Göttingen 2025 – wissenschaftliches Programm

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

T: Fachverband Teilchenphysik

T 77: Data, AI, Computing, Electronics VIII (Fast ML, Triggers)

T 77.2: Vortrag

Donnerstag, 3. April 2025, 16:30–16:45, VG 2.102

Convolutional Neural Networks on FPGAs for Processing of ATLAS Liquid Argon Calorimeter Signals — Anna Franke, Manuel Gutsche, Markus Helbig, Rainer Hentges, Arno Straessner, •Johann Christoph Voigt, and Philipp Welle — Institut für Kern- und Teilchenphysik, TU Dresden

During the Phase-II upgrade of the ATLAS Liquid Argon Calorimeter, more than 500 high-performance FPGAs will be installed to allow for the energy reconstruction of all 182468 detector cells at the LHC bunch crossing frequency of 40 MHz. We trained 1-dimensional convolutional neural networks (CNNs) to improve the energy reconstruction under high-luminosity conditions with respect to the currently used Optimal Filter. The network architecture has been optimized with a hyperparameter search, where the network size is constrained to 400 parameters. This is motivated by resource estimates from the FPGA firmware prototype implementation. Quantization aware training using QKeras is used to adapt the CNNs to 18 bit fixed point numbers. A revised simulation pipeline is in development to produce training samples for clusters of similar cells. To better evaluate the physics impact of the CNN based readout, the networks are being integrated into the ATLAS common detector simulation and analysis framework, Athena. The inference code of these networks has been implemented in the hardware description language VHDL targeting an Intel Agilex FPGA. A test project targeting a Stratix-10 development kit is available to verify the behaviour of the implementation. Recent results of the CNN training and its firmware realisation will be presented.

Keywords: LAr; CNN; detector; calorimeter; readout