Machine Learning for Real-Time Processing of ATLAS Liquid Argon Calorimeter Signals with FPGAs

Abstract

With the High-Luminosity upgrade of the LHC, the number of simultaneous proton-proton collisions will be increased to up to 200. This requires an extensive overhaul of the detector systems. For the ATLAS Liquid Argon calorimeter electronics, 556 high performance FPGAs will be installed to reconstruct the energy for all 182 468 cells at the LHC bunch crossing frequency of 40 MHz. However, the current digital filter used for energy reconstruction (optimal filter) decreases in performance under these high pileup conditions. We demonstrate, that small recurrent or convolutional neural networks can outperform the optimal filter. Prototype implementations of the respective inference code in VHDL show, that the use of these networks on FPGAs is feasible and the resulting firmware fits onto the planned Intel Agilex devices. The full design is capable of processing 384 detector cells per FPGA, by combining parallel instances of the firmware with time division multiplexing.