Reinterpreting the ATLAS HHH to 6b Search with CheckMATE and Rivet: Validation, TRSM Benchmarks, and HL-LHC Prospects
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Reinterpreting the ATLAS HHH to 6b Search with CheckMATE and Rivet".
Title and authors: Kai: So we’re looking at this paper today. The full title is "Reinterpreting the ATLAS HHH to 6b Search with CheckMATE and Rivet: Validation, TRSM Benchmarks, and HL-LHC Prospects" <ref:2606.01966#pg1,6b Search with CheckMATE and Rivet: Validation, TRSM Benchmarks, and HL-LHC>.
Mira: It’s about taking something ATLAS actually did—that search for triple Higgs production in six b-jets—and putting it through these new tools called CheckMATE and Rivet.
Kai: Basically, they aren't just running the search again; they are reinterpreting it. They use a deep neural network and a statistical model based on the HS3 format to check if the input features make sense across different Standard Model scenarios.
Lev: So, what this means is they are testing how well those machine learning inputs actually map onto what physics we expect in different models, like the Standard Model or extensions with extra scalar fields.
Mira: Exactly. It’s about validating the neural network features and setting exclusion limits for these new models using the ATLAS validation material and these new frameworks.
The paper's summary: Kai: To summarize what they did, this paper takes the ATLAS 6b search from HIGP-two thousand twenty-four-thirty-two and applies it to CheckMATE and Rivet <ref:2606.01966#pg1>. They compare the results against a lot of ATLAS validation stuff, like neural networks in ONNX format and a statistical model in HS3 format.
Mira: And they specifically look at constraints on two Beyond the Standard Model models: the Two Real Scalar Model or TRSM, and also the Dark Matter with CP Violation Model or DM-CPV.
Kai: In this TRSM, you extend the Standard Model by adding two extra real singlet scalar fields, X and S. The paper mentions that they use benchmark points called Benchmark Plane three or BP3 which are chosen to maximize triple Higgs boson production while still fitting theoretical and experimental constraints <ref:2606.01966#pg1>.
Lev: So, the core of the analysis is setting up these TRSM benchmarks, defining how many free parameters you have—seven input parameters including masses and mixing angles—and then seeing how those models fit into the existing search structure.
The paper's improvements: Mira: Now they point out some improvements they’ve made to this study. One big thing is building an interface for HS3-format likelihoods so other people can look at complex systematic uncertainty scenarios, like the reduced background uncertainty version of the baseline scenario.
Kai: That interface is important because it lets you dig into those messy things, which is where a lot of the real physics happens in these searches.
Lev: I agree. It’s about making sure that when we actually try to run this on real hardware, we account for those uncertainties properly, not just the simplest case.
Mira: They also added the capability to compare cross-section limits from ATLAS with those from reinterpretation studies, which gives a level of detail that is usually only available inside the experimental collaborations.
Conclusion: Kai: So wrapping up this paper on "Reinterpreting the ATLAS HHH to 6b Search with CheckMATE and Rivet: Validation, TRSM Benchmarks, and HL-LHC Prospects," they show that while the Run two analysis doesn't exclude these new benchmark points yet, extrapolating to the High Luminosity LHC shows sensitivity <ref:2606.01966#pg1,6b Search with CheckMATE and Rivet: Validation, TRSM Benchmarks, and HL-LHC>.
Mira: They found that even without retraining the neural networks, the High Luminosity LHC will be sensitive to many of these models across several systematic uncertainty scenarios.
Lev: The study also showed that you can access some points with just a small data set, like three ab minus one data set in the no-systematics scenario, but almost half of those benchmark points are accessible with the combined six ab minus one data set <ref:2606.01966#pg1>.
Kai: So it suggests that for new physics models like these TRSMs, we can get stronger limits at the future collider by combining this reinterpretation with other results from CMS in both the six b-jets and four b-gamma channels.
Mira: It really boils down to saying that this methodology helps us figure out which new physics models are actually testable at the High Luminosity LHC, and it opens up a path for more comprehensive searches.
Lev: For me, what this implies is that the structure of these constraints is robust; even when you throw in uncertainty scenarios, the underlying physics still shows up in a way that we can probe further if we have enough data.
