Speakers
Description
The IceCube Neutrino Observatory is a cubic-kilometer-scale detector located beneath the Antarctic ice, designed to capture Cherenkov light emitted by secondary particles produced in neutrino interactions. Reconstructing the physical parameters of such events—such as energy, direction, and interaction vertex—from sparse and noisy photon signals is a key task for physics analyses in IceCube.
In recent machine learning research, score-based diffusion models and flow matching techniques have gained attention as powerful generative approaches. These models are capable of capturing complex data distributions and enabling stable conditional generation or inference, making them promising tools for a variety of scientific applications.
In this work, we explore the application of these generative frameworks to IceCube. Specifically, we are developing conditional score-based diffusion models and flow matching models trained on simulated IceCube data. These models are designed to operate in both directions: generating detector responses from particle-level parameters, and reconstructing physical properties from observed detector signals. Our approach aims to provide a new, flexible framework for both event simulation and reconstruction in IceCube.
