[Paper Review] A multivariate study of mass composition for simulated showers at the Auger South Observatory

G. A. Medina Tanco, Sergio J Sciutto|arXiv (Cornell University)|Sep 10, 2001

Astrophysics and Cosmic Phenomena2 references3 citations

TL;DR

This study develops multivariate diagnostics for ultra-high-energy cosmic ray primary composition using simulated air showers at the Auger South Observatory. Employing principal component analysis and neural networks on ground array data from AIRES simulations with QGSJet and Sibyll models, it achieves ~87% iron and ~80% proton identification accuracy, improving to ~91% with hybrid Xmax information, demonstrating robustness across hadronic interaction models.

ABSTRACT

The output parameters from the ground array of the Auger South observatory, were simulated for the typical instrumental and environmental conditions at its Malargüe site using the code sample-sim. Extensive air showers started by photons, protons and iron nuclei at the top of the atmosphere were used as triggers. The study utilized the air shower simulation code Aires with both QGSJet and Sibyll hadronic interaction models. A total of 1850 showers were used to produce more than 35,000 different ground events. We report here on the results of a multivariate analysis approach, including principal component analysis and neural networks, to the development of new primary composition diagnostics.

Motivation & Objective

• To develop robust, multivariate diagnostics for distinguishing primary cosmic ray composition (photons, protons, iron) in ultra-high-energy showers.
• To evaluate the performance of principal component analysis (PCA) and neural networks in separating photon, proton, and iron primaries from ground array data.
• To assess the stability of these diagnostics under different hadronic interaction models (QGSJet vs. Sibyll).
• To determine the impact of including hybrid Xmax information on classification accuracy.
• To provide a practical, statistically sound method for primary composition identification at the Auger Observatory.

Proposed method

• Simulated extensive air showers (EAS) were generated using the AIRES code for photons, protons, and iron nuclei with energies from 10^17.5 to 10^20.5 eV and zenith angles up to 60°.
• Ground array responses were simulated using the sample-sim code, producing 36,620 events by reusing each shower 20 times at different array positions.
• Principal component analysis (PCA) was applied to orthogonalize parameter spaces derived from direct observables and reconstructed quantities (e.g., energy, zenith angle, number of triggered stations).
• A four-layer feed-forward neural network with tan-sigmoid and log-sigmoid transfer functions was trained on surface array parameters: P1–P5 (time and amplitude-weighted signal profiles), energy (P6), zenith angle (P7), and number of triggered stations (P8).
• The network was trained using resilient backpropagation and tested on independent control samples of 11,600 events to evaluate classification performance.
• An additional test included Xmax from hybrid fluorescence detection to assess the impact of combined data on discrimination accuracy.

Experimental results

Research questions

• RQ1 Can multivariate techniques like PCA and neural networks effectively separate primary cosmic ray composition (photon, proton, iron) from ground array data alone?
• RQ2 How does the inclusion of Xmax information from hybrid events affect the accuracy of primary composition classification?
• RQ3 How stable are the classification results when the underlying hadronic interaction model (QGSJet vs. Sibyll) is changed?
• RQ4 What is the optimal set of observable parameters for maximizing primary composition discrimination using surface array data?
• RQ5 To what extent can neural networks reduce ambiguity in primary identification compared to univariate or simple multivariate methods?

Key findings

• The neural network achieved approximately 80% correct classification for protons and 87% for iron nuclei when trained and tested on surface array data alone using the QGSJet model.
• Incorporating Xmax information from hybrid events improved classification accuracy to approximately 90% for protons and 91% for iron nuclei, with a significant reduction in ambiguous intermediate outputs.
• The neural network trained on QGSJet-based simulations maintained high performance when tested on Sibyll-based showers, indicating robustness to hadronic interaction model uncertainties.
• Principal component analysis successfully identified the most informative orthogonal components in the parameter space, enabling dimensionality reduction and revealing the dominant sources of variance in shower development.
• The study demonstrates that multivariate approaches, particularly neural networks with hybrid data, offer a powerful and stable method for primary composition diagnosis at the Auger Observatory.
• The results suggest that surface array data alone can provide substantial composition discrimination, but hybrid detection significantly enhances accuracy and reliability.