Potraga za novim česticama pomoću tehnika strojnog učenja

master's thesis

• Title (english): The search for new particles with the use of machine learning techniques
• Author: Tin Hadži Veljković
• Mentor: Nikola Poljak (mentor)
• Committee member: Nikola Poljak (predsjednik povjerenstva)
• Committee member: Davor Horvatić (član povjerenstva)
• Committee member: Mirko Planinić (član povjerenstva)
• Committee member: Krešimir Kumerički (član povjerenstva)
• Committee member: Ivica Smolić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2021-05-27, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

U ovom radu predstavljene su dvije metode za detekciju novih čestica na sudarivačima poput LHC-a. Obje metode se zasnivaju na uporabi neuralnih mreža, točnije autoenkodera i U-Net mreže. Podaci korišteni u ovom radu dani su u sklopu natjecanja LHC Olympics 2020. Dobivena su dva skupa podataka: podaci koji ne sadrže nove čestice, odnosno opisuju pozadinu i podaci koji sadrže pozadinu i nepoznat broj novih čestica. Treniranjem autoenkoderskih modela na podacima koji sadrže samo pozadinu pokušali smo naučiti njene bitne značajke. Nadalje, evaluacijom modela na drugim podacima pokušali smo detektirati nove čestice tražeći događaje sa lošom autoenkoderskom rekonstrukcijom. Ako događaji s novim česticama odstupaju od pozadine, onda se očekuje da će ih autoenkoder gore rekonstruirati. Uporabom U-Net mreže pokušali smo naučiti model da oponaša anti-kT algoritam grupiranja čestica u mlazove te detektirati nove čestice koristeći istu metodologiju kao i za autoenkoderske modele.

Abstract (english)

In this work we will present the use of machine learning techniques for detecting new particles. Two different types of neural networks were used, namely autoencoders and U-Net. The data used was provided as a part of the competition LHC Olympics 2020. We were given two datasets: a dataset containing only background events and a dataset containing an unknown number of new particles along with the background events. Autoencoder models were trained using the first dataset with the goal of learning crucial background information via latent variables. The new particles are considered as anomalies, and therefore we expect the model to poorly reconstruct the data when evaluated on the events containing new particles. The U-Net model was used to learn the anti-kT jet clustering algorithm, and anomaly detection was conducted using the same methodology as for the autoencoders.

• Keywords: fizika izvan standardnog modela; eksperimentalna fizika čestica; strojno učenje; autoenkoderi; U-Net mreža
• Keywords (english): physics beyond the Standard Model; experimental Particle Physics; machine learning; autoencoders; U-Net
• Language: croatian
• URN:NBN: urn:nbn:hr:217:601043
• Study programme: Title: Physics; specializations in: Research; Course: Research; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: magistar/magistra fizike (magistar/magistra fizike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2021-06-17 12:08:55