⚛️ high-energy experiments

Search for Higgs boson production in association with another Higgs or an exotic scalar boson in the $b\bar{b}W^{+}W^{-}$ single lepton final state at $\sqrt{s}=13$ TeV with the ATLAS experiment

Using the full ATLAS Run 2 dataset of 140 fb $^{-1}$ at $\sqrt{s}=13$ TeV, this paper presents a search for both non-resonant Higgs boson pair production and resonant production of heavy scalars decaying into $b\bar{b}W^{+}W^{-}$ final states, setting a 95% confidence level upper limit of 35 times the Standard Model prediction for the former and establishing mass-dependent constraints for the latter.

Original author: ATLAS Collaboration

Published 2026-11-06

📖 4 min read🧠 Deep dive

CC BY 4.0

Original author: ATLAS Collaboration

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

The universe is held together by a set of invisible rules that dictate how particles interact and how matter gains weight. At the heart of this system sits a particle known as the Higgs boson, a discovery that confirmed a decades-old theory about why the building blocks of the cosmos have mass. While scientists have spent years studying this particle in isolation, a deeper mystery remains: how does the Higgs boson interact with itself? In the standard model of physics, the accepted framework for understanding the subatomic world, the Higgs boson should be able to produce a pair of its own kind when high-energy protons smash together. This self-interaction is a crucial piece of the puzzle, revealing the shape of the energy field that permeates the universe. However, this process is incredibly rare and difficult to spot, buried beneath a mountain of more common particle collisions.

A researcher using the ATLAS detector at the Large Hadron Collider in Switzerland has now taken a fresh look at this elusive phenomenon. They analyzed a massive collection of data recorded between 2015 and 2018, capturing the aftermath of trillions of proton collisions at an energy level of 13 trillion electron volts. The scientist was hunting for a very specific signature: a pair of Higgs bosons appearing in the same event, where one decays into a pair of bottom quarks and the other transforms into a pair of W bosons, which subsequently decay into a single charged particle and other debris. This specific combination of decay products is challenging to identify because it looks remarkably similar to the background noise produced by the decay of top quarks, a far more common occurrence. To separate the signal from the noise, they employed advanced computer algorithms trained to recognize subtle patterns in the angles and energies of the particles flying out of the collision.

The search yielded no evidence of the Higgs boson pairs appearing more frequently than the standard model predicts. The researcher found that the number of events matching their search criteria was consistent with what would be expected from known background processes alone. Consequently, they established a new upper limit on how often this pair production can occur. Their analysis indicates that if these pairs are being created, they happen no more than 35 times the rate predicted by the standard model. This result is a significant improvement over previous attempts, which had set a limit 300 times the predicted rate, thanks to the larger dataset and refined analysis techniques. While this does not confirm the existence of new physics, it tightens the constraints on theories that propose the Higgs boson behaves differently than expected.

In addition to looking for pairs of standard Higgs bosons, the researcher also searched for a different kind of exotic scenario. They investigated the possibility that a heavy, undiscovered particle could decay into a lighter heavy particle and a standard Higgs boson. This hypothetical chain of events would also leave behind the same tell-tale signature of bottom quarks and W bosons. The researcher scanned a wide range of possible masses for these hypothetical particles, looking for any unexpected spikes in the data that would signal their presence. Just as with the search for standard Higgs pairs, no such signals were found. The data remained smooth and consistent with the background, allowing them to rule out the existence of these heavy particles within the specific mass ranges they tested.

The absence of these discoveries is, in its own way, a vital piece of scientific progress. By confirming that the Higgs boson behaves largely as the standard model predicts, even in these rare and complex interactions, the researcher has narrowed the path for future theories. The work demonstrates that the current understanding of the universe's fundamental forces is robust, even as it leaves the door open for more sensitive searches in the future. The techniques developed to sift through this data, particularly the use of machine learning to distinguish between similar-looking particle showers, represent a major step forward in the ability to probe the deepest layers of reality. As the Large Hadron Collider continues to operate and collect more data, these refined methods will be essential for uncovering the next layer of secrets hidden within the subatomic world.

