Primjene metoda dubokog učenja u kompleksnim sustavima

master's thesis

• Title (english): Applications of deep learning methods in complex systems
• Author: Filip Mirković
• Mentor: Davor Horvatić (mentor)
• Committee member: Davor Horvatić (predsjednik povjerenstva)
• Committee member: Matko Glunčić (član povjerenstva)
• Committee member: Andrej Novak (član povjerenstva)
• Committee member: Nikola Poljak (član povjerenstva)
• Committee member: Matko Milin (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2024-02-22, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

U ovom radu bavimo se modeliranjem kompleksnih sustava korištenjem neuronskih mreža nad grafovima. Posebnu pažnju posvećujemo ugradnji simetrija fizikalnog sustava u model, koje su presudne za kvalitetno funkcioniranje modela. Koristimo O(3)- ekvivarijatnu inačicu transformera nad grafovima za predviđanje međuatomskih sila unutar malih organskih molekula: benzena, uracila, etanola i aspirina. Provjeravamo generalizacijska svojstva modela pri predviđanju sila na atome unutar molekule paracetamola. Na kraju rada, istražujemo robusnost modela u režimima malog broja podataka. Taj režim je ostvaren treniranjem modela na deset puta manjem skupu podataka od inicijalnog skupa podataka za trening.

Abstract (english)

In this work, we examine the use of graph neural networks in modelling complex systems. We also emphasise the use of built-in symmetries in our model. We demonstrate these ideas on the problem of molecular dynamics. A version of an O(3)- equivariant graph attention network is used to predict interatomic forces in small organic molecules: benzene, ethanol, uracil, and aspirin. Furthermore, we test the model’s generalization capabilities on various configurations of paracetamol. Lastly, we demonstrate the power of equivariance in low-data regimes. This regime is achieved by training the model with a ten times smaller dataset than the initial training set.

• Keywords: Kompleksni sustavi; Molekularna dinamika; Neuronske mreže nad grafovima; Ekvivarijantnost
• Keywords (english): Complex systems; Molecular dynamics; Graph neural networks; Equivariance
• Language: croatian
• URN:NBN: urn:nbn:hr:217:027408
• 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: Embargoed access; Embargo expiration date: 2025-02-22
• Created on: 2024-04-05 12:39:19