Invisible Higgs search through Vector Boson Fusion: A deep learning approach

TL;DR

This paper demonstrates that deep learning applied to low-level calorimeter data significantly enhances the search for invisibly decaying Higgs bosons in vector boson fusion, surpassing traditional methods in sensitivity.

Contribution

It introduces a deep learning approach using low-level calorimeter data for invisible Higgs searches, outperforming existing techniques and setting more stringent bounds.

Findings

Deep learning improves the invisible Higgs branching ratio bounds by a factor of three.

The method surpasses traditional event kinematics-based techniques in sensitivity.

It provides the most stringent bounds on the invisible Higgs decay with the same data volume.

Abstract

Vector boson fusion proposed initially as an alternative channel for finding heavy Higgs has now established itself as a crucial search scheme to probe different properties of the Higgs boson or for new physics. We explore the merit of deep-learning entirely from the low-level calorimeter data in the search for invisibly decaying Higgs. Such an effort supersedes decades-old faith in the remarkable event kinematics and radiation pattern as a signature to the absence of any color exchange between incoming partons in the vector boson fusion mechanism. We investigate among different neural network architectures, considering both low-level and high-level input variables as a detailed comparative analysis. To have a consistent comparison with existing techniques, we closely follow a recent experimental study of CMS search on invisible Higgs with 36 fb$^{-1}$ data. We find that sophisticated deep-learning techniques have the impressive capability to improve the bound on invisible branching ratio by a factor of three, utilizing the same amount of data. Without relying on any exclusive event reconstruction, this novel technique can provide the most stringent bounds on the invisible branching ratio of the SM-like Higgs boson. Such an outcome has the ability to constraint many different BSM models severely.