Exploring exotic decays of the Higgs boson to multi-photons at the LHC via multimodal learning approaches

The Standard Model (SM) Higgs boson, the most recently discovered elementary particle, may still serve as a mediator between the SM sector and a new physics sector related to dark matter (DM). The Large Hadron Collider (LHC) has not yet fully constrained the physics associated with the Higgs boson, leaving room for such possibilities. Among the various potential mass scales of the dark sector, the sub-GeV mass range is particularly intriguing. This parameter space presents significant challenges for DM direct detection experiments that rely on nuclear recoils. Various innovative experimental methods are currently under investigation to explore this sub-GeV dark sector. The LHC, functioning as a Higgs factory, could explore this sector once the challenge of identifying DM signals is resolved. Due to the significantly lower mass of particles in the dark sector compared to the Higgs boson, these particles are expected to be highly boosted following the Higgs boson’s decay. However, detecting and identifying these highly boosted particles remains a considerable challenge at the LHC, despite their eventual decay into SM particles. We employ a well-motivated leptophobic $Z_B^{\prime }$ \( {Z}_B^{\prime } \) model as a prototype to analyze the distinctive signatures from Higgs boson exotic decays into multi-photons. These signatures consist of collimated photons that fail to meet the photon isolation criteria, forming jet-like objects. Conventional analyses relying solely on the purity of energy deposits in the electromagnetic calorimeter would fail to detect these signatures, as they would be overwhelmed by background events from Quantum Chromodynamics. To effectively distinguish between such novel signal signatures and SM background events, we leverage advanced machine learning techniques, specifically the transformer encoder in a multimodal network structure. This neural network successfully separates dark sector signals from SM backgrounds, outperforming traditional event selection analyses.