Toward a generative modeling analysis of CLAS exclusive $2\pi$ photoproduction

TL;DR

This paper demonstrates that generative adversarial networks can effectively unfold detector effects in high-energy physics experiments, accurately reconstructing multidimensional correlations in two-pion photoproduction data.

Contribution

It introduces the novel application of GANs for unfolding detector effects in high-energy physics, preserving complex correlations in multidimensional phase space.

Findings

GANs successfully reproduce correlated multidifferential cross sections

The method accurately accounts for detector-induced distortions

Uncertainty quantification via bootstrap estimates systematic errors

Abstract

AI-supported algorithms, particularly generative models, have been successfully used in a variety of different contexts. In this work, we demonstrate for the first time that generative adversarial networks (GANs) can be used in high-energy experimental physics to unfold detector effects from multi-particle final states, while preserving correlations between kinematic variables in multidimensional phase space. We perform a full closure test on two-pion photoproduction pseudodata generated with a realistic model in the kinematics of the Jefferson Lab CLAS g11 experiment. The overlap of different reaction mechanisms leading to the same final state associated with the CLAS detector's nontrivial effects represents an ideal test case for AI-supported analysis. Uncertainty quantification performed via bootstrap provides an estimate of the systematic uncertainty associated with the procedure. The test demonstrates that GANs can reproduce highly correlated multidifferential cross sections even in the presence of detector-induced distortions in the training datasets, and provides a solid basis for applying the framework to real experimental data.