Search for Higgs bosons produced in association with a high-energy photon via vector-boson fusion and decaying to a pair of $b$-quarks in the ATLAS detector

Using 133 fb$^{-1}$ of 13 TeV proton-proton collision data, the ATLAS collaboration performed a search for Standard Model Higgs bosons produced via vector-boson fusion in association with a high-energy photon and decaying to $b\bar{b}$, employing improved analysis techniques to measure a signal strength of $0.2 \pm 0.7$ with an observed significance of 0.3 standard deviations, consistent with the background-only hypothesis.

The Gist

The Big Picture: Hunting for a Ghost in a Storm

Imagine the Large Hadron Collider (LHC) is a massive, high-speed car race where particles are the cars. When they crash, they create a chaotic explosion of debris. Physicists are trying to find a very specific, rare "ghost" car in that explosion: the Higgs boson.

The Higgs boson is famous for giving other particles mass, but it's tricky to catch. It usually decays (falls apart) almost instantly into a pair of bottom quarks (let's call them "b-quarks"). The problem is that the race track is littered with millions of other "b-quark" debris from normal crashes. Finding the Higgs is like trying to spot a specific red marble in a pile of a million identical red marbles.

The New Strategy: The "Flashlight" Trick

In this new study, the ATLAS team decided to change their search strategy. Instead of just looking for the red marbles, they decided to look for a red marble that was hit by a bright flashlight at the exact moment of the crash.

• The Flashlight: This is a high-energy photon (a particle of light).
• The Trick: In the physics of these collisions, if a Higgs boson is created alongside a photon, it happens in a very specific way called Vector-Boson Fusion (VBF). This process is rare, but it has a superpower: it naturally suppresses the "noise" (the background debris).
• The Result: By demanding a photon be present, the team filters out 99% of the junk. It's like turning on a spotlight in a dark, crowded room; suddenly, the specific person you are looking for stands out much more clearly against the dark background.

The Detective Work: Upgrading the Tools

The team used data from 2015 to 2018 (133 "inverse femtobarns" of data, which is a fancy way of saying "a huge amount of collision records"). To find the signal, they had to upgrade their detective toolkit:

1. The Neural Network (The Super-Sleuth): In previous searches, they used a standard decision tree (like a flowchart) to guess which events were Higgs bosons. In this paper, they upgraded to a Neural Network (a type of AI). Think of the old method as a junior detective following a checklist, while the new Neural Network is a seasoned detective who can look at the whole picture, sense patterns, and spot subtle clues that the checklist would miss.
2. Better Background Modeling: They realized their computer simulations of the "junk" background weren't perfect. They developed a new method to "reweight" their simulations, essentially teaching the computer to mimic the real-world noise more accurately before they started looking for the signal.
3. Direct Fitting: Instead of just counting how many events fell into a "Higgs zone," they looked at the entire distribution of the AI's confidence scores. It's like not just counting how many people fit a description, but analyzing the probability that every single person in the crowd is the suspect.

The Results: A Quiet Room

After running all the data through their new, high-tech system, here is what they found:

• The Expectation: Based on the Standard Model (our best theory of physics), they expected to see a signal with a significance of 1.5 standard deviations. In detective terms, this means they expected a "strong hint" or a "likely suspect," but not enough to arrest anyone yet.
• The Reality: They observed a signal strength of 0.2 (relative to what was predicted). The statistical significance was only 0.3 standard deviations.
• The Translation: This is essentially a "null result." It's like the detective looking at the suspect list and saying, "I don't see anyone here who matches the description better than random chance." The data looks almost exactly like the background noise.

Why This Matters (Even if they didn't find it)

You might wonder, "If they didn't find it, why write a paper?"

1. Proving the Method Works: They successfully demonstrated that their new "Flashlight + AI" strategy works. They showed they can model the background noise incredibly well and that their new tools are more sensitive than the old ones.
2. Setting the Bar: They measured the "signal strength" to be 0.2 ± 0.7. This means the true value is likely somewhere between -0.5 and +0.9. Since the Standard Model predicts 1.0, their result is compatible with the theory (it's within the margin of error), but it doesn't prove the theory is right either.
3. Future Proofing: This analysis is a dress rehearsal. The techniques they perfected here—especially the neural network and the background modeling—are now ready to be applied to even more data in the future. They are sharpening their knives for the next hunt.

The Bottom Line

The ATLAS team took a massive dataset, used a clever "photon tag" to clean up the noise, and deployed a super-smart AI to look for the Higgs boson. They didn't find a definitive discovery this time (the signal was too faint to distinguish from random fluctuations), but they proved their new methods are powerful and ready for the next round of the race. They are still looking, and they are looking smarter than ever.

Technical Summary

Technical Summary: Search for Higgs bosons produced in association with a high-energy photon via vector-boson fusion and decaying to a pair of $b$-quarks in the ATLAS detector

Problem and Motivation
The search for the Standard Model (SM) Higgs boson decaying into a pair of $b$-quarks ($H \to b\bar{b}$) is challenging due to the overwhelming non-resonant Quantum Chromodynamics (QCD) multijet background. While the $H \to b\bar{b}$ decay is the dominant mode, isolating it requires significant background suppression. This analysis targets the associated production of a Higgs boson with a high-energy photon ($H\gamma$) in the vector-boson fusion (VBF) topology. The requirement of an associated photon serves two critical functions: it provides a clean trigger signature and, through destructive interference effects between initial-state and final-state radiation diagrams, significantly reduces the non-resonant QCD multijet background. Furthermore, in the presence of an associated photon, VBF becomes the dominant Higgs production mechanism, surpassing gluon-gluon fusion ($ggH$). This channel offers a unique opportunity to study the $HWW$ interaction and VBF production dynamics more closely, as $ZZ$-fusion is suppressed by similar interference effects.

