Improving Neutrino Energy Reconstruction with Machine Learning

If you experience any problem watching the video, click the download button below

Preprint
Report number	arXiv:2405.15867 ; CERN-TH-2024-066 ; MITP-24-052 ; FERMILAB-PUB-24-0276-T ; IPPP/24/26
Title	Improving Neutrino Energy Reconstruction with Machine Learning
Author(s)	Kopp, Joachim (CERN ; Mainz U., Inst. Phys. ; U. Mainz, PRISMA) ; Machado, Pedro (Fermilab) ; MacMahon, Margot (University Coll. London) ; Martinez-Soler, Ivan (Durham U., IPPP)
Imprint	2024-05-24
Number of pages	13
Note	13 pages, 8 figures
Subject category	hep-ex ; Particle Physics - Experiment ; hep-ph ; Particle Physics - Phenomenology
Abstract	Faithful energy reconstruction is foundational for precision neutrino experiments like DUNE, but is hindered by uncertainties in our understanding of neutrino--nucleus interactions. Here, we demonstrate that dense neural networks are very effective in overcoming these uncertainties by estimating inaccessible kinematic variables based on the observable part of the final state. We find improvements in the energy resolution by up to a factor of two compared to conventional reconstruction algorithms, which translates into an improved physics performance equivalent to a 10-30% increase in the exposure.
Other source	Inspire
Copyright/License	preprint: (License: CC BY 4.0)

$Example of a Pearson correlation matrix between several observables in charged-current neutrino--argon interactions with at least one proton and one neutron in the final state, for the DUNE neutrino-mode flux and before including detector responses. We define the proton ($p$) and neutron ($n$) systems by adding up their kinetic energies and three-momenta. We include the total kinetic energy of the proton and systems ($K_p$, $K_n$), the energies of the neutrino and outgoing lepton ($E_\nu$, $E_\ell$), the directions $\cos\theta_{\ell,p,n}$ relative to the beam axis, and the opening angles $\cos\theta_{\ell p,\ell n, pn}$. Note that neutrons are very challenging to reconstruct, so information on the neutron system is typically not available in realistic event records.$ $DNN loss functions when trained on beam neutrino events (upper panel) and atmospheric neutrino events (lower panel) for training (darker) and validation (brighter) data. These results are for the ``0n'' (no neutron reconstuction) scenario, but results are similar for the ``En'' and ``$E_n + \theta_n$'' cases.$ $: \emph{Top:} fractional neutrino energy resolution $\sigma(E_\nu)/E_\nu$ as a function of neutrino energy from DNN-based analyses with no information on final-state neutrons (solid purple), with limited information on the neutron energy (dashed purple), and with information on the neutron energy and direction (dotted purple). For comparison, we also show the energy resolutions anticipated in the DUNE CDR and TDR simulations (magenta), and the resolution of a simple calorimetric method assuming invisible neutrons (gray). \emph{Bottom:} reconstruction bias for the DNN compared to the calorimetric method. \emph{DNNs significantly outperform conventional approaches to energy reconstruction. When information on the energy of final-state neutrons is available, the improvement is more than a factor of two at high energies.} : Fractional neutrino energy resolution, $\sigma(E_\nu)/E_\nu$ (top), and angular resolution, $\sigma(\theta_\nu)$ (bottom) from DNN-based analyses of atmospheric neutrino events with no information on final-state neutrons (solid cyan), and with information on neutron energies and directions included (dashed/dot-dashed cyan). For comparison, we also show the resolutions achievable with simple calorimetric methods (gray curves). The dotted black curve in the bottom panel is based on only the charged lepton kinematics, as in Cherenkov detectors at low energy.$ Show more plots

Back to search

Record created 2024-05-29, last modified 2024-12-14

Similar records

Fulltext:
2405.15867 - Download fulltext

PDF
bcc7e79ac3b9066341510bfaf1bcadb7 - Download fulltext

PDF

External link:
Download fulltext

Fermilab Library Server

Add to personal basket
Export as BibTeX, MARC, MARCXML, DC, EndNote, NLM, RefWorks

CERN Document Server

Access articles, reports and multimedia content in HEP

Main menu

CERN Accelerating science