Jagiellonian University · Faculty of Physics, University of Warsaw
hep-ph
Submitted: 2026-06-01
Updated: 2026-11-17
Comments: 23 pages, 10 figures; minor changes; to appear in EPJC
Code: https://github.com/scipp-atlas/pyhs3
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: We present an implementation in CheckMATE and Rivet of the ATLAS collaboration search for triple Higgs boson production in a six b-jets final state which relies on event selection using a deep neural…
Key concepts
- Triple Higgs Production (HHH)
- This is a rare process where three Higgs bosons are produced simultaneously in particle collisions. The search focuses on finding this specific signature in ATLAS data, particularly when it results in a final state with six b-jets.
- TRSM Benchmark Models
- These models extend the Standard Model by adding two real singlet scalar fields (X and S). The analysis uses these models to test how well the search can distinguish between the Standard Model and these new physics scenarios, using specific benchmark points like BP3 and ODRB.
- CheckMATE/Rivet Frameworks
- These are computational tools used for reinterpreting experimental data. They allow researchers to take raw ATLAS results, apply machine learning methods (like Deep Neural Networks), and perform statistical fits to rigorously test the search results against various theoretical predictions.
Terminology
Summary
We present an implementation in CheckMATE and Rivet of the ATLAS collaboration search for triple Higgs boson production in a six b-jets final state which relies on event selection using a deep neural network and a statistical model based on the HS3 format to perform a thorough validation of neural network input features and exclusion limits in the Standard Model and its extension with two additional singlet scalar fields (TRSM). p. 0 The search relies on event selection using a deep
neural network and a statistical model based on the HS3 format. Owing to a rich validation material
provided by ATLAS, we perform a thorough validation of neural network input features and exclusion
limits in the Standard Model and its extension with two additional singlet scalar fields (TRSM).Where the paper says it
Pruning redundant reasoning tokens
The search focuses on reinterpreting the ATLAS 6b search, HIGP-2024-32 which utilizes advanced experimental techniques including machine learning methods and a statistical model with a simultaneous fit to a histogram of Deep Neural Network (DNN) outputs in the signal and control regions. The implementation is studied in two frameworks widely used for reinterpretation studies, CheckMATE and Rivet. p. 1 1 On the
technical side, the ATLAS search is very interesting as it utilises advanced experimental techniques, including
both machine learning methods and a statistical model with a simultaneous fit to a histogram of Deep Neural
Network (DNN) outputs in the signal and control regions. The implementation is studied in two frameworks
widely used for reinterpretation studies: CheckMATE and Rivet. The implementation is studied in two frameworks
widely used for reinterpretation studies: CheckMATE and Rivet. We perform a detailed comparison
between the implementations and the rich set of ATLAS validation material: the ATLAS collaboration
provided neural networks in the ONNX format; a statistical model in the HS3 [19] format; many validation
histograms [20]; as well as a code snippet in the SimpleAnalysis framework [21].Where the paper says it
Where the paper says it
How the training signal is built
The analysis uses two sets of TRSM benchmark models:
-
The Two Real Scalar Model (TRSM) extends the SM by adding two Z2-symmetric real singlet scalar fields, X and S. p. 1
Where the paper says it
Apart from the SM context, the ATLAS search also looked at constraints on two BSM models: the Two Real Scalar Model (TRSM) [22, 23] and the Dark Matter with CP Violation Model (DM-CPV) [24]. In the TRSM, the SM is extended by adding two Z2-symmetric real singlet scalar fields, X and S.
-
The ATLAS search looked at constraints on two BSM models: the TRSM [24] and the Dark Matter with CP Violation Model (DM-CPV) [24]. p. 1
Where the paper says it
We perform a detailed comparison between the implementations and the rich set of ATLAS validation material: the ATLAS collaboration provided neural networks in the ONNX format; a statistical model in the HS3 [19] format; many validation histograms [20]; as well as a code snippet in the SimpleAnalysis framework [21]. Apart from the SM context, the ATLAS search also looked at constraints on two BSM models: the Two Real Scalar Model (TRSM) [22, 23] and the Dark Matter with CP Violation Model (DM-CPV) [24].
