Publications
Publication Information
Title | Artificial Intelligence for reduction and interpretation of subatomic particle production data |
Authors | Y. Alanazi Awadh, P. Ambrozewicz, M. Battaglieri, G. Costantini, A. Hiller Blin, E. Isupov, T. Jeske, Y. Li, L. Marsicano, W. Melnitchouk, Victor Mokeev, N. Sato, A. Szczepaniak, T. Viducic |
JLAB number | JLAB-PHY-21-3519 |
LANL number | (None) |
Other number | (None) |
Document Type(s) | (Journal Article) |
Associated with EIC: | No |
Supported by Jefferson Lab LDRD Funding: | Yes |
LDRD Numbers: | 2019-LDRD-13 |
Funding Source: | Nuclear Physics (NP) |
Other Funding: | STRONG - 2020 - No 824093 |
Journal Compiled for Nature | |
Publication Abstract: | In nuclear and particle physics scattering reactions the most direct observable connecting rates measured in experiments to the underlying probability of the scattering process is the cross section. Cross sections are interpreted in terms of fundamental interactions described by the Standard Model (SM) of nuclear and particle physics. Reactions with more than two particles in the final state are described in a multi-dimensional space in order to fully account for the correlations between the observed particles, which in turn reflect the underlying dynamics. Practical manipulation of the data, however, requires projection to low dimensions, making it difficult to preserve the original correlations between the particle variables. We present a novel approach based on generative adversarial networks (GANs) to encode correlations contained in a multi-dimensional cross section into neural networks (NNs). The NNs trained on event-level experimental data are used as the core of a Monte Carlo event generator (MCEG) that mimics the original dataset, preserving the original information in an efficient and compact way. The procedure was tested on a photo-induced reaction at GeV-energy scale, where strong-interaction dynamics produce hadronic resonances with highly nontrivial correlations between the observed particle momenta. An uncertainty quantification procedure was developed for GAN-based MCEGs, and selected observables extracted from experimental and GAN-generated synthetic data systematically compared. The proposed method can be extended to higher multiplicity events of interest to hadrons spectroscopy and hadron structure studies, as well as to high-statistics multi-dimensional analysis in other branches of subatomic physics. |
Experiment Numbers: | E04-021 |
Group: | Hall B |
Document: | docx |
DOI: | |
Accepted Manuscript: | |
Supporting Documents: | |
Supporting Datasets: |