[Paper Review] A Neural Network-Based Monoscopic Reconstruction Algorithm for H.E.S.S. II

Thomas Murach, M. Gajdus+1 more|arXiv (Cornell University)|Sep 2, 2015

Astrophysics and Cosmic Phenomena9 references17 citations

TL;DR

This paper presents MonoReco, a novel monoscopic reconstruction algorithm for the H.E.S.S. II experiment that uses multilayer perceptron neural networks to reconstruct the energy, direction, and particle type (gamma vs. hadron) of very high-energy cosmic rays from data collected by the single large CT 5 telescope alone. The method achieves competitive performance with a 10x speedup over existing methods, enabling real-time on-site analysis and extending H.E.S.S. sensitivity to energies as low as 50 GeV, with successful validation using Crab Nebula data.

ABSTRACT

The H.E.S.S. experiment entered its phase II with the addition of a new, large telescope named CT 5 that was added to the centre of the existing array of four smaller telescopes. The new telescope is able to detect fainter air showers due to its larger mirror area, thereby lowering the energy threshold of the array from a few hundred GeV down to $\mathcal{O}(50\, extrm{GeV})$. Due to the power-law decrease of typical γ-ray and cosmic-ray spectra of astrophysical sources a majority of detected air showers are of low energies, thus being detected by CT 5 only, which motivates the need for a reconstruction algorithm based on information from CT 5 alone. By exploiting such monoscopic events the H.E.S.S. experiment in phase II becomes sensitive in an energy range not covered by H.E.S.S. I and in which the Fermi LAT runs out of statistics. Furthermore the chance of detecting transient phenomena like γ-ray bursts is increased significantly due to the large effective area of CT 5 at low energies. In this contribution a newly developed reconstruction algorithm for monoscopic events based on neural networks is presented. This algorithm uses multilayer perceptrons to reconstruct the direction and energy of the particle initiating the air shower and also to discriminate between gamma rays and hadrons. The performance of this algorithm is evaluated and compared to other existing reconstruction algorithms. Furthermore results of first applications of the algorithm to measured data are shown.

Motivation & Objective

• To develop a monoscopic reconstruction algorithm capable of fully reconstructing particle energy, direction, and type (gamma/hadron) using data from a single telescope, CT 5, in H.E.S.S. Phase II.
• To extend the energy sensitivity of H.E.S.S. II to lower energies (down to ~50 GeV), filling the gap between H.E.S.S. I and the Fermi LAT, especially for transient sources.
• To enable real-time on-site analysis of data by significantly reducing CPU time per event compared to existing stereoscopic reconstruction methods.
• To improve detection sensitivity for low-energy air showers, particularly those initiated by gamma rays from sources like the Crab Nebula, which are predominantly detected by CT 5 alone.

Proposed method

• The algorithm uses multilayer perceptrons (MLPs) within the TMVA framework to perform multivariate analysis on Hillas parameters extracted from the Cherenkov light images recorded by CT 5.
• Image preprocessing includes noise suppression by requiring at least two adjacent pixels to have 5 or 10 p.e. intensity, followed by cuts on number of pixels, total intensity, nominal distance, and image shape parameters (ζ and Θ).
• Three cut configurations (std, loose, extraloose) are applied to balance sensitivity and background rejection, with cuts on pixel number (>3), intensity (>35–60 p.e.), nominal distance (<1.15°), and image shape (ζ > 0.55–0.9, Θ > 0.13°–0.23°).
• Neural networks are trained to reconstruct the primary particle's energy, direction (right ascension and declination), and to classify the event as gamma or hadron using the preprocessed Hillas parameters.
• The reconstruction pipeline is validated using 7.2 hours of Crab Nebula observations with a wobble offset of 0.5° and a mean zenith angle of 47°.
• A forward-folding technique is applied to derive a differential energy spectrum, correcting for energy bias and resolution effects, and the results are compared to the MAGIC collaboration's published spectrum.

Experimental results

Research questions

• RQ1 Can a monoscopic reconstruction algorithm based on neural networks achieve performance comparable to stereoscopic methods in reconstructing energy, direction, and particle type from single-telescope data?
• RQ2 What is the minimum detectable energy threshold achievable with CT 5 alone, and how does it compare to H.E.S.S. I and Fermi LAT?
• RQ3 Can the algorithm enable real-time on-site analysis of H.E.S.S. data, and by how much is processing speed improved over existing methods?
• RQ4 How well does the reconstructed energy spectrum of the Crab Nebula match the published MAGIC spectrum, and what does this imply about the algorithm’s accuracy?
• RQ5 Can the algorithm effectively distinguish gamma-ray events from hadronic background at low energies, especially in the 50–300 GeV range?

Key findings

• The MonoReco algorithm reconstructs particle energy, direction, and gamma/hadron classification with performance competitive to existing stereoscopic methods, enabling full event reconstruction from CT 5 alone.
• The algorithm reduces average CPU time per 5000 events from ~20 seconds (ImPACT) to 0.8 seconds, enabling real-time on-site analysis for rapid source variability checks.
• The energy threshold for H.E.S.S. II is reduced to ~50 GeV, significantly extending sensitivity into the energy range where Fermi LAT loses statistical power.
• The reconstructed differential energy spectrum of the Crab Nebula is well-fit by a log-parabolic function and is compatible with the MAGIC collaboration’s published spectrum within systematic uncertainties.
• A 91σ significance excess was observed toward the Crab Nebula using the reflected background method, with no significant excess in OFF regions, confirming the algorithm’s sensitivity and background rejection capability.
• The sky map shows a point-like source centered on the known Crab position, with no visible artifacts, confirming the accuracy of directional reconstruction.