Several methods can be used to perform statistical inference of primary composition of cosmic rays measured with water Cerenkov detectors as those in use at the Pierre Auger Southern Observatory. In the present work we assess the impact of additional information about the number of muons in the air shower, on the problem of statistical primary mass discrimination. Several tools are studied, including neural networks, principal component analysis and traditional methods in current use in the field. For our case study we use hypothetical plastic scintillators as muon counters, buried at the side and outside the shade of the water Cerenkov tanks. The study is extended to protons and Fe nuclei impinging on an array with two different spacings, 750 and 1500 m and, therefore, suitable to the 1-10 EeV energy range. A prototype of such a detector is under construction.

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 Malarg\"ue 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.