15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

Bayesian Deep Learning for Jet-Level Uncertainty Quantification of Jet Energy Scale and Resolution

Not scheduled
15m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for Data Analysis and Theory AI/ML for Data Analysis and Theory

Speaker

Hemalata Nayak

Description

Precise reconstruction of jet kinematics is essential for the physics program of the future Electron-Ion Collider (EIC). Deep-learning-based methods have recently shown strong performance in improving jet energy scale and resolution; however, standard deterministic approaches do not provide the uncertainty quantification required for robust precision measurements. In this work, we introduce a Bayesian neural network (BNN) based on Multiplicative Normalizing Flows to reconstruct the jet transverse momentum (pT) from a hybrid representation combining global jet observables with low-level constituent-level particle features. Compared with traditional jet reconstruction methods and a deterministic deep neural network baseline, the BNN achieves comparable kinematic response while providing uncertainty estimates.
BNN provides epistemic uncertainty estimates that exhibit a positive relationship with the true reconstruction error. These uncertainties can be leveraged through inverse-variance weighting and relative-uncertainty selections to explore the trade-off between statistical efficiency and reconstruction precision. These results suggest that Bayesian deep learning has the potential to provide competitive jet reconstruction performance alongside jet-level uncertainty information, offering a framework for assessing reconstruction reliability in precision jet measurements at the EIC.

Author

Presentation materials

There are no materials yet.