Mainz 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
HK: Fachverband Physik der Hadronen und Kerne
HK 42: Hadron Structure and Spectroscopy VII
HK 42.5: Talk
Wednesday, March 30, 2022, 15:15–15:30, HK-H8
Performance Improvement of Deep Machine Learning for the PANDA Software Trigger — •Peiyong Jiang1,2, Klaus Goetzen1, Ralf Kliemt1, Klaus Peters1, and Frank Nerling1 for the PANDA collaboration — 1GSI Helmholtzzentrum fuer Schwerionenforschung GmbH, Darmstadt, Germany — 2Institute of Modern Physics, Chinese Academy of Sciences, Lanzhou, China
Deep machine learning methods have been studied for the PANDA software trigger with data sets from full Monte Carlo simulation using PandaRoot. Following the first comparison of multiclass and binary classification, the binary classification has been selected because of higher signal efficiencies. In total seven neural network types have been compared and the residual convolutional neural network with 4 residual blocks has been chosen. The results of optimized neural networks and those of the conventional method have been compared, showing an efficiency gain of up to 140% for the deep machine learning method. The flatness quality parameters on Dalitz plots and theta-phi projections have been obtained.