Heidelberg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 4: Deep Learning
AKPIK 4.2: Vortrag
Donnerstag, 24. März 2022, 16:30–16:45, AKPIK-H13
Amplifying Calorimeter Simulations with Deep Neural Networks — •Sebastian Guido Bieringer1, Anja Butter2, Sascha Diefenbacher1, Engin Eren3, Frank Gaede3, Daniel Hundshausen1, Gregor Kasieczka1, Benjamin Nachman4, Tilman Plehn2, and Mathias Trabs5 — 1Institut für Experimentalphysik, Universität Hamburg, Germany — 2Institut für Theoretische Physik, Universität Heidelberg, Germany — 3Deutsches Elektronen-Synchrotron, Hamburg, Germany — 4Physics Division, Lawrence Berkeley National Laboratory, Berkeley, CA, USA — 5Fachbereich Mathematik, Universität Hamburg, Germany
Speeding up detector simulation over the computationally expensive Monte Carlo tools is a key effort for upcoming studies at the LHC and future colliders. Machine-learned generative surrogate models show great potential to accelerate such and other simulations. However, estimating the relation between the statistics of the training data and the generated distribution of the model is essential to determine the gains and use-cases of these methods.
We present a detailed study of this relation for the concrete physics example of photon showers in a highly granular calorimeter. For established metrics on calorimeter images, the amplification properties of a VAE-GAN model are examined in terms of an approximation to the Jenson-Shannon-divergence between generated data, training data and a high-statistics batch.