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

TL;DR

This paper investigates non-resonant di-Higgs production in the $b\bar{b}\gamma\gamma$ final state at the HL-LHC within a BSM scenario featuring new colored scalars (squark-like or leptoquark-like). The authors implement a complete simulation framework, generate signal and backgrounds, and train deep neural networks to maximize discovery significance, finding that two dedicated classifiers (one for QCD backgrounds and one for single-Higgs backgrounds) paired with high-level observables such as $m_{\gamma\gamma}$, $m_{bb}$, and $m_{hh}$ substantially enhance sensitivity. For BM$_{\rm L}$ with $m_{\rm NP}=464$ GeV, they achieve a discovery-level significance of $S\approx 7.3$ at $3\ \text{ab}^{-1}$ and reach discovery at $\mathcal{L}\approx 1.7\ \text{ab}^{-1}$, while BM$_{\rm H}$ with $m_{\rm NP}=621$ GeV yields $S\approx 3.1$ at the same luminosity, indicating reduced sensitivity in that heavier scenario. The results highlight the potential of ML-based strategies to extend the HL-LHC reach for studying the Higgs potential in the presence of new colored states, while also showing that reliance on high-level features is crucial for challenging cases.

Abstract

We investigate di-Higgs production in the $b\bar{b}γγ$ 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. We consider two benchmarks characterized by the mass of the lightest colored scalar, BM$_{\mathrm{L}}$ and BM$_{\mathrm{H}}$, corresponding to 464 GeV and 621 GeV, respectively. Using Monte Carlo simulations for both the signal and the dominant backgrounds, we perform a discovery analysis with deep neural networks, exploring various architectures and input variables. 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. Furthermore, we demonstrate that including high-level features -- such as the invariant masses $m_{γγ}$, $m_{bb}$, and $m_{hh}$, 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 as the invariant mass and momenta of the final particles. For the latter case, we find that architectures processing photon and $b$-jet variables separately can enhance the significance for BM$_{\mathrm{H}}$. Projecting for an integrated luminosity of 3 ab$^{-1}$, we obtain a significance of 7.3 for BM$_{\mathrm{L}}$, while it drops to 3.1 for BM$_{\mathrm{H}}$. In the particular case of BM$_{\mathrm{L}}$, discovery level significance can be reached at 1.7 ab$^{-1}$.

Figures (12)

• Figure 1: Representative Feynman diagrams for di-Higgs production from gluon fusion with colored scalar fields.
• Figure 2: Normalized distributions of the di-Higgs invariant mass $m_{hh}$ at detector level for the two BMs considered in this work.
• Figure 3: DNN architecture used for the analysis with all the features.
• Figure 4: Ranking of features based on the Shapley values approach for classifiers trained against the QCD (upper panels) or Single-$h$ backgrounds (lower panels).
• Figure 5: Probability distributions given by the output of DNN classifiers trained with background events corresponding only to QCD (upper panels) or Single-$h$ (lower panels) processes and signal events of BM$_{\mathrm{L}}$ (left panels) or BM$_{\mathrm{H}}$ (right panels).