Machine learning techniques in searches for $t\bar{t}h$ in the $h \rightarrow b\bar{b}$ decay channel

TL;DR

This study systematically evaluates machine learning methods, especially gradient boosted trees and neural networks, for detecting the $t\bar{t}h$ process in the challenging $h \rightarrow b\bar{b}$ decay channel at the LHC, highlighting the effectiveness of ensemble models.

Contribution

It provides a comparative analysis of ML techniques for $t\bar{t}h$ detection, emphasizing the success of shallow ensemble models over deeper architectures.

Findings

Gradient boosted trees and neural networks outperform other ML methods.

Extended feature sets and larger data improve model performance.

Shallow models can match or surpass deep models in effectiveness.

Abstract

Study of the production of pairs of top quarks in association with a Higgs boson is one of the primary goals of the Large Hadron Collider over the next decade, as measurements of this process may help us to understand whether the uniquely large mass of the top quark plays a special role in electroweak symmetry breaking. Higgs bosons decay predominantly to \bbbar, yielding signatures for the signal that are similar to $t\bar{t}$ + jets with heavy flavor. Though particularly challenging to study due to the similar kinematics between signal and background events, such final states ($t\bar{t} b \bar{b}$) are an important channel for studying the top quark Yukawa coupling. This paper presents a systematic study of machine learning (ML) methods for detecting $t\bar{t}h$ in the $h \rightarrow b\bar{b}$ decay channel. Among the eight ML methods tested, we show that two models, extreme gradient boosted trees and neural network models, outperform alternative methods. We further study the effectiveness of ML algorithms by investigating the impact of feature set and data size, as well as the structure of the models. While extended feature set and larger training sets expectedly lead to improvement of performance, shallow models deliver comparable or better performance than their deeper counterparts. Our study suggests that ensembles of trees and neurons, not necessarily deep, work effectively for the problem of $t\bar{t}h$ detection.