Technical Summary: Search for Higgs boson production in association with another Higgs or an exotic scalar boson in the $b\bar{b}W^+W^-$ single lepton final state at $\sqrt{s}=13$ TeV with the ATLAS experiment

Problem and Motivation
The Higgs boson ( $H$ ) is central to the Standard Model (SM) electroweak symmetry breaking mechanism. However, the shape of the Higgs potential, which determines the Higgs self-coupling, remains experimentally unprobed with high precision. Measuring Higgs boson pair ( $HH$ ) production is the primary method to access this self-coupling. The SM predicts an extremely small cross-section for non-resonant $HH$ production via gluon-gluon fusion ( $\sigma \approx 31.1$ fb at 13 TeV), dominated by top-quark loop contributions and the Higgs self-interaction vertex.

Furthermore, Beyond the Standard Model (BSM) theories, such as Two Higgs Doublet Models (2HDM) extended with singlets, predict the resonant production of heavy scalar resonances ( $X$ ) decaying into a pair of scalars ( $X \to SH$ ), where $S$ is a heavy scalar and $H$ is the SM-like Higgs. The $b\bar{b}W^+W^-$ final state is a critical channel for these searches. While it possesses the second-highest branching fraction after the $4b$ final state, it is experimentally challenging due to the overwhelming background from top-quark pair ( $t\bar{t}$ ) production, which shares the same final state topology ( $b\bar{b}\ell\nu q\bar{q}'$ ). This analysis aims to probe both non-resonant $HH$ production and resonant $X \to SH$ production in the single-lepton ( $e$ or $\mu$ ) channel using the full ATLAS Run 2 dataset.

Methodology
The analysis utilizes proton-proton collision data recorded by the ATLAS detector at $\sqrt{s}=13$ TeV between 2015 and 2018, corresponding to an integrated luminosity of 140 fb $^{-1}$ . The search targets events with exactly one charged lepton ( $e$ or $\mu$ ), large missing transverse momentum ( $E_T^{\text{miss}}$ ), and at least four jets, including exactly two $b$ -tagged jets.

Event Selection and Kinematic Variables:
To discriminate the signal from the dominant $t\bar{t}$ background, the analysis employs kinematic variables designed to test the compatibility of the event with either the $HH$ or $t\bar{t}$ hypotheses.
- Higgsness ( $H$ ) and Topness ( $T$ ): These variables are constructed by minimizing a $\chi^2$ -like function that compares reconstructed invariant masses of jet combinations to the known masses of the $W$ boson, top quark, and Higgs boson. $T$ tests the $t\bar{t}$ hypothesis, while $H$ tests the $HH$ hypothesis.
- Kinematic Fits: Likelihood-based kinematic fits ( $\Lambda_{HH}$ and $\Lambda_{t\bar{t}}$ ) are performed to further constrain the event kinematics under both hypotheses.
- Additional Variables: A set of 40 input variables, including jet multiplicities, transverse momenta, angular distances ( $\Delta R$ , $\Delta \phi$ ), and transverse masses ( $m_T$ ), are utilized.
Machine Learning Discrimination:
A Deep Neural Network (DNN) is trained to maximize the separation between signal ( $HH$ or $SH$ ) and the three main background processes: $t\bar{t}$ , $W$ +jets, and single-top. The DNN architecture consists of five dense layers with 512 nodes (except the final hidden layer with 256 nodes), utilizing ReLU activation functions and dropout layers to prevent overfitting. The output is a discriminant variable $d_{HH}$ (or $d_{SH}$ for resonant searches), defined as the logarithm of the ratio of the signal probability to the sum of background probabilities.
Background Estimation:
- Prompt Leptons: Backgrounds from $t\bar{t}$ , $W$ +jets, and single-top are modeled using Monte Carlo (MC) simulations (Powheg-Box, MadGraph5_aMC@NLO, Sherpa) and normalized using data-driven control regions (CRs).
- Fake/Non-prompt Leptons: These are estimated using the "fake-factor" method. Fake factors are derived in data-enriched control regions and applied to events in the signal regions (SRs) containing loose leptons.
- Validation: Validation regions (VRs) are defined to ensure the MC modeling of kinematic distributions and background normalizations agrees with data before looking at the SRs.
Signal Regions and Categorization:
The analysis is divided into non-resonant ( $HH$ ) and resonant ( $X \to SH$ ) searches.
- Non-resonant: Events are categorized based on the trigger type (single-lepton or MET) and lepton flavor. The SRs are defined by cuts on the DNN discriminant $d_{HH}$ and $\log(T)$ .
- Resonant: The search covers a grid of scalar masses ( $m_X \in [350, 3000]$ GeV, $m_S \in [170, 2500]$ GeV). To handle the wide range of kinematic properties, signal samples are grouped into six categories (SH0–SH5) based on the boost of the decay products. Dedicated DNN discriminators ( $d_{SHi}$ ) are trained for each group.

