Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 16: Data, AI, Computing 1 (anomaly detection)

T 16.7: Vortrag

Montag, 4. März 2024, 17:30–17:45, Geb. 30.33: MTI

Machine-Learning-based Background Identification for the Radio Neutrino Observatory Greenland — •Philipp Laub for the RNO-G collaboration — ECAP, FAU Erlangen-Nürnberg

The Radio Neutrino Observatory Greenland (RNO-G) is currently being built to detect ultra-high energy (UHE) neutrinos above 10 PeV. UHE neutrinos are detected by measuring radio waves, which are created via the Askaryan effect when UHE neutrinos interact in the ice. Located at Summit Station in Greenland, RNO-G is designed as a wide-spread station array, where each station is equipped with several radio antennas, which are positioned either close to the surface or deep in the ice. Despite the remote and relatively radio-quiet location at Summit, various backgrounds, often of anthropogenic origin, occur in the radio frequency regime. Among them is a rather unexplored background that is correlated to time periods of high wind speed. These “wind events” are impulsive signals of unknown origin, appear in different forms, and can pose a considerable threat to analyses such as cosmic ray searches. Since only little is known about this wind-correlated background and other backgrounds are present in the data, it is difficult to analyze wind events separately from other events.

In this contribution, multiple wind event classes are identified by clustering recorded events into clusters and analyzing event clusters primarily with respect to wind speed. The clustering is performed by first representing events in a low-dimensional latent feature space using a variational autoencoder and then applying the clustering algorithm HDBSCAN to these representations to obtain event clusters.

Keywords: machine learning; deep learning; clustering