Methodology
The analysis utilizes 133 fb$^{-1}$ of proton-proton collision data at $\sqrt{s} = 13$ TeV collected by the ATLAS detector during Run 2 (2015–2018).

• Event Selection: The target signature consists of two $b$-tagged jets (from the Higgs decay), two additional high-invariant-mass jets (VBF-candidate jets), and a high-energy isolated photon. Events are selected with a photon $p_T > 30$ GeV, at least two $b$-tagged jets with $p_T > 40$ GeV, and two additional jets with $p_T > 40$ GeV. The dijet invariant mass of the VBF candidates is required to exceed 800 GeV. To maintain orthogonality with other analyses, events containing isolated electrons or muons with $p_T > 25$ GeV are vetoed.
• Signal and Background Modeling: Signal events ($H \to b\bar{b} + \gamma$) are simulated at next-to-leading order (NLO) using MadGraph5_aMC@NLO and Herwig 7. The dominant backgrounds are non-resonant $b\bar{b}\gamma jj$ and $c\bar{c}\gamma jj$ production, modeled at leading order (LO) with MadGraph5_aMC@NLO and Pythia 8. Top-quark ($t\bar{t}$) and $Z\gamma jj$ backgrounds are also included.
• Multivariate Analysis (MVA): A key improvement over previous analyses is the replacement of the Boosted Decision Tree (BDT) with a densely connected Neural Network (NN). The NN is trained to distinguish signal-like events from background using 14 kinematic input variables, including the di-$b$-jet invariant mass ($m_{bb}$), jet multiplicity, and angular correlations. The NN output score is used as the primary discriminant.
• Background Modeling and Reweighting: To address discrepancies between data and Monte Carlo (MC) simulations in the control region (defined by $m_{bb}$ sidebands), kinematic reweighting scale factors are extracted for the dominant $b\bar{b}\gamma jj$ and $c\bar{c}\gamma jj$ backgrounds. These factors are derived from the ratio of data to MC distributions for the 14 NN input variables and applied to the MC samples to improve agreement.
• Statistical Analysis: A binned maximum-likelihood fit is performed directly on the NN output score distribution across the signal region ($100 \le m_{bb} \le 150$ GeV) and control region. The fit extracts the Higgs signal strength ($\mu_H$) and the normalization of the dominant non-resonant backgrounds simultaneously. Systematic uncertainties, including theory modeling, parton shower variations, and experimental effects (jet energy scale, $b$-tagging efficiency), are incorporated via 78 nuisance parameters.

Key Contributions and Improvements
This analysis introduces several methodological advancements compared to previous ATLAS searches in this channel:

1. Neural Network Classifier: The use of a deep neural network provides superior signal-background discrimination compared to the previously used BDT.
2. Direct Fitting Strategy: Unlike previous approaches that fitted the $m_{bb}$ distribution, this analysis extracts the signal strength directly from the NN score distribution, utilizing $m_{bb}$ only to define signal and control regions. This reduces modeling-related systematic uncertainties associated with the background shape parametrization.
3. Enhanced Background Modeling: The implementation of kinematic reweighting for the dominant non-resonant backgrounds significantly improves the agreement between data and simulation in the control region.
4. Larger MC Samples: The use of significantly larger simulated samples for the main backgrounds reduces the statistical uncertainty associated with the MC templates.

Results
The analysis yields an observed Higgs boson signal strength of $\mu_H = 0.2 \pm 0.7$ relative to the Standard Model prediction. The expected signal strength under the SM hypothesis is $1.0 \pm 0.7$.

• Significance: The observed significance is 0.3 standard deviations ($\sigma$), compared to an expected significance of 1.5$\sigma$ assuming the SM.
• Background Normalization: The post-fit normalization of the dominant non-resonant backgrounds is approximately 85% of the leading-order MC estimate.
• Goodness-of-Fit: The fit yields a $p$-value of 30%, indicating good agreement between the data and the background-plus-signal model.
• Uncertainties: The total uncertainty is dominated by statistical limitations (0.57), followed by MC background template statistics (0.25) and theory/modelling uncertainties (0.24).

Significance
The paper concludes that while no significant evidence for the VBF $H(b\bar{b}) + \gamma$ process was observed, the analysis demonstrates improved sensitivity over previous measurements. The expected significance of 1.5$\sigma$ represents a nearly 50% improvement over the previous full Run-2 analysis of the same channel. The modest result (0.3$\sigma$ observed) is consistent with the SM expectation given the small production cross-section and the statistical limitations of the dataset. The study validates the effectiveness of the new multivariate techniques, kinematic reweighting, and direct fitting strategies, establishing a robust framework for future searches with the full Run-2 dataset and beyond. The results are compatible with the previous measurement ($\mu_H = 1.3 \pm 1.0$) within uncertainties.