Exploring enhanced non-resonant di-Higgs production at the HL-LHC with neural networks

arXiv:2511.15897 · hep-ph · Submitted 2025-11-19 · Read on arXiv

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: "Exploring enhanced non-resonant di-Higgs production at the HL-LHC with neural networks".

Paper summary: Kai: So we're looking at this paper, "Exploring enhanced non-resonant di-Higgs production at the HL-LHC with neural networks." Essentially, it's about trying to find these di-Higgs events in the b bar b gamma gamma final state at the Large Hadron Collider.

Mira: That’s right. The main idea is they are focusing on scenarios where a new kind of particle—something colored, like a squark or leptoquark—enhances that gluon fusion process that makes di-Higgs production happen.

Lev: So, they're using two specific benchmarks for these new particles, BML and BMH. One has the lightest scalar mass at four hundred sixty-four GeV, and the other is six hundred twenty-one GeV <ref:2511.15897#pg3>. That’s what you can actually build into your simulation setup.

Kai: And what they claim is that they use deep neural networks to analyze the data. They explore different network architectures and input variables to see which ones work best for discrimination between the signal and the background noise.

Mira: The key finding from this paper is that using two separate classifiers actually works better than one single classifier trying to handle everything at once <ref:2511.15897#pg1>. One classifier targets QCD backgrounds, and another focuses on backgrounds involving a single Higgs production process.

Lev: That makes sense from an engineering standpoint because running two specialized models usually gives you better performance than one general model when the signal and background shapes are very different <ref:2511.15897#pg1>. They also point out that using high-level features like invariant masses for gamma gamma, b bar b, and Higgs Higgs, along with transverse momenta and angular separations, significantly improves performance compared to just using low-level features <ref:2511.15897#pg2>.

Kai: And when they project this out for a three ab minus one integrated luminosity at the LHC, they get a discovery significance of seven point three for the BML scenario and it drops down to three point one for the BMH scenario <ref:2511.15897#pg2>.

Mira: So, looking at the title of "Exploring enhanced non-resonant di-Higgs production at the HL-LHC with neural networks," it’s really about using these machine learning tools to push the limits on Higgs physics <ref:2511.15897#pg1>.

Lev: From a hardware perspective, that means if we were building a detector to measure this, we'd need to be very precise with those angular separations and invariant masses they mentioned <ref:2511.15897#pg2>.

Kai: It suggests that the potential for finding these new colored scalars through di-Higgs production is really high, even if it's challenging to get the right signal out of the noise <ref:2511.15897#pg2>.

Mira: The authors also make sure to stress that none of their chosen benchmarks, BML and BMH, are ruled out by other experiments like electroweak precision observables or flavor physics searches <ref:2511.15897#pg3>.

Lev: That's important because it means these scenarios are still viable targets even if we don't see the signal immediately <ref:2511.15897#pg3>.

Kai: The paper suggests that machine learning approaches have substantial potential for maximizing the physics reach of the High Luminosity LHC <ref:2511.15897#pg2>.

Mira: They highlight how using constructed variables, like the invariant masses around two times the mass of these new particles, shows particular efficiency in finding signal events <ref:2511.15897#pg2>.

Lev: The main limitation they mention is that they are using Monte Carlo simulations for both the signal and the dominant backgrounds when performing this discovery analysis <ref:2511.15897#pg3>.

Kai: So, in short, this work shows that smart machine learning techniques, specifically using two dedicated classifiers and rich input features, can significantly improve our ability to search for non-resonant di-Higgs production at the HL-LHC <ref:2511.15897#pg1>.

Mira: It really gets back to how important it is to use these neural network tools when we are trying to probe the Higgs potential and the mechanism of electroweak symmetry breaking <ref:2511.15897#pg2>.

Conclusion: Kai: So we've been looking at how these neural networks help us search for di-Higgs events in this paper about enhanced production at the High Luminosity LHC <ref:2511.15897#pg1>.

Mira: Right, it boils down to using two separate classifiers instead of one general one to really sharpen the signal against those tricky backgrounds <ref:2511.15897#pg2>.

Lev: From where I stand on hardware, having those dedicated classifiers makes sense because you can tune each one specifically for the type of noise it’s dealing with, which is crucial when you're talking about real-world detector constraints <ref:2511.15897#pg3>.

Kai: The authors are pushing these specific benchmarks, BML and BMH, and they show that even in that three ab minus one integrated luminosity scenario, the potential significance is pretty high for BML <ref:2511.15897#pg2>.

