1h_1 \to 4b$ channel is accessible in the VBF production mode within the NMSSM, a scenario where the heavy Higgs coupling to top quarks is suppressed.
• Deep Learning Application: It introduces a Deep Learning strategy specifically tailored for this chain decay, utilizing jet images to capture the substructure of boosted di-Higgs decays.
• Algorithm Comparison: The study systematically compares fixed-R and variable-R jet clustering algorithms, showing that variable-R clustering better captures the physics of boosted objects with varying momentum scales.
• Kinematic Integration: It validates the efficacy of combining low-level calorimeter information (jet images) with high-level kinematic features to maximize signal significance.

Results
The analysis evaluates performance at an integrated luminosity of $300 \text{ fb}^{-1}$:

• Cut-and-Count: A conventional analysis using mass window cuts around $m_{h_1}$ and $m_{h_2}$ yields a statistical significance of only 1.7$\sigma$, insufficient for discovery.
• CNN on Jet Images (Fixed-R): Using only the Multi-layer CNN on jet images improves the significance to 2.8$\sigma$.
• Hybrid Model (Fixed-R): Combining jet images with kinematic features increases the significance to 4.2$\sigma$ (dropping to 3.7$\sigma$ with 10% systematic uncertainty).
• Hybrid Model (Variable-R): The optimal configuration employs the variable-R jet clustering algorithm combined with the hybrid CNN model. This achieves a statistical significance of 4.5$\sigma$ (maintaining 4.0$\sigma$ with 10% systematic uncertainty).

Significance and Claims
The authors claim that their findings represent a significant improvement over previous studies (specifically Ref. [14]) which analyzed the same NMSSM benchmark point but focused on the $h_2 \to ZZ \to 4\ell$ channel, finding less than 1$\sigma$ significance for that mode at Run 3.

The paper asserts that by leveraging advanced Deep Learning methodologies on low-level calorimeter data, the VBF production mode for heavy Higgs chain decays ($h_2 \to h_1h_1 \to 4b$) becomes a viable and competitive discovery channel at the LHC Run 3, potentially offering sensitivity comparable to or better than the ggF channel in specific BSM scenarios. The authors advocate for ATLAS and CMS to adopt these specific analysis strategies (variable-R clustering and hybrid CNNs) to probe Higgs self-couplings and BSM physics in the VBF regime.