Particle jet classification using edge machine learning

Abstract

This thesis focuses on studying machine learning models for particle jet classification based on Large Hadron Collider (LHC) data. The analysis of the computational challenges of the high-luminosity upgrade (HL-LHC) was aimed at by means of deep learning, with the results being adaptable for the deployment on field-programmable gate arrays (FPGAs) HL-LHC is an upgrade to the present LHC that will boost its luminosity, needed for physicists to explore basic particles and forces with unprecedented granularity and possibly make groundbreaking discoveries.

The used machine learning model for the particle jet classification is based on the Deep Sets B-tagging system. The use of Deep Sets is promising as it is capable of dealing with unordered inputs. Such is the case for the compact muon solenoid (CMS) experiment at LHC, where the order of jet constituents is not fixed and can vary. The model is quantized and pruned as part of the training so that fits to the FPGA-based hardware constraints. Due to the efficiency of the quantization and pruning methods, the task of subsequently deploying the model on the FPGA can be considered feasible