Generative Unfolding of Jets and Their Substructure

TL;DR

The paper tackles high-dimensional unfolding of LHC jet substructure by learning the posterior over particle-level phase space given detector-level observations with conditional flow matching. It decomposes the task into multiplicity, jet kinematics, and constituent generation to manage hundreds of dimensions while preserving physical correlations. It demonstrates percent-level accuracy for jet-level observables and several substructure distributions on Z+jets and boosted-top benchmarks, using both Transformer and Lorentz-equivariant L-GATr architectures. The approach offers a scalable, unbinned alternative to discriminative unfolding, enabling post-hoc measurements of arbitrary jet observables and improved cross-experiment comparisons.

Abstract

Unfolding, for example of distortions imparted by detectors, provides suitable and publishable representations of LHC data. Many methods for unbinned and high-dimensional unfolding using machine learning have been proposed, but no generative method scales to the several hundred dimensions necessary to fully characterize LHC collisions. This paper proposes a 3-stage generative unfolding framework that is capable of unfolding several hundred dimensions. It is effective to unfold the jet-level kinematics as well as the full substructure of light-flavor jets and of top jets, and is the first generative unfolding study to achieve high precision on high-dimensional jet substructure.

Figures (8)

• Figure 1: Visualization of the CFM training procedure for a generative unfolding task. The network output $v_\theta$ can be estimated by a neural network equipped with the L-GATr architecture. See text for details.
• Figure 2: Three-step generative substructure unfolding: multiplicity, jet, and jet constituents.
• Figure 3: Constituents distributions for Z+jets events in the $(\delta_{\log p_T}, \delta_\phi, \delta_\eta)$ parametrization at particle level and detector level.
• Figure 4: Unfolded distributions for the jet multiplicity as predicted by the multiplicity unfolding network and the jet transverse momentum and mass predicted directly with the jet unfolding network.
• Figure 5: Distributions of jet (substructure) observables computed from the unfolded constituents.