Karlsruhe 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 119: Data, AI, Computing 8 (foundational & transformer models)

T 119.3: Talk

Friday, March 8, 2024, 09:30–09:45, Geb. 30.33: MTI

Leveraging Transformer Models for Gamma-Hadron Separation in SWGO — •Markus Pirke, Jonas Glombitza, Martin Schneider, and Christopher van Eldik for the SWGO collaboration — ECAP, FAU Erlangen-Nürnberg

The Southern Wide-field Gamma-ray Observatory (SWGO) is a proposed next-generation water-Cherenkov gamma-ray observatory in the Southern Hemisphere, thus being complementary to other water-Cherenkov detectors like HAWC (Mexico) and LHAASO (China), which are both located in the Northern Hemisphere. One of the primary challenges of the water-Cherenkov technique, is the effective discrimination of gamma-ray signals from the prevalent hadronic background. Several techniques have been developed in the past, primarily relying on human-designed discrimination variables.

In other scientific areas, recent advancements in deep learning have revealed that employing an end-to-end learning approach, which involves using raw data without the inclusion of handcrafted designed features, frequently improves the performance. One specific deep learning architecture is the Transformer. The self-attention mechanism of the Transformer, initially developed for tasks in natural language processing, offers a promising approach to efficiently handle the complex and variable-sized data in a ground-based observatory with high multiplicities. In this work, this approach will be investigated specifically for Gamma-Hadron separation in SWGO. Performance will be evaluated and additionally the inner workings, meaning the individual building blocks and their functions, of the Transformer will be explained.

Keywords: SWGO; Gamma-ray Observations; Deep Learning; Air Showers; Transformers