Mira: They stress that none of those benchmark scenarios are already excluded by other searches like flavor physics or electroweak precision observables, which means this search space is still open <ref:2511.15897#pg3>.

Lev: That's a good point, because if the signal could be found there, it would tell us something fundamental about the Higgs potential itself <ref:2511.15897#pg2>.

Kai: So the whole implication is that these machine learning tools aren't just an academic exercise; they are essential for maximizing our physics reach at this next generation of collider <ref:2511.15897#pg2>.

Mira: It really shows how important those constructed variables, like those invariant masses around two times the new particle mass, are for pinpointing where the signal is strongest <ref:2511.15897#pg2>.

Lev: And what this means for us on the hardware side is that if we build detectors capable of resolving those precise angles and masses, we have a clear target for optimizing our analysis pipelines <ref:2511.15897#pg3>.

Kai: So they've shown that the method works well in simulation, and now we're looking at how robust this becomes when we actually get data from the LHC <ref:2511.15897#pg2>.

Centro Atómico Bariloche Institute Balseiro CONICET Department of Physics University Nacional de La Plata

hep-ph

Submitted: 2025-11-19

Updated: 2026-11-18

Code: https://github.com/manuepele/SM_LQs

Importance score: 80/100

The gist: We investigate di-Higgs production in the b¯bγγ final state at the LHC, focusing on scenarios where the gluon fusion process is enhanced by new colored scalars…

Key concepts

Di-Higgs Production
This refers to the process where two Higgs bosons are produced simultaneously in a specific way, resulting in a final state of bbarγγ. This process is important for understanding the Higgs potential and how electroweak symmetry breaking occurs at high energies.
Colored Scalars
These are hypothetical new particles, such as supersymmetric partners or leptoquarks, that carry color charge. The paper explores how their presence can enhance the gluon fusion process that creates di-Higgs events.
Discriminatory Classifiers
Instead of one classifier for all data, the study used two specialized neural networks: one trained to distinguish signal from QCD backgrounds and another trained against single-Higgs production backgrounds. This dual approach maximized the system's ability to separate the rare signal from noise.
High-Level Features
These are detailed measurements used by the AI, such as invariant masses (mγγ, mbb) and angular separations. The paper demonstrated that using these complex variables provided much better discrimination power than using only basic, low-level features.

Terminology

Summary

We investigate di-Higgs production in the b¯bγγ final state at the LHC, focusing on scenarios where the gluon fusion process is enhanced by new colored scalars, which could be identified as squarks or leptoquarks. p. 0 Our results show that the discrimination power is maximized by employing two dedicated classifiers, one trained against QCD backgrounds and another against backgrounds involving single-Higgs processes. p. 0 Furthermore, we demonstrate that including high-level features – such as the invariant masses mγγ, mbb, and mhh, as well as the transverse momenta and angular separations of the photon and b-jet pairs – significantly improves the performance compared to using only low-level features. p. 0 Projecting for an integrated luminosity of 3 ab−1, we obtain a significance of 7. p. 03 for BML, while it drops to 3. p. 01 for BMH.

Model and Benchmark Scenarios

The study considers a model with colored scalar fields in the same representations as one generation of quarks, which can be associated with either supersymmetric partners of one generation of quarks or with a set of leptoquarks. Two benchmark (BM) scenarios are chosen: BML and BMH, where the mass of the lightest state is mNP = 464 GeV and 621 GeV, respectively. p. 2 It is important to emphasize that none of the chosen BM points are ruled out by experimental constraints from electroweak precision observables, flavor physics or collider searches, in particular from single Higgs and QCD-induced direct leptoquark productions. p. 2 The di-Higgs production cross-sections from gluon fusion in this model were calculated in Refs. p. 2 [22, 28] (see also Ref. p. 2 [19]).

