Transformer Neural Networks in the Measurement of $t\bar{t}H$ Production in the $H\,{\to}\,b\bar{b}$ Decay Channel with ATLAS

Using 140 fb⁻¹ of 13 TeV proton-proton collision data from the ATLAS detector, this paper presents a measurement of $t\bar{t}H$ production in the $H\to b\bar{b}$ decay channel that leverages transformer neural networks to significantly improve signal-background discrimination and Higgs transverse momentum reconstruction, resulting in an observed excess of 4.6 standard deviations over the background-only hypothesis.

The Gist

Imagine the Large Hadron Collider (LHC) as a massive, high-speed particle collider where scientists smash protons together to recreate the conditions of the early universe. Inside this chaos, physicists are hunting for a very specific, rare event: a "Higgs boson" (a fundamental particle that gives other particles mass) being born at the same time as a pair of "top quarks" (the heaviest known particles).

This paper describes a new, smarter way to find this rare event using data collected by the ATLAS detector between 2015 and 2018. Here is the breakdown of what they did and what they found, using everyday analogies.

The Challenge: Finding a Needle in a Haystack

The specific event they are looking for is the Higgs boson decaying into two "bottom quarks." The problem is that the universe produces a massive amount of "background noise"—specifically, top quark pairs produced with random jets of particles—that looks almost exactly like the signal they want.

Think of it like trying to hear a specific song playing in a crowded, noisy stadium. The song is the signal (the Higgs + top quarks), and the crowd cheering is the background noise (top quarks + random jets). In previous attempts, the "crowd" was so loud that it was hard to tell if the song was actually playing.

The New Tool: "Transformer" Neural Networks

The biggest innovation in this paper is the use of Transformer Neural Networks. You might know Transformers from AI tools that write essays or translate languages. In this context, the scientists used them as a super-smart sorting machine.

• Why Transformers? In a particle collision, there is no fixed order to the particles flying out. A Transformer is special because it doesn't care about the order; it looks at the whole picture at once. It's like a detective who can look at a messy crime scene with 50 scattered clues and instantly understand the story, whereas an older method might have tried to look at the clues one by one in a specific order.
• The Sorting Job: The AI was trained to look at every collision and decide: "Is this the rare Higgs song, or is it just the noisy crowd?" It sorts events into different categories (like "Signal," "Crowd Type A," "Crowd Type B") with incredible precision.

The Strategy: Loosening the Net

Because the AI is so good at telling the difference between the signal and the noise, the scientists could change their strategy.

• The Old Way: In the past, they had to set a very strict "pre-selection" filter (like a bouncer at a club) to keep the noise out. This meant they only looked at the cleanest, most obvious events, but they missed a lot of the actual signal because the filter was too tight.
• The New Way: With the AI acting as a super-smart bouncer, they could let more people into the club (loosen the pre-selection). They let in three times as many events as before. The AI then did the heavy lifting, sorting the good ones from the bad ones later in the process. This tripled their ability to catch the signal.

They also built a "reconstruction" network. Imagine trying to figure out the speed of a car just by looking at the skid marks it left behind. The AI looks at the debris from the collision and calculates the exact speed (transverse momentum) of the Higgs boson, allowing them to study how it behaves at different speeds.

The Results: A Clear Signal

After feeding 140 units of collision data (a massive amount of information) through this new system, the results were clear:

1. They found the signal: They observed an excess of events that matched the Higgs boson prediction.
2. Statistical Confidence: The chance that this was just a random fluke of the background noise is incredibly low. The result has a significance of 4.6 standard deviations.
   • Analogy: If you flipped a coin and got heads 4.6 times in a row more often than chance would allow, you'd be pretty sure the coin was rigged. Here, the "coin" is the data, and it strongly suggests the Higgs boson is there.
3. Comparison to Theory: The number of Higgs bosons they found matches what the Standard Model of physics predicted (within the margin of error). It's like the AI predicted the weather, and the actual weather matched the forecast perfectly.

The Bottom Line

This paper presents a re-analysis of old data using a new, powerful AI technique. By using "Transformer" neural networks to sort through the chaos of particle collisions, the ATLAS team was able to triple their sensitivity, reduce the noise, and confirm the existence of Higgs bosons produced alongside top quark pairs with high confidence. It is currently the most precise measurement of this specific process ever made.

