Identification of $tqg$ flavor-changing neutral current interactions using machine learning techniques

TL;DR

This paper explores machine learning techniques, including deep learning and boosted decision trees, to improve the detection of flavor-changing neutral current interactions involving a top quark and a gluon at collider experiments.

Contribution

It introduces the SAJA deep learning network for jet-parton assignment, outperforming traditional BDT classifiers in identifying $tqg$ FCNC events.

Findings

SAJA network with qg-discrimination variables achieves the best performance.

Expected upper limits on Br($t o qg$) are 25-35 ext% lower with SAJA.

Deep learning enhances sensitivity to rare FCNC processes.

Abstract

Flavor-changing neutral currents (FCNCs) are forbidden at tree level in the Standard Model (SM), but they can be enhanced in physics Beyond the Standard Model (BSM) scenarios.In this paper, we investigate the effectiveness of deep learning techniques to enhance the sensitivity of current and future collider experiments to the production of a top quark and an associated parton through the $tqg$ FCNC process, which originates from the $tug$ and $tcg$ vertices. The $tqg$ FCNC events can be produced with a top quark and either an associated gluon or quark, while SM only has events with a top quark and an associated quark. We apply machine learning techniques to distinguish the $tqg$ FCNC events from the SM backgrounds, including $qg$-discrimination variables. We use the Boosted Decision Tree (BDT) method as a baseline classifier, assuming that the leading jet originates from the associated parton. We compare with a Transformer-based deep learning method known as the Self-Attention for Jet-parton Assignment (SAJA) network, which allows us to include information from all jets in the event, regardless of their number, eliminating the necessity to match the associated parton to the leading jet. The \SaJa\ network with qg-discrimination variables has the best performance, giving expected upper limits on the branching ratios Br($t \to qg$) that are 25-35\% lower than those from the BDT method.