SMuK 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
HK: Fachverband Physik der Hadronen und Kerne
HK 53: AI Topical Day – Computing II (joint session HK/AKPIK)
HK 53.6: Vortrag
Donnerstag, 23. März 2023, 15:15–15:30, HSZ/0103
Deep Leaning Based PID with the HADES detector — •Waleed Esmail1 and James Ritman1,2,3 for the HADES collaboration — 1GSI Helmholtzzentrum für Schwerionenforschung GmbH, 64291 Darmstadt, Germany — 2Forschungszentrum Jülich, 52428 Jülich, Germany — 3Ruhr-Universität Bochum, 44801 Bochum, Germany
The main purpose of a particle identification (PID) algorithm is to provide a clean sample of particle species needed to conduct a physics analysis. The conventional approach used in the HADES experiment is to apply the so-called "graphical cuts" around the theoretical Bethe-Bloch curves of the energy loss as a function of the particle momentum. However, this approach is not optimal, since the distributions resulting from the different particle species overlap. A better approach is based on deep learning algorithms. In our preliminary studies done with the p(4.5 GeV)+p data recently collected by HADES, we were able to improve the separation power of the particle species. The algorithm is based on Domain Adversarial Neural Networks (DANN) trained in a semi-supervised way to simultaneously look at simulated and real data to learn the discrepancies between the two data domains. In this talk we will present our preliminary results, which show that this technique significantly improves the classification of particle species in the experimental data.