Cosmic rays and ultra-high-energy photons provide complementary information on the origin, acceleration, and propagation of the most energetic particles in the Universe. In the energy range of a few tens of PeV, several models predict a diffuse photon flux hidden below the dominant background of hadron-induced extensive air showers. A measurement or improved constraint of this flux would provide insight into cosmic-ray interactions in the interstellar medium, the transition from Galactic to extragalactic cosmic rays, and possible exotic scenarios such as the decay of super-heavy dark matter. This thesis investigated the detector response, updated the relevant reconstruction sequences, and developed machine-learning methods required to improve PeV photon searches with the Pierre Auger Observatory in the AugerPrime era, exploiting the combined information of the Surface Detector (SD) and the Underground Muon Detector (UMD) in the 433 m configuration.
The first part of the work focused on hadron-induced showers, which constitute the dominant background for photon searches. The energy response of the SD-433 array was characterised in the region of the second knee, including its efficiency, energy bias, and energy resolution, providing the basis for quantitative comparisons between data and simulations for hadron-induced showers. ... mehrSubstantial effort was also devoted to the cleaning of the UMD data, including major improvements to the reconstruction sequence related to the rejection of faulty modules and the identification of spurious measurements.
The hadronic background was then characterised through a measurement of the muon content with the UMD-433. The mean shape of the muon lateral-distribution function (MLDF) was studied and found to be consistent with historical measurements of muons in the same energy range. An event-by-event muon observable was constructed from the event-level parametrisation of the MLDF and interpreted using detector-level proton and iron baselines. The resulting measurement around the second knee indicated a gradual evolution of the muon content from heavier to lighter primaries. The interpretation depends on the hadronic interaction model used to define the baselines, but the measured mean muon content is, on average, compatible with that expected for an oxygen-like composition over much of the studied range. The comparison with expectations based on shower-maximum measurements also suggests a residual muon deficit in simulations, although less severe than that inferred from some higher-energy observables. Comparisons with other experiments led to different levels of agreement, showing that compatibility depends strongly on the chosen composition references.
The second part of the thesis focused on photon-induced showers, corresponding to the signal component of the search. A dedicated reconstruction chain for photon-induced showers was developed, including a photon-specific lateral distribution function, geometry validation, and an energy calibration adapted to the predominantly electromagnetic nature of photon showers. In the reference phase space, the reconstructed photon energy was found to be stable, with a mean bias consistent with zero and a relative resolution of about 14\%. The energy response of the detector was therefore determined for both hadronic and photon showers, setting the baseline for the search analysis.
The central methodological contribution of this thesis was the development of a photon--hadron discriminator based on graph neural networks. Each event was represented as a detector graph constructed from the triggered surface stations and the available underground-muon information, allowing the classifier to exploit the spatial, temporal, and muonic structure of the shower footprint. A multi-view architecture, combining SD-only, UMD-only, and joint SD--UMD branches, was shown to provide strong discrimination power, with the joint branch yielding the best overall separation. Performance studies with proton and iron showers, together with tests under reduced detector information, indicated that the method was robust and largely interpretable. A comparison with a traditional observable employed by the Pierre Auger Collaboration for the same SD-433 and UMD-433 detector combination shows that, when an ideal UMD configuration is available, the gain in separation power is negligible. In that regime, adequate UMD area coverage already provides strong photon--hadron discrimination without the need for neural-network methods. The main advantage of the graph-based approach therefore lay in its robustness: it makes it possible to exploit less favourable SD--UMD configurations and thereby extend photon searches to a broader event selection.
Finally, the analysis chain was validated on a burn sample of real Phase~I data as an intermediate step towards a photon search. The selected events populate the expected hadron-like region of the network-score distributions, with no unexpected instability observed as a function of energy, zenith angle, time, or score correlation. The graph-based selection doubled the number of usable burn-sample events relative to the previous SD-433+UMD selection, showing that the method can recover a substantially larger event sample while keeping the background contamination under control. Under a Phase~I data unblinding scenario, the projected exposure achievable with this discriminator indicates that, in the absence of photon candidates, the existing photon limits in this energy range could be improved by approximately a factor of two.
