SMuK 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
T: Fachverband Teilchenphysik
T 86: ML Methods IV
T 86.5: Vortrag
Mittwoch, 22. März 2023, 18:30–18:45, HSZ/0405
Super-resolution of photon calorimeter images using generative adversarial networks — Johannes Erdmann1, Aaron van der Graaf2, •Florian Mausolf1, and Olaf Nackenhorst2 — 1III. Physikalisches Institut A, RWTH Aachen University — 2TU Dortmund University, Department of Physics
Photons are important objects at collider experiments as, for example, the Higgs boson can be studied with high precision in the diphoton decay channel. For this purpose, it is crucial to achieve the best possible spatial resolution for photons and to discriminate against other particles which can mimic the photon signature.
In this talk, a method to generate photon calorimeter images at increased resolution is presented. The energy depositions of single photons and photon pairs from neutral pion decays are simulated in a lead tungstate crystal calorimeter. Each shower is obtained pairwise, for a calorimeter with a crystal width of 2.2 cm and for a calorimeter with higher resolution, where the number of crystals is increased by a factor of 16. Wasserstein generative adversarial networks are trained to estimate the high-resolution images from their low-resolution counterparts, with a deep residual convolutional neural network used as generator. The properties of the super-resolved calorimeter images are analysed and it is shown that their barycentres can be significantly better localised in the calorimeter. Moreover, classifiers are trained on either super-resolution or low-resolution images to separate single photons from neutral pion decays and their performances are compared.