Primjena neuronskih mreža u modeliranju svojstava atomskih jezgara

master's thesis

• Title (english): Application of neural networks in modelling the properties of atomic nuclei
• Author: Krešimir Šura
• Mentor: Nils Paar (mentor)
• Committee member: Nils Paar (predsjednik povjerenstva)
• Committee member: Tamara Nikšić (član povjerenstva)
• Committee member: Nikola Poljak (član povjerenstva)
• Committee member: Matko Milin (član povjerenstva)
• Committee member: Krešimir Kumerički (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2023-09-28, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

Energija vezanja temeljno je svojstvo atomske jezgre. Može se mjeriti eksperimentalno, ali ne za sve jezgre koje su potrebne za modeliranje nukleosinteze. Razvijeni su mnogi teorijski modeli koji opisuju energije vezanja kroz cijelu mapu nuklida, a najbolji medu njima postižu srednje kvadratno odstupanje RMS od eksperimentalnih vrijednosti od oko 500 keV-a. U ovome radu proučen je pristup određivanja energije vezanja atomskih jezgara pomoću strojno naučenih neuronskih mreža koristeći eksperimentalne podatke iz baze nuklearnih podataka AME2020. Prediktivna moć strojno naučenih modela uspoređena je s eksperimentalnim podacima, kao i s drugim metodama određivanja energija vezanja. Ključni korak razvoja neuronskih mreža za primjenu u nuklearnim istraživanjima je odabir značajki modela, kojima unosimo relevantne fizikalne informacije u proces strojnog učenja. Neuronska mreža s deset odabranih značajki postigla je RMS = 248 keV. Ovakav rezultat bolji je od tradicionalnih modela nuklearne fizike.

Abstract (english)

Binding energy of an atomic nucleus is one of its fundamental properties. It can be measured in experiment, however not for all nuclei necessary in nucleosynthesis modeling applications. Many theoretical models that calculate binding energies for all nuclei have been developed, the best of which describe binding energies with a root mean square RMS deviation from experimental data of 500 keV. In this work a novel approach to modelling binding energies has been investigated: machine learning using neural networks utilizing experimental data from the AME2020 mass evaluation. Predictive power of machine learning models has been compared with experiment, as well as other nuclear models. Crucial step of developing neural networks for use in nuclear physics has been feature engineering, a way to introduce and encode physically important information in the process of machine learning. Best performing neural network had ten selected features, achieving an RMS = 248 keV, a result surpassing traditional nuclear physics models.

• Keywords: nuklearna struktura; model energija vezanja atomske jezgre; strojno učenje; neuronske mreže; duboko učenje; odabir značajki; fizikalno informirane neuronske mreže
• Keywords (english): nuclear structure; nuclear models; binding energy; machine learning; neural networks; deep learning; feature engineering; physics informed neural network
• Language: croatian
• URN:NBN: urn:nbn:hr:217:519531
• Study programme: Title: The University Integrated Undergraduate and Graduate Programme in Physics; Research Physics; specializations in: Research Physics; Course: Research Physics; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: sveučilišni magistar fizike (sveučilišni magistar fizike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2023-11-29 09:49:52