Enhancing the Sensitivity for Triple Higgs Boson Searches with Deep Learning Techniques

TL;DR

The paper tackles resonant triple-Higgs production in the all-hadronic $6b$ final state, a channel severely challenged by jet-pairing combinatorics and overwhelming QCD backgrounds. It deploys Spa-Net, a symmetry-preserving attention network that jointly performs jet pairing and event classification, trained on larger $4b$ samples and evaluated on $6b$ samples within the DM-CPV singlet and TRSM frameworks. Spa-Net delivers up to a 40% improvement in expected cross-section limits over a Dense-NN baseline and yields higher pairing efficiency and classification AUC across most benchmark points, signaling a substantial sensitivity gain for multi-Higgs searches at the HL-LHC. By integrating permutation symmetry and multi-task learning, the method offers a practical path to enhance discovery potential in complex multi-jet final states and could generalize to other resonance searches in hadronic environments.

Abstract

Using two benchmark models containing extended scalar sectors beyond the Standard Model, we study deep learning techniques to enhance the sensitivity of resonant triple Higgs boson searches in the fully hadronic $6b$ channel, which suffers from the combinatorial challenge of reconstructing the Higgs bosons correctly from the multiple $b$-jets. More specifically, we employ the framework of Symmetry Preserving Attention Network (\textsc{Spa-Net}), which takes into account the permutational symmetry when a correct pairing of $b$-jets is achieved, to tackle both jet pairing and event classification. Significantly improved efficiency is achieved in signal and background discrimination. When comparing with the conventional Dense Neural Networks, \textsc{Spa-Net} results in up to 40\% more stringent limits on resonant production cross-sections. These results highlight the potential of using advanced machine learning techniques to significantly improve the sensitivity of triple Higgs boson searches in the fully hadronic channel.

Figures (5)

• Figure 1: Feynman diagram for resonant triple Higgs production via gluon fusion in the TRSM, where the heavy scalar $h_3$ decays through $h_2$ to produce three SM-like Higgs bosons.
• Figure 2: The high-level model structure of Spa-Net. Each $E$ is an embedding layer, $T_i$ is the transformer encoder, and $h_i$ is the jet assignment result, which contains two jets $j_i$ for the Higgs decay. The particle transformer is a stack of transformer encoders.
• Figure 3: Jet pairing efficiency of various methods across different benchmark mass points.
• Figure 4: Classification performance at different mass points. The AUCs of various classifiers are shown. All models are trained on the combined $4b$ dataset described in section \ref{['sec:sample_preparation']}.
• Figure 5: Expected 95% CL upper limits on the triple Higgs production cross-section, $\sigma(pp \to hhh)$, at the 14-TeV LHC for different classifiers and luminosities. The limits are computed using the $\text{CL}_{\text{s}}$ method.