Key Contributions

Updated ATLAS Result in Single-Lepton $b\bar{b}W^+W^-$ Channel: This paper presents an updated ATLAS search for $HH \to b\bar{b}W^+W^-$ in the single-lepton channel using the full Run 2 dataset (140 fb $^{-1}$ ), significantly improving upon the previous analysis based on 36 fb $^{-1}$ .
Resonant Search in $b\bar{b}W^+W^-$ : It provides the first investigation of the resonant $X \to SH \to b\bar{b}W^+W^-$ process in the single-lepton decay channel by ATLAS, setting limits on a wide range of scalar masses.
Advanced Background Rejection: The implementation of a multi-output DNN utilizing 40 kinematic variables significantly improves the separation between the signal and the irreducible $t\bar{t}$ background compared to previous analyses relying on simpler kinematic cuts.
Resolved Topology Focus: The analysis explicitly focuses on the "resolved" topology where the $b$ -quarks from the Higgs decay are reconstructed as two distinct jets, excluding "boosted" topologies where they merge into a single large-radius jet.

Results

Non-Resonant $HH$ Search:
No significant excess of events over the SM background expectation was observed. The observed (expected) upper limit at 95% confidence level (CL) on the non-resonant $HH$ production cross-section is:
$\sigma(pp \to HH) < 1.08 \text{ (0.72}^{+0.31}_{-0.22}\text{) pb}$
This corresponds to 35 (23 $^{+10}_{-7}$ ) times the SM predicted cross-section. The observed limit is within +1.1 $\sigma$ of the expected limit.
Resonant $X \to SH$ Search:
No statistically significant discrepancy from the SM prediction was found in any of the resonant signal regions. Exclusion limits at 95% CL were set on the production cross-section $\sigma(pp \to X \to SH) \times \mathcal{B}(SH \to b\bar{b}W^+W^-)$ for various combinations of $m_X$ and $m_S$ .

Significance and Claims
The paper claims a substantial improvement in sensitivity compared to previous ATLAS results in the same channel. The limit on non-resonant $HH$ production is stronger by a factor of 13 compared to the 36 fb $^{-1}$ analysis. The author attributes this improvement to two main factors:

Data Volume: A factor of $\sim$ 2 improvement due to the larger dataset (140 fb $^{-1}$ vs 36 fb $^{-1}$ ).
Analysis Optimization: A factor of $\sim$ 6.6 improvement derived from optimized analysis procedures, specifically the use of the Neural Network discriminant, improved multijet background rejection and estimation, and a refined design of background control regions.

The author notes that while the analysis is not yet competitive with other $HH$ decay channels (primarily due to the large $t\bar{t}$ background), the use of Machine Learning techniques has significantly enhanced the search capability. The results are consistent with the SM, and no evidence for new physics in the form of non-resonant $HH$ or resonant $X \to SH$ production was found.

More like this