Improving heavy Dirac neutrino prospects at future hadron colliders using machine learning

TL;DR

This paper employs machine learning techniques to enhance the detection prospects of heavy Dirac neutrinos at future high-energy hadron colliders, significantly improving sensitivity to neutrino mixing parameters.

Contribution

It introduces the use of ML methods like MLP and BDT to optimize signal-background discrimination in heavy neutrino searches at colliders.

Findings

ML methods improve heavy-light neutrino mixing sensitivity by orders of magnitude.

Reconstructed Z boson and neutrino masses are crucial for background suppression.

Sensitivity to |V_{lN}|^2 reaches 10^{-6} at 14 TeV and 10^{-4} at 100 TeV colliders.

Abstract

In this work, by using the machine learning methods, we study the sensitivities of heavy pseudo-Dirac neutrino $N$ in the inverse seesaw at the high-energy hadron colliders. The production process for the signal is $pp \to \ell N \to 3 \ell + E_T^{\rm miss}$, while the dominant background is $p p \to W Z \to 3 \ell + E_T^{\rm miss}$. We use either the Multi-Layer Perceptron or the Boosted Decision Tree with Gradient Boosting to analyse the kinematic observables and optimize the discrimination of background and signal events. It is found that the reconstructed $Z$ boson mass and heavy neutrino mass from the charged leptons and missing transverse energy play crucial roles in separating the signal from backgrounds. The prospects of heavy-light neutrino mixing $|V_{\ell N}|^2$ (with $\ell = e,\,\mu$) are estimated by using machine learning at the hadron colliders with $\sqrt{s}=14$ TeV, 27 TeV, and 100 TeV, and it is found that $|V_{\ell N}|^2$ can be improved up to ${\cal O} (10^{-6})$ for heavy neutrino mass $m_N = 100$ GeV and ${\cal O} (10^{-4})$ for $m_N = 1$ TeV.