Implementation of Deep Neural Networks for the Level 1 Trigger system of the future High-Granularity Calorimeter (HGCAL)

The CMS detector (Compact Muon Solenoid) is a general-purpose particle detector recording the output of proton-proton collisions at the CERN Large Hadron Collider (LHC). It has been recording collision data since 2010 and has led to the discovery of the Higgs boson in 2012, together with the ATLAS detector. In 2026 the LHC will enter a new phase, the high-luminosity phase (HL-LHC). For the HL-LHC, the end-cap calorimeters of the CMS detector will be replaced by a new radiation-resistant and highly granular calorimeter (HGCAL). The high granularity of this detector poses severe challenges on the trigger system of the experiment, a two-stage online system which selects collision events of interest to be recorded. The first stage (L1 trigger) is based on custom FPGAs and takes as input data coming from the detector at an event rate of 40MHz. The purpose of the HGCAL trigger subsystem is to reconstruct clusters of energy deposited in the detector and to classify them as belonging to one of the two fundamental classes: Electromagnetic or Hadronic. This binary classification task needs to be done with a tight latency constraint of a few microseconds. Only simple hand-made clustering algorithms combined with classifiers based on cluster shapes have been developed so far in order to comply with this constraint. This thesis is focused on a new area of development now emerging, consisting in using Deep Neural Network models to perform this classification task. Two possibilities are considered. The first one consists in using the Multilayer Perceptron (MLP) instead of the traditional classifier, while the second one is based on the use of a Convolutional Neural Network (CNN) to analyse raw data avoiding the need for additional hand-made clustering algorithms. This last idea is particularly promising because the possibility to avoid any pre-processing is expected to speed up the process. This approach is interesting because High Level Synthesis tools for Neural Network (NN) models have been recently developed, facilitating the implementation of deep learning models on FPGA. This allows for the creation of dedicated hardware, reducing the time needed to perform a classification task and hopefully reaching real-time efficiency. Therefore, the goal of the thesis is mainly to develop CNN models for particle classification and to evaluate them in terms of performances, logic usage and latency on a target FPGA. Firstly, it is shown that the traditional algorithm plus the MLP is an efficient way of performing the classification. Then, by comparing the performances achieved with those of the more challenging CNN approach, its effective applicability is verified. It is then designed and applied a method for the optimization of the hyper-parameters characterising the CNN. This includes hand-tuning and more advanced techniques such as the Bayesian Optimization for the optimization of the performance of the Python model. Nonetheless, when implementing NNs on FPGA, also the latency and the resources usage need to be considered. Therefore, approaches to find a trade-off between these three constraints are analysed. These can be grouped into two classes. On one hand network compression strategies are studied in order to reduce as much as possible the logic footprint and the latency of the models while keeping high classification performance. On the other hand, a new type of networks called Quantized NNs is considered for further reducing the resources usage.