Speaker
Description
Neutron-antineutron transitions are a baryon number violating process with ΔB=2, providing a unique insight into potential explanations of the baryon asymmetry of the universe, particularly via post-sphaleron baryogenesis scenarios. The Deep Underground Neutrino Experiment (DUNE), whose primary physics program includes neutrino oscillation measurements, searches for proton decay, and detection of supernova neutrinos, also offers strong prospects to constrain (or discover) this rare process through its high-resolution liquid argon time projection chamber technology.
This work presents a new machine learning approach to neutron-antineutron transition detection at DUNE using a transformer-based neural network architecture via event classification. Our model jointly processes pixel-level detector inputs and latent representations derived through feature embedding modules, with separate pathways for event-level and prong-level (track or shower) information. The architecture employs multi-head attention mechanisms to capture correlations between event topology and prong characteristics, enabling more effective discrimination of signal as opposed to background processes. We demonstrate how this deep learning framework achieves superior classification performance compared to more traditional convolutional neural network based approaches, leveraging a much more complete dimensionality of detector information. The results demonstrate the great potential of transformer-based architectures to significantly enhance background rejection performance within large neutrino detectors.
| Contribution types | Poster |
|---|
