Light charged Higgs search with deviation neural networks

Abstract

In particle physics, search for signals of new particles in the proton–proton collisions is an ongoing effort. The energies and luminosities have reached a level where new search techniques are becoming a necessity. In this work, we develop a search technique for light-charged Higgs boson (nearly degenerate with [Formula: see text]-boson), which is extremely hard to do with the traditional cut-based methods. To this end, we employ a deep anomaly detection approach to extract the signal (light-charged Higgs particle) from the vast [Formula: see text]-boson background. We construct a Deviation Network (DevNet) to directly obtain anomaly scores used to identify signal events using background data and few labeled signal data. Our results show that DevNet is able to find regions of high efficiency and gives better performance than the autoencoders, the classic semi-supervised anomaly detection method. It shows that employing Deviation Networks in particle physics can provide a distinct and powerful approach to search for new particles.