Strojno učenje svojstava materijala na podacima različite točnosti

master's thesis

• Title (english): Multi-fidelity machine learning of material properties
• Author: Fran Krčelić
• Mentor: Ivor Lončarić (mentor)
• Mentor: Nikola Poljak (komentor)
• Committee member: Nikola Poljak (predsjednik povjerenstva)
• Committee member: Ivor Lončarić (član povjerenstva)
• Committee member: Davor Horvatić (član povjerenstva)
• Committee member: Danko Radić (član povjerenstva)
• Committee member: Maro Cvitan (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2024-10-21, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

Ovaj rad istražuje primjenu ekvivarijantnih graf neuralnih mreža s nasumičnim Fourierovim značajkama u predikciji energije molekula. Kombinirajući dva seta podataka različite točnosti, ANI-1x i ANI-1ccx, sastavljenih od energija organskih molekula baziranih na H, C, N, O elementima, model je usmjeren na poboljšanje prediktivne moći kroz dublje razumijevanje međuatomske interakcije. ANI-1x sadrži podatke generirane teorijom funkcionala gustoće za molekule izabrane aktivnim učenjem, omogućujući... More bolju generalizaciju s manjim brojem podataka, dok ANI-1ccx uključuje visoko precizne energije računane teorijom spregnutih grozdova. Treniranjem s različitim omjerima ovih setova pokazano je da povećanje količine ANI-1x podataka u trening setu rezultira boljom prediktivnom moći modela. Također, diskutirana je primjena ovoga pristupa na složenije sustave kao što su molekularni kristali. Ovakva metoda otvara nove mogućnosti za preciznije modeliranje u fizici, kemiji i znanosti o materijalima. Less

Abstract (english)

This thesis explores the application of equivariant graph neural networks with random Fourier features in predicting molecular energy. By combining two datasets of different accuracies, ANI-1x and ANI-1ccx, consisting of energies of organic molecules based on H, C, N, and O elements, the model aims to enhance predictive power through a deeper understanding of interatomic interactions. ANI-1x contains data generated with density functional theory for molecules selected through active... More learning, which allows for better generalization with fewer data points, while ANI-1ccx provides highly accurate energy values calculated using coupled-cluster methods. Training with various ratios of these datasets demonstrated that increasing the amount of ANI-1x data in the training set results in improved predictive power of the model. Additionally, the application of this approach to more complex systems, such as molecular crystals, was discussed. This method opens new possibilities for more accurate modeling in physics, chemistry and materials science. Less

• Keywords: neuralne mreže; ekvivarijantne graf neuralne mreže; nasumične Fourierove značajke; molekularna energija
• Keywords (english): neural networks; equivariant graph neural networks; random Fourier features; molecular energy
• Language: croatian
• URN:NBN: urn:nbn:hr:217:460693
• 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: 2025-02-01 09:18:28