Speakers
Description
One of the aims of large water Cherenkov neutrino experiments like Hyper-Kamiokande (Hyper-K) experiment is to detect low energetic neutrinos to allow studies in the solar and supernova sector. This requires pattern recognition of very faint signals on top of the inevitable background noise. Such developments have to begin already at the stage of data acquisition, where Hyper-K is targeting neutrino energies down to 3 MeV. The fully software-based trigger suite provides a good basis for this, allowing to apply an array of different triggers to target different signals. This talk will describe two machine-learning based trigger approaches aimed at the Hyper-K far detector: a supervised transformer encoder, and an unsupervised anomaly-detection method trained purely on detector noise. We describe details of the two models and show performance estimates for trigger efficiencies as well as runtimes on GPU and CPU. Both approaches show promising gains relative to standard non-ML trigger algorithms: the supervised transformer significantly improves low-energy signal efficiency, and the noise-trained anomaly-detection method also performs competitively and exceeds some conventional trigger strategies.
| I read the instructions above | Yes |
|---|