The Hyper Kamiokande experiment, a large neutrino detector in Japan, aims to study neutrinos and their role in the evolution of the universe, focusing on phenomena like neutrino oscillations and proton decay. Neutrino oscillations, where neutrinos change flavor as they travel, are central to the project’s objectives. This project seeks to apply deep learning techniques to classify these oscillations, with the goal of improving the identification and characterization of events in neutrino interactions using convolutional neural networks within the WatChMaL software.

The main goal of this project is to develop a data analysis methodology that enhances the performance of deep learning models used for neutrino classification. By tracing input and output data to its source, the project aims to visualize patterns and identify cases where the model succeeds or fails. The methodology will explore three hypotheses: the impact of energy ranges on model performance, the influence of the number of detected hits, and the effect of event positions relative to the detection radius. These insights are expected to improve the model’s accuracy.