Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 9: AI Topical Day – New Methods (joint session AKPIK/T)
AKPIK 9.2: Talk
Thursday, March 23, 2023, 17:45–18:00, HSZ/0004
Transformer-Based Eventwise Reconstruction of Cosmic-Ray Masses at the Pierre Auger Observatory — Martin Erdmann, •Niklas Langner, and Dominik Steinberg — III. Physikalisches Institut A, RWTH Aachen University
As one aspect of the AugerPrime upgrade, scintillators (SSDs) will be added to the water Cherenkov detectors (WCDs) that form the surface detector of the Pierre Auger Observatory. This combined measurement offers the possibility to distinguish individual components of extensive air showers, potentially increasing the mass sensitivity. To efficiently exploit this new potential, novel methods are needed.
We introduce a Transformer-based neural network to reconstruct cosmic-ray masses from joint WCD and SSD measurements that outperforms both recurrent and convolutional networks. Efficient Transformers are employed to analyze and relate the two different sets of time traces on station level while ensuring a reasonable degree of computational demands. A Vision Transformer is then applied to the hexagonal grid of detector stations to process the whole shower footprint.
The Transformer network is trained to simultaneously reconstruct the depth of the shower maximum Xmax as well as the shower’s number of muons on ground Rµ. Both observables can be combined to estimate the primary cosmic-ray mass with an accuracy higher than what can be achieved individually.