SciPost Submission Page
Neutrino Classification Through Deep Learning
by María Fernanda Romo-Fuentes, Luis Eduardo Falcón-Morales
Submission summary
| Authors (as registered SciPost users): | María Fernanda Romo Fuentes |
| Submission information | |
|---|---|
| Preprint Link: | scipost_202410_00066v2 (pdf) |
| Date submitted: | Aug. 31, 2025, 11:44 p.m. |
| Submitted by: | María Fernanda Romo Fuentes |
| Submitted to: | SciPost Physics Proceedings |
| Proceedings issue: | 22nd International Symposium on Very High Energy Cosmic Ray Interactions (ISVHECRI 2024) |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approach: | Computational |
Abstract
Neutrinos are a type of sub-atomic particle whose study is expected to allow us to gain a better understanding of cosmic phenomena and the universe itself. The study of these particles begins with the detection of their passing through a Water Cherenkov detector and, once the data has been collected it is analyzed to determine properties such as its energy, direction of travel and its class. In this project we implemented 4 deep learning methods for the classification of neutrino events as one of three classes: gamma, electron and muon, with the objective of determining which algorithm works best, state of the art methods include custom Convolutional Neural Networks (CNNs) or deep learning algorithms, such as ResNet50 itself, but with other hyper-parameters. Our results show that among the implemented methods, ResNet 50 yielded the best results, with an accuracy of 72.48% and an Area Under the Curve for the efficiency plot of 0.71. These results were obtained by employing the largest dataset available which showed the importance of having a big enough representation of all types of events of all classes in the analysis.
List of changes
-Improving the redaction
-Correcting misspelling and not properly worded sentences
-Including the description of the models
-Citing the appropriate sources for the deep learning architectures
-Clarifying the slicing of data for training, validation and testing
The suggested attaching of a table with the accuracies of all the models as well as showing samples of correctly and incorrectly classified events was not donde gioven the limitation of 6 pages for the submission, though the results that support our conclusions are shown by the graphs in Figures 2 and 3.
I can further provide clarification to any doubt or question you may have. Thank you.
Anonymous on 2025-09-01 [id 5767]
The authors have implemented the suggestions.