Machine-Learning Performance on Higgs-Pair Production Associated with Dark Matter at the LHC

Aug 4, 2026

Upload by Ivy

Di-Higgs production at the LHC with missing transverse energy is investigated using simplified models that parametrize scenarios with heavy scalars and dark matter candidates. The study quantifies how boosted decision trees and deep neural networks improve over cut-based rectangular selections when selecting events with four b-jets and large EmissT. Performance comparisons use kinematic feature sets at 14 TeV, evaluating maximum significance for signal discovery or evidence levels across considered mass benchmarks. A conservative case with 30% background systematics is also examined, yielding consistently promising sensitivities.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages34

Reading Time1 h 26 min

AccessFree to download

Introduction

Phenomenological framework

Machine-learning algorithms for collider analyses

Results

Conclusions

Appendix

Cross sections

Relevant kinematic distributions

Tables of acceptance

arXiv :2401 .03178v2 [hep-ph] 21 Nov 2024 IFT-UAM/CSIC-23-60 Machine-Learning Performance on Higgs-Pair Production Associated with Dark Matter at the LHC Ernesto Arganda 1∗ , Manuel Epele2†, Nicolas I. Mileo2‡, and Roberto A. Morales2§ 1 Departamento de F´ısica Te´orica and Instituto de F´ısica Te´orica UAM-CSIC, Universidad Aut´onoma de Madrid, Cantoblanco, 28049 Madrid, Spain 2 IFLP, CONICET-Dpto. de F´ısica, Universidad Nacional de La Plata, C. C. 67, 1900 La Plata, Argentina Abstract Di-Higgs production at the LHC associated with missing transverse energy is explored in the context of simplified models that generically parameterize a large class of models with heavy scalars and dark matter candidates. Our aim is to figure out the improvement capability of machine-learning tools over traditional cut-based analyses. In particular, boosted decision trees and neural networks are implemented in order to determine the parameter space that can be tested at the LHC demanding four b-jets and large missing energy in the final state. We present a performance comparison between both machine-learning algorithms, based on the maximum significance reached, by feeding them with different sets of kinematic features corresponding to the LHC at a center-of-mass energy of 14 TeV. Both algorithms present very similar performances and substantially improve traditional analyses, being sensitive to most of the parameter space considered for a total integrated luminosity of 1 ab −1, with significances at the evidence level, and even at the discovery level, depending on the masses of the new heavy scalars. Amore conservative approach with systematic uncertainties on the background of 30% has also been contemplated, again providing very promising significances. ∗ ernesto .arganda@uam .es †[manuepele@fisica.unlp.edu.ar](manuepele@fisica.unlp.edu.ar)[ ](manuepele@fisica.unlp.edu.ar)‡[mileo@fisica.unlp.edu.ar](mileo@fisica.unlp.edu.ar) §[roberto.morales@fisica.unlp.edu.ar](roberto.morales@fisica.unlp.edu.ar) Contents 1 Introduction 1 2 Phenomenological Framework 2 3 Machine-Learning Algorithms for Collider Analyses 8 3.1 XGBoost Overview and Architecture ........................... 8 3.2 DNN Overview and Architecture ............................. 12 4 Results 14 5 Conclusions 17 A Cross Sections 18 B Relevant Kinematic Distributions 19 C Tables of Acceptance 22 1 Introduction In the last few years machine learning (ML) has become an standard and basic tool for experimental and phenomenological high-energy physics studies (for reviews see, for instance, [1–11]) . Indeed, ML may be crucial to take full advantage of the data collected at the LHC in order to probe the standard model (SM) and new physics. In this sense, a key issue is whether ML techniques could replace traditional counting methods based on rectangular cuts. The ATLAS and CMS Collaborations have shown in many experimental analyses the potential of ML tools, as boosted decision trees (BDT) and neural networks (NN), to improve the LHC sensitivity, understood as the signal-to-background ratio, to beyond the SM (BSM) physics [12–55] . The main goal of this work is to find out the improvement capability of modern ML tools over cut-based analyses applied toa case study of physical interest: the production at the LHC of Higgs boson pairs associated with dark matter (DM) particles. After the Higgs boson discovery [56, 57], with a mass value of mh = 125.09 ± 0.24 GeV [58] and which seems to be the scalar Higgs boson predicted by the SM [59], the search for extended Higgs sectors represents an intensive experimental program carried out at the LHC by ATLAS and CMS (for recent analyses see, for instance, [44, 50 , 60–85]) . Interestingly, these additional Higgs bosons may serve as portals to dark sectors [86–104], which could manifest in multi-Higgs final states with a large amount of missing transverse energy (EmissT), coming from the potential emi

FAQs

What LHC process and final-state signature are studied?

The work focuses on di-Higgs production associated with missing transverse energy, targeting events with four b-jets and large EmissT.

Which machine-learning methods are compared, and against what baseline?

Boosted decision trees (XGBoost) and deep neural networks (DNN) are compared, with performance measured relative to traditional cut-based search strategies.

How is performance evaluated and what scenarios are considered?

Performance is assessed by the maximum signal significance for different kinematic feature sets at 14 TeV, using benchmarks over the parameter space of the simplified heavy-scalar plus dark-matter setup. A conservative option including 30% systematic uncertainty on the background is also considered.

Aug 4, 2026

Upload by Ivy