Using deep learning for particle identification and energy estimation in CMS HGCAL L1 trigger

In run 4 of the LHC, the extreme high luminosity is expected to generate an enormous pileup of up to 200 proton-proton collisions for each bunch crossing. This has to be read out at 750 kHz with a maximum latency of 12.5s. In order to disentangle the energy from pileup collision, the upgraded CMS detector for Run-4 will feature a new High Granularity Calorimeter (HGCAL) with unprecedented lateral and longitudinal segmentation. The total number of channels read out into the Level-1 trigger processor will be of the order of 106. To process this data with such small latency, we need to develop sophisticated algorithms. In this report, we aim to use machine learning techniques for electron-photon identification and energy estimation in the L1 Trigger. The idea is to implement the architectures on FPGA boards that will have fast inference, enough to cope with the requirements of the HGCAL