Ponente
Descripción
The IceCube Neutrino Observatory is a cubic-kilometer Cherenkov detector located deep in the Antarctic ice at the geographic South Pole. High-energy neutrinos are of great interest as they may originate from powerful astrophysical sources and provide insight into the most extreme environments in the universe. However, at energies above a few hundred TeV, neutrinos traversing the Earth are significantly attenuated, making IceCube primarily sensitive to the southern sky at high energies. A major challenge is that while astrophysical neutrinos are expected to arrive at a rate of roughly 1 $\mu\text{Hz}$, atmospheric muons trigger the detector at a rate of about 2 kHz, leading to a background-to-signal ratio as high as $10^9$.
This extreme imbalance poses a significant computational challenge for high-energy neutrino analyses, which are often limited by the statistical precision of background simulations. The vast majority of simulated air showers fail to pass even the earliest stages of event selection, resulting in a substantial waste of computational resources. In this work, we present a method based on a custom neural network architecture that applies rejection sampling to air showers generated by CORSIKA. The model is trained on existing IceCube simulations to learn and reject events unlikely to survive later processing. Applying rejection sampling before the propagation of Cherenkov photons in ice—the most computationally intensive stage of the simulation chain—enables more efficient use of resources and allows for the generation of higher-statistics, signal-like background samples at fixed computational cost.
