The CMS (Compact Muon Solenoid) experiment is one of the particle detectors at CERN's Large Hadron Collider (LHC). Among its objectives is the exploration of physics at the TeV energy scale, focusing on both the study of the Standard Model and the search for potential new physics beyond the Standard Model. The upcoming Phase 2 of the detector upgrade will enhance the Level-1 (L1) trigger system and, for the first time, will allow access to tracking information at this stage, enabling the use of sophisticated algorithms for hadronic jet reconstruction and the suppression of noise caused by multiple proton-proton interactions (pile-up). The goal of this thesis project is to study simulations of the response of the future L1 trigger and to develop algorithms based on machine learning techniques aimed at distinguishing jets originating from b-quarks from those produced by the hadronization of light quarks or gluons. This is intended to improve the detector’s ability to detect physical phenomena involving b-jets. The dataset used comes from simulated proton-proton collisions at the High Luminosity LHC (HL-LHC) and contains all the information that will be available at the L1 trigger level after the upgrade. These data were used to train a binary feed-forward neural network designed to distinguish b-quark jets from light jets. The obtained results, measured through ROC curves with an area under the curve (AUC) of 0.95, demonstrate that the classifier effectively distinguishes between the two classes. This allowed for the definition of three possible working points to select b-jets with different purity levels, depending on the allowed false positive rate. As a starting point for future developments, a preliminary second algorithm based on a convolutional neural network (CNN) has been developed, whose results will be discussed and compared with those produced by the feed-forward network.
L’esperimento CMS (Compact Muon Solenoid) è uno dei rivelatori di particelle al Large Hadron Collider (LHC) del CERN. L’esperimento ha tra i suoi obiettivi quello di esplorare la fisica nella scala energetica del TeV, approfondendo sia lo studio del Modello Standard sia la ricerca di possibile fisica oltre il Modello Standard. La futura Fase 2 dell’aggiornamento del rivelatore produrrà un miglioramento del trigger di primo livello (L1) e consentirà l’accesso alle informazioni del tracciatore per la prima volta a questo stadio, consentendo l’uso di algoritmi sofisticati per la ricostruzione dei jet adronici e la riduzione del rumore causato dalle collisioni multiple (pile-up). L’obiettivo del progetto di tesi è studiare le simulazioni della risposta del futuro trigger di L1 e realizzare algoritmi basati su tecniche di machine learning mirati a distinguere i jet originati da b-quark dai jet provenienti dall’adronizzazione di quark “leggeri” o gluoni al fine di migliorare la capacità del detector di rilevazione di fenomeni fisici comprendenti un jet da quark b. Il dataset utilizzato proviene da simulazioni di collisioni protone-protone ad HL-LHC, e contiene tutte le informazioni che saranno disponibili a livello di L1 trigger dopo l’aggiornamento. Questi dati sono stati utilizzati per l’addestramento di una rete neurale binaria feed-forward, che distingue jet da adronizzazione di quark b da jet leggeri. I risultati ottenuti, misurati per mezzo di ROC curve, la cui area è pari a AUC=0.95, mostrano che il classificatore è in grado di distinguere efficacemente le due classi. Questo ha permesso di definire tre possibili punti di lavoro al fine di selezionare jet da b con diversa purezza, in base alla frazione di falsi positivi ammessa. Come punto di partenza per sviluppi futuri del lavoro, è stato sviluppato in via preliminare un secondo algoritmo basato su una rete neurale convoluzionale (CNN), i cui risultati saranno discussi e confrontati con quelli prodotti dalla rete feed-forward.
Sviluppo di algoritmi di identificazione di Jet per l'upgrade del trigger di primo livello dell'esperimento CMS
D'AGOSTINO, RAFFAELE
2023/2024
Abstract
The CMS (Compact Muon Solenoid) experiment is one of the particle detectors at CERN's Large Hadron Collider (LHC). Among its objectives is the exploration of physics at the TeV energy scale, focusing on both the study of the Standard Model and the search for potential new physics beyond the Standard Model. The upcoming Phase 2 of the detector upgrade will enhance the Level-1 (L1) trigger system and, for the first time, will allow access to tracking information at this stage, enabling the use of sophisticated algorithms for hadronic jet reconstruction and the suppression of noise caused by multiple proton-proton interactions (pile-up). The goal of this thesis project is to study simulations of the response of the future L1 trigger and to develop algorithms based on machine learning techniques aimed at distinguishing jets originating from b-quarks from those produced by the hadronization of light quarks or gluons. This is intended to improve the detector’s ability to detect physical phenomena involving b-jets. The dataset used comes from simulated proton-proton collisions at the High Luminosity LHC (HL-LHC) and contains all the information that will be available at the L1 trigger level after the upgrade. These data were used to train a binary feed-forward neural network designed to distinguish b-quark jets from light jets. The obtained results, measured through ROC curves with an area under the curve (AUC) of 0.95, demonstrate that the classifier effectively distinguishes between the two classes. This allowed for the definition of three possible working points to select b-jets with different purity levels, depending on the allowed false positive rate. As a starting point for future developments, a preliminary second algorithm based on a convolutional neural network (CNN) has been developed, whose results will be discussed and compared with those produced by the feed-forward network.File | Dimensione | Formato | |
---|---|---|---|
DAgostino_Raffaele.pdf
accesso aperto
Dimensione
8.33 MB
Formato
Adobe PDF
|
8.33 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/71931