Novel Jet Observables from Machine Learning

TL;DR

This paper introduces a machine learning-based method to construct new jet observables for particle discrimination, achieving comparable or superior performance to traditional methods by analyzing the jet's phase space.

Contribution

The paper presents a novel approach to derive jet observables directly from machine learning insights, optimizing discrimination power based on phase space analysis.

Findings

Constructed observables match or outperform traditional ones in discrimination tasks.

Applied method to distinguish H→bb decays from g→bb splittings.

Identified the dominant gluon emission angle as a key discriminating feature.

Abstract

Previous studies have demonstrated the utility and applicability of machine learning techniques to jet physics. In this paper, we construct new observables for the discrimination of jets from different originating particles exclusively from information identified by the machine. The approach we propose is to first organize information in the jet by resolved phase space and determine the effective $N$-body phase space at which discrimination power saturates. This then allows for the construction of a discrimination observable from the $N$-body phase space coordinates. A general form of this observable can be expressed with numerous parameters that are chosen so that the observable maximizes the signal vs.~background likelihood. Here, we illustrate this technique applied to discrimination of $H\to b\bar b$ decays from massive $g\to b\bar b$ splittings. We show that for a simple parametrization, we can construct an observable that has discrimination power comparable to, or better than, widely-used observables motivated from theory considerations. For the case of jets on which modified mass-drop tagger grooming is applied, the observable that the machine learns is essentially the angle of the dominant gluon emission off of the $b\bar b$ pair.