How the training signal is built

Signal and backgrounds events are simulated with MadGraph5 aMC@NLO (MG5) [53, 54] at √s = 14 TeV with the PDF set PDF4LHC15 nnlo mc including one extra jet with the MLM scheme for the jet-parton matching. The Higgs decay is simulated with MadSpin [56], and parton showering and hadronization are handled with Pythia 8 [57–59] and a fast detector simulation is carried out with Delphes 3 [60] using the ATLAS HL-LHC card. p. 3 The DNNs are trained by splitting the simulated sample of events surviving the acceptance requirements into training and testing subsamples, where the training subsample contains an equal proportion of signal and background events. p. 4 All events are preprocessed to normalize the range of variation of every input feature in such a way that the training set has zero mean and unit variance. p. 4

Where the gains come from

The discrimination power is maximized by training two separate classifiers, one against QCD-induced backgrounds and another against backgrounds involving the production of one Higgs. The invariant masses mγγ, mbb, and mhh are effective discriminating observables. p. 1 In the case of BML, discovery level significance can be reached at 1. p. 07 ab−1. The All Features (AF) classifier shows particular efficiency in identifying signal events within the distinctive region of the mhh distribution around 2 mNP, where BSM contributions are most pronounced. p. 6

Comparison with Alternative Methods

The comparison with classifiers using only low-level features (LLF) revealed the crucial importance of high-level variables. The significant performance gap, especially evident for the more challenging BMH scenario where the LLF significance dropped to 2. p. 9 p. 94, underscores the value of physically meaningful constructed variables. p. 0 p. 0 The AF classifier outperforms the LLF regardless the fraction of events, but the gap between them decreases with the number of events in the training set until the fraction of events reaches the ∼ 50% for BML (∼ 60% for BMH). p. 9 The LLF γ/b classifier outperforms the LLF for all the fraction of events, although its maximum significance achieved is 2. p. 97, which corresponds to the complete training sample being used and entails a slight improvement of 10% with respect to the LLF classifier. p. 9 The AF classifier obtained promising results: significances well above the discovery level (7. p. 93) for BML and at the evidence level (3. p. 91) for BMH, with an integrated luminosity of 3 ab−1. p. 0 Looking forward, these results highlight the substantial potential of machine learning approaches to maximize the physics reach of the HL-LHC.

Conclusions

In this work we have explored the power of machine learning techniques, specifically multiple layer perceptron neural networks, for improving the search for non-resonant di-Higgs production in the b¯bγγ final state at the HL-LHC. A key finding from our analysis is the superior performance achieved by employing two dedicated DNN classifiers—one targeting QCD-induced backgrounds and another focusing on single-Higgs backgrounds—compared to a single classifier handling all backgrounds simultaneously. p. 0 The AF classifier showed particular efficiency in identifying signal events within the distinctive region of the mhh distribution around 2 mNP, where BSM contributions are most pronounced. p. 6 While an alternative architecture with dedicated sub-networks for photons and b-jets provided modest improvement, it could not match the performance of the AF classifier. p. 8 p. 8 The significant performance gap, especially evident for the more challenging BMH scenario where the LLF significance dropped to 2.4, underscores the value of physically meaningful constructed variables. While an alternative architecture with dedicated sub-networks for photons and b-jets provided modest improvement, it could not match the performance of the AF classifier. The measurement of di-Higgs production represents a central goal in the exploration of the Higgs potential and the mechanism of electroweak symmetry breaking. p. 1

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

  1. Implement a two-classifier strategy by training one DNN against QCD backgrounds and another against Single-h backgrounds to improve discrimination power, as stated in Our results show that the discrimination power is maximized by employing two dedicated classifiers, one trained against QCD backgrounds and another against backgrounds involving single-Higgs processes.

  2. Incorporate high level features such as the invariant masses mγγ, mbb, and mhh along with transverse momenta and angular separations of photon and b-jet pairs to significantly improve performance over using only low level features. This is supported by the finding that "including high-level features – such as the invariant masses mγγ, mbb, and mhh, as well as the transverse momenta and angular separations of the photon and b-jet pairs – significantly improves the performance compared to using only low-level features."

  3. Utilize dedicated sub networks within a DNN architecture that separately process photon and b-jet variables to potentially enhance significance for BMH. This is suggested by the statement: We also find that, in some cases, architectures with dedicated sub-networks that separately process the photon and b-jet variables can lead to an improved significance.

  4. Develop a machine learning pipeline capable of maximizing discovery significance by training classifiers using only low level features, testing their performance against those incorporating high level features. This allows for testing the classification power of the algorithm without the external physical inputs provided by the high-level features.

  5. Implement a hybrid classifier trained with both QCD and Single-h backgrounds simultaneously, which was found to be superior to using two separate classifiers in some contexts, as implied by the conclusion that "the significance corresponding to the LLF classifier is systematically above that obtained with the alternative low-level features classifier, which shows the advantage of combining classifiers trained exclusively with each background."

Sources

Related papers