13–15 Dec 2026
Indian institute of Science Education and Research (IISER), Tirupati
Asia/Kolkata timezone

Interpreting Graph Neural Networks for Quark-Gluon Jet Tagging

Not scheduled
20m
Indian institute of Science Education and Research (IISER), Tirupati

Indian institute of Science Education and Research (IISER), Tirupati

Srinivasapuram Venkatagiri Road, Jangalapalli G, P), Yerpedu, Mandal, Durgiperi, Andhra Pradesh 517619
Oral Presentation Machine Learning and AI for High Energy Nuclear Physics

Speaker

Sunny Kumar Singh (IIT Bombay)

Description

Quark vs Gluon jet tagging is a key handle on the colour-charge dependence of jet quenching. Machine-learning (ML) taggers such as ParticleNet by now outperform classifiers built from standard QCD observables. However, what these networks learn remains an area of active investigation. Recent studies of their internal representations reveal structures closely correlated with constituent multiplicity, radial energy flow and fragmentation, alongside more subtle features that remain hard to characterize. This motivates studying how the reliance of the network on these structures vary across multiplicity and jet kinematics, particularly for regions where misidentification is enhanced.

In this talk, we address this differentially in jet $p_T$ and multiplicity, using a ParticleNet architecture trained on Z+jet samples. We find that the network's decision is largely driven by constituent multiplicity, with additional sensitivity to jet width, and that misidentification is largely prominent for low-multiplicity gluons and high-multiplicity quarks. Furthermore, by probing the network's response to controlled deformations of the jet substructure, we identify the substructure changes that shift the output of the network in the truthward direction. We further examine how the magnitude and direction of these responses depend on constituent multiplicity and jet $p_T$.

Authors

Ankita Budhraja (Nikhef) Balbeer Singh (LIP, Lisbon) Sunny Kumar Singh (IIT Bombay)

Presentation materials

There are no materials yet.