The TRSM model is defined by a scalar potential where the SM is extended by two Z2-symmetric real singlet scalar fields, X and S. p. 1 Finally, we use another
set of TRSM benchmark points to demonstrate a physics motivated case of reinterpetation of the triple-Higgs
search, in LHC Run-2 and at the HL-LHC. In the TRSM, the SM is extended by adding two Z2-symmetric real singlet scalar fields, X and S,
resulting in the following scalar potential before the electroweak symmetry breaking [22]:
V = µ
2
ΦΦ
†Φ + λΦ(Φ†Φ)2
+ µ
2
SS
2 + λSS
4 + µ
2
XX2 + λXX4 + λΦSΦ
†ΦS
2 + λΦXΦ
†ΦX2 + λSXS
2X2
,
(1)
where Φ is the SM doublet and all couplings are real. [22] for a general discussion). The nine free parameters of the TRSM
can be expressed in terms of the physical masses, mixing angles, and vevs. the size of triple
and quartic couplings, different diagrams can dominate the process. Reference [22] defined “Benchmark Plane 3” (BP3), which fulfills theoretical (correct scalar potential
minimum) and experimental (Higgs physics) constraints, while maximizing triple Higgs boson production. [25] a new set of benchmark points was constructed, which obey the
perturbativity constraint at the EW scale and provide an enhanced cross section, at least 100 times, with
respect to the SM prediction for HHH production. The points originate from a scan of a seven-dimensional
parameter space, see Eq.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The validation process involved comparing input features from ATLAS with equivalents from Rivet and CheckMATE and found that the agreement is consistently very good, especially in most of the distributions. The study also extensively validated the statistical inference carried out based on the results of these implementations, which marks the first usage of HS3 in the CheckMATE framework. p. 5 Typically, statistical inference using the results of Rivet analyses is carried out using the Contur [59]
package. However, in this case, this was not practical: Contur does not interface with ROOT, and (for
the moment) the HS3 likelihoods provided by the experiment can only be used for inference with ROOT.Where the paper says it
The performance comparison showed that while overall agreement is very good, some discrepancies were found in variables like aplanarity, sphericity, and transverse sphericity of the leading six jets. p. 5 However, for the sake
of completeness and to aid other reinterpretation efforts going forward, we will briefly describe the issues
encountered. The most significant discrepancy related to the aplanarity, sphericity, and transverse sphericity of the
leading six jets (defined in Table 1): variables which were used as inputs to the resonant and non-resonant
4An example script is contained in the Zenodo repository [62] associated with this publication. Nevertheless, across all the sets of inputs, there were a handful of distributions that still suffer from slightly
poorer agreement. We suspect cos θ is particularly sensitive to jet reconstruction, and hence the simplified
detector simulation in CheckMATE (using Delphes) and in Rivet (using four-vector smearing) is likely
to explain the discrepancies with ATLAS in Figure 3g.Where the paper says it
Where the paper says it
Extrapolation to HL-LHC run
The paper discusses the extrapolation of the ATLAS search to the High Luminosity LHC (HL-LHC) by studying several scenarios of uncertainty evolution at the future collider. The results for all 140 ODRB points under the uncertainty scenarios considered are shown in Figure 10. p. 2 They generally fall into a doubly resonant
category:
pp → h3 → h2(→ h1h1)h1, (5)
shown in Figure 1a. These points, which we will describe as the Optimised Double-Resonant Benchmark
(ODRB) points, to help distinguish from BP3, form an interesting case for a reinterpretation study.Where the paper says it
The expected exclusion tended to be stronger for points with a higher cross-section (as expected), and higher values of m3. p. 1 We
study several scenarios of uncertainty evolution at the future collider, following the ATLAS HL study [17]. We also test the set of TRSM benchmarks points and conclude that at least some of them might be accessible
for exclusion using the full data set expected at the HL-LHC. for the model
point (275, 550) the non-resonant statistical model is published. Interestingly the limit from the non-resonant DNN is stronger
than the resonant one, even though the point is clearly double-resonant. We
study several scenarios of uncertainty evolution at the future collider, following the ATLAS HL study [17]. We also test the set of TRSM benchmarks points and conclude that at least some of them might be accessible
for exclusion using the full data set expected at the HL-LHC.Where the paper says it
Where the paper says it
Where the paper says it
The results for all 140 ODRB points under the uncertainty scenarios that we considered are shown in Figure 10. p. 2 They generally fall into a doubly resonant
category:
pp → h3 → h2(→ h1h1)h1, (5)
shown in Figure 1a. These points, which we will describe as the Optimised Double-Resonant Benchmark
(ODRB) points, to help distinguish from BP3, form an interesting case for a reinterpretation study. We