Technical Summary

Technical Summary: Transformer Neural Networks in the Measurement of $t\bar{t}H$ Production

Problem and Context
The direct measurement of the Yukawa coupling between the top quark and the Higgs boson is critical for understanding electroweak symmetry breaking and the stability of the Higgs vacuum potential. While indirect measurements exist via quantum loops, direct observation requires the associated production of a Higgs boson with a top quark pair ($t\bar{t}H$). The ATLAS collaboration previously published a measurement of this process in the $H \to b\bar{b}$ decay channel using the full Run 2 dataset (140 fb$^{-1}$ at $\sqrt{s}=13$ TeV). However, this channel faces a significant challenge: a large irreducible background from top quark pair production in association with additional jets ($t\bar{t} + \text{jets}$), particularly those containing heavy-flavour jets. Previous analyses struggled with signal acceptance and background modelling, necessitating a re-analysis with improved strategies to achieve higher precision.

Methodology
This paper presents a re-analysis of the $t\bar{t}H(b\bar{b})$ process utilizing a sophisticated multi-variate analysis (MVA) strategy driven by transformer neural networks. The core methodological innovations include:

• Transformer Architecture: The analysis employs transformer networks, originally developed for natural language processing, for both event classification and Higgs boson reconstruction. These networks are chosen for their invariance to the permutation of input objects and their ability to handle variable cardinality, which is essential for particle collision events where the number of final-state particles varies and no fixed ordering exists.
• Input Variables: The networks utilize low-level four-vector and $b$-tagging information from all jets, leptons, and missing transverse energy.
• Event Classification and Region Definition: Unlike previous analyses that relied on strict kinematic cuts, this strategy uses neural network outputs to define regions. The background ($t\bar{t} + \text{jets}$) is split into five categories based on the flavour of jets not originating from top decays: $t\bar{t} + \text{light}$, $t\bar{t} + \geq 1c$, $t\bar{t} + \geq 2b$, $t\bar{t} + 1B$ (jets with collimated $b$-hadrons), and $t\bar{t} + 1b$.
  • Discriminants: Multi-class classification probabilities are converted into six class discriminants to separate the $t\bar{t}H$ signal from the five background categories. Events are assigned to signal regions (SR) or control regions (CR) based on these discriminants.
  • Preselection Loosening: The robustness of the classification allows for a loosening of kinematic preselections, increasing signal acceptance by a factor of three compared to the first full Run 2 analysis. Selections target single-lepton ($\geq 5$ jets) and dilepton ($\geq 3$ jets) channels with $\geq 3$ $b$-tags, plus a dedicated "boosted" channel for high-$p_T$ Higgs bosons where both $b$-quarks are contained in a single jet.
• Higgs Reconstruction: A separate reconstruction network uses a pairing layer (similar to SPANet) to identify the two jets originating from the Higgs decay. This allows for the reconstruction of the Higgs transverse momentum ($p_T^H$), which is used as a fit variable in the boosted channel and for differential measurements.
• Statistical Analysis: A simultaneous binned profile likelihood fit is performed across all regions. The fit extracts the inclusive $t\bar{t}H$ cross-section and a differential cross-section in six Simplified Template Cross-Section (STXS) bins defined by the truth $p_T^H$.

Key Contributions
The primary contribution of this work is the successful implementation of transformer neural networks to simultaneously constrain the dominant $t\bar{t} + \text{jets}$ background (split into five disjoint categories) and measure the $t\bar{t}H$ signal. This approach supersedes previous ATLAS measurements by:

1. Implementing an overhauled analysis strategy that relies on multi-variate discriminants for region definition rather than fixed kinematic cuts.
2. Achieving a three-fold increase in signal acceptance through loosened preselections.
3. Providing a differential measurement of the cross-section in $p_T^H$ bins, utilizing the reconstructed $p_T^H$ from the neural network.

Results
The analysis yields an observed event excess of 4.6 standard deviations over the background-only hypothesis, compared to an expected significance of 5.4 standard deviations.

• Inclusive Cross-Section: The measured inclusive cross-section is $\sigma_{t\bar{t}H}^{\text{obs}} = 411^{+101}_{-92} \text{ fb}$ (statistical and systematic uncertainties combined). This is compatible with the Standard Model (SM) prediction of $\sigma_{t\bar{t}H}^{\text{SM}} = 507^{+35}_{-50} \text{ fb}$ at $m_H = 125.09$ GeV.
• Differential Cross-Section: The differential cross-section results across the six STXS bins show compatibility with the SM prediction, with a $p$-value of 89%. Correlation coefficients between bins do not exceed 30%.
• Uncertainties: Unlike previous analyses where $t\bar{t} + \geq 1b$ background modelling was dominant, the leading sources of systematic uncertainty in this measurement are the modelling of the $t\bar{t}H$ signal itself.

Significance
The paper claims that this measurement represents the most precise single-channel $t\bar{t}H$ cross-section measurement to date. By leveraging transformer networks, the analysis demonstrates that advanced machine learning techniques can significantly improve the discrimination between complex signal and background processes in high-energy physics. The work validates the use of such networks for simultaneous background constraint and signal extraction, setting a new standard for precision measurements in the $H \to b\bar{b}$ channel.