Erlangen 2026 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 73: Calorimeters I

T 73.5: Vortrag

Donnerstag, 19. März 2026, 17:15–17:30, KH 01.012

Processing of ATLAS Liquid Argon Calorimeter Signals by Convolutional Neural Networks and its Impact on Calorimeter Energy Reconstruction — •Manuel Gutsche, Markus Helbig, Arno Straessner, Johann Christoph Voigt, and Philipp Welle — Technische Universität Dresden

During the Phase-II upgrade of the ATLAS detector, over 500 high-performance FPGAs will be installed in the off-detector electronics of the Liquid Argon Calorimeter to cope with the increased luminosity and, therefore, pileup. Under these challenging conditions, the energies of the 182468 detector cells will be reconstructed by the FPGAs. Different methods are being considered. One possible approach is the implementation of 1-dimensional convolutional neural networks (CNNs), which are limited by resource constraints of the Intel Agilex-7 FPGAs to about 400 parameters.

A total of 23 dedicated CNNs are applied to account for differences between cells, mainly in terms of pulse shape and noise. This is achieved by grouping similar cells into clusters, and then training a model of fixed architecture on simulated data expected for a representative cell of each cluster.

These CNNs are integrated into the ATLAS simulation and analysis framework, Athena, in order to compare the performance of the energy reconstruction by CNNs with the current implementation based on an optimal filter. The impact of the two methods on the calculated cell energies, as well as on reconstructed physics objects, is studied.

Keywords: LAr; CNN; detector; calorimeter; readout