study several scenarios of uncertainty evolution at the future collider, following the ATLAS HL study [17]. We also test the set of TRSM benchmarks points and conclude that at least some of them might be accessible
for exclusion using the full data set expected at the HL-LHC.Where the paper says it
Where the paper says it
Conclusion and outlook
The validated codes were used to study the new (optimised) doubly-resonant benchmark points suggested by Ref. [25]. We showed that the Run 2 analysis is not sensitive to exclude any of these points, but by extrapolating our results to the HL-LHC we found that even without re-training of the neural networks, the HL-LHC will be sensitive to many of the models in at least some of the systematic scenarios considered. p. 0 Finally,
we discuss expected performance of the search at the High Luminosity LHC in several scenarios of
systematic uncertainty. We present projected exclusion limits for a set of TRSM benchmark models
beyond those considered by ATLAS. (6), we expect this distribution to peak sharply
at approximately 120 GeV, and the fact that two wide bins are divided on this value means that even a
small disagreement in the central value or an asymmetry in one of the distributions can lead to much larger
disagreements in the counts of the adjacent bins. 2 Likelihoods
Another method of validating the implementations in CheckMATE and Rivet is to compare the crosssection limits obtained by ATLAS for some benchmark points with those we achieve from the reinterpretation.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved implementation of CheckMATE and Rivet for triple Higgs boson search by including an interface for HS3-format likelihoods, allowing external users to examine complex systematic uncertainty scenarios such as the “reduced-background-uncertainty” version of the baseline scenario.
-
Capability to compare cross section limits obtained by ATLAS for some benchmark points with those achieved from reinterpretation studies, providing a
level of detail hitherto impractical outside the experimental collaborations.
-
Ability to perform HL-LHC projections using both resonant, non-resonant, and heavy-resonant neural networks across three systematic uncertainty scenarios: baseline HL-LHC scenario (with reduced b-tag/theory uncertainties), an advanced scenario with reduced data-driven background uncertainties, and a no systematics scenario.
-
Enhanced sensitivity to new physics models beyond the doubly resonant benchmark points by allowing application of the reinterpretation to various new physics models with extended scalar sectors, like NMSSM or N2HDM.
Abstract
We present an implementation in CheckMATE and Rivet of the ATLAS collaboration search for triple Higgs boson production in a six b-jets final state. The search relies on event selection using a deep neural network and a statistical model based on the HS3 format. Owing to a rich validation material provided by ATLAS, we perform a thorough validation of neural network input features and exclusion limits in the Standard Model and its extension with two additional singlet scalar fields (TRSM). Finally, we discuss expected performance of the search at the High Luminosity LHC in several scenarios of systematic uncertainty. We present projected exclusion limits for a set of TRSM benchmark models beyond those considered by ATLAS.
Sources
- Observation of a new particle in the search for the Standard Model Higgs boson with the ATLAS detector at the LHC
- Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC
- A detailed map of Higgs boson interactions by the ATLAS experiment ten years after the discovery
- A portrait of the Higgs boson by the CMS experiment ten years after the discovery
- Constraints on the trilinear and quartic Higgs couplings from triple Higgs production at the LHC and beyond
- The Higgs field and early universe cosmology: a (brief) review
- Combination of searches for Higgs boson pair production in $pp$ collisions at $\sqrt{s}=13$ TeV with the ATLAS detector
- Constraints on the Higgs boson self-coupling from the combination of single and double Higgs boson production in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Combined measurements and interpretations of Higgs boson production and decay in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- A search for triple Higgs boson production in the $6b$ final state using $pp$ collisions at $\sqrt{s}=13$ TeV with the ATLAS detector
- Two-real-scalar-singlet extension of the SM: LHC phenomenology and benchmark scenarios
- Triple Higgs Boson Production at the Large Hadron Collider with Two Real Singlet Scalars
- A Simple Model of Dark Matter and CP Violation
- Triple Higgs boson production and electroweak phase transition in the two-real-singlet model
- The Next-to-Minimal Two Higgs Doublet Model
- Benchmarking Di-Higgs Production in Various Extended Higgs Sector Models
- Multi-Higgs Boson Production with Anomalous Interactions at Current and Future Proton Colliders
- HHH Whitepaper
- The anti-k_t jet clustering algorithm
- Combination of searches for nonresonant Higgs boson pair production in proton-proton collisions at $\sqrt{s}$= 13 TeV
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses