Speaker
Description
The GRAPES-3 experiment aims to study high-energy cosmic rays through their production mechanisms, propagation, and sources. Located in Ooty at an altitude of 2200 m, it spans an area of 25,000 m$^2$ and comprises about 400 plastic scintillator detectors (SDs) arranged with 8 m spacing to measure the charged component of extensive air showers, along with a dedicated muon detector consisting of 3712 proportional counters. The SD photomultiplier tubes (PMTs), operated at gains of $(3–5)\times10^6$, are sensitive to primary energies up to $\sim$1 PeV, beyond which their response becomes non-linear, necessitating a dedicated correction strategy. GRAPES-3 enables such a correction but only through 105 SDs placed at 16 m separation, each equipped with an additional PMT operated at approximately 50 times lower gain. So to regress a low-gain correspondence for all SDs, in this work, we develop a graph neural network based feature transformation algorithm that maps the high-gain density manifold of all SDs to the corresponding low-gain manifold learned from the 16 m array. We compare multiple models, to understand which feature configuration and architecture captures PMT gain fluctuations and different operating characteristics of non-linear and saturated regions, more effectively. An informed loss function and a detector-independent validation strategy are introduced, enabling performance evaluation even without ground-truth information. We demonstrate improved agreement of the reconstructed density spectrum and Nishimura–Kamata–Greisen parameters with low-gain expectations.