Development of a Neural Network-Based Background Suppression Technique for $\Sigma N$ Cusp Spectroscopy at J-PARC

Abstract

A clear spectral enhancement, known as the ``$\Sigma N$ cusp'', has been observed near the $\Sigma N$ threshold in the $d(K^-, \pi^-)$ reaction. To understand the dynamical origin of this enhancement, the J-PARC E90 experiment aims to investigate the missing-mass spectrum with an unprecedented resolution of 0.4 MeV ($\sigma$). In this experiment, a Hyperon Time Projection Chamber (HypTPC) is utilized to detect charged decay products and suppress severe contamination from quasi-free (QF) background processes. While a conventional track multiplicity condition of three (Mt=3) effectively suppresses these QF events, it restricts the signal statistics to approximately 17\% and introduces a mass-dependent acceptance bias that distorts the spectrum. In contrast, events with a track multiplicity of two (Mt=2) offer roughly double the statistical power ($\sim$39\%) with minimal mass dependence, but they suffer from heavy background contamination. To fully exploit the Mt=2 events, we developed an innovative background suppression technique based on a neural network. By constructing a binary classification model using the HypTPC track topology and energy loss ($dE/dx$) as input features, we successfully discriminated the signal from QF backgrounds. This machine learning approach achieves a signal-to-noise ratio comparable to the strict Mt=3 condition while preserving the integrity of the spectral shape. By combining this independent ML-selected Mt=2 sample with the conventional Mt=3 sample, the total usable statistics are effectively doubled compared to traditional methods, significantly enhancing the sensitivity for determining the $\Sigma N$ cusp parameters.