Developing Machine Learning Models for Jet Pileup Energy Prediction and Correction at the CMS Level-1 Trigger Subsystem

Abstract

The Compact Muon Solenoid (CMS) at the Large Hadron Collider (LHC) in CERN is designed to detect and identify the particles produced in high-energy proton-proton collisions to explore Standard Model (SM) behavior at high energies and search for Beyond Standard Model phenomena. The Level-1 Trigger, the first of CMS’s two-tier trigger system, uses custom hardware processors to select proton-proton interactions whose final state includes signatures of new physics such as high transverse energy electrons, photons, and jets. In order to perform high-precision measurements and searches of new physics, the LHC will undergo an upgrade increasing the instantaneous luminosity, otherwise known as the High Luminosity-LHC (HL-LHC) Project. This increase in luminosity brings about an increase in the pileup, which are the low-energy “soft” collisions that occur simultaneously (within one proton-proton bunch crossing) with the primary high-energy “hard” collision of interest (leading vertex). Pileup adds background to the measured energy interaction. This increase in pileup reduces the performance of many calorimeter algorithms in Level-1 Trigger. One of these algorithms is Princeton’s Calorimeter Image Convolutional Anomaly Detection Algorithm (CICADA), a calorimeter-based event-level Level-1 Trigger autoencoder designed to select anomalous events and assign anomaly scores. In this thesis, we develop several FPGA-deployable machine-learning algorithms under the name Stabilization Network for Anomaly Interference at L1T (SNAIL) ranging from linear regression algorithms to novel Sequential API and Functional API Convolutional Neural Network structures designed to be deployed in series with CICADA in order to find the most effective pileup prediction and correction algorithm. When correcting for pileup energy, we focus on jets due to their high susceptibility to pileup and test the accuracy of our correction algorithms by analyzing the Monte Carlo simulation of H → b anti-b decay events. It is found that the Linear Regression algorithm has the smallest correction error with a Root Mean Squared Error (RMSE) of 2.15 for Zero-Bias events and 3.94 for H → b anti-b, Convolutional Neural Networks utilizing Strip ϕ-ring subtraction can be used as a pileup prediction algorithm as it performs better than Linear Regression in ranges where there is a high number and ET of Calorimeter Trigger Primitives.