Neural Networks as a Composition Diagnostic for Ultra-high Energy Cosmic Rays

We analyze here the possibility of studying mass composition in the Auger data sample using neural networks as a diagnostic tool. Extensive air showers were simulated using the AIRES code, for the two hadronic interaction models in current use: QGSJet and Sibyll. Both, photon and hadron primaries were simulated and used to generate events. The output parameters from the ground array were simulated for the typical instrumental and environmental conditions at the Malargue Auger site using the code SAMPLE. Besides photons, hydrogen, helium, carbon, oxygen, magnesium, silicon, calcium and iron nuclei were also simulated. We show that Principal Components Analysis alone is enough to separate individual photon from hadron events, but the same technique cannot be applied to the classification of hadronic events. The latter requires the use of a more robust diagnostic. We show that neural networks are potentially powerful enough to discriminate proton from iron events almost on an event-by-event basis. However, in the case of a more realistic multi-component mixture of primary nuclei, only a statistical estimate of the average mass can be reliably obtained. Although hybrid events are not explicitly simulated, we show that, whenever hybrid information in the form of $X_{max}$ is introduced in the training procedure of the neural networks, a considerable improvement can be achieved in mass discrimination analysis.