Machine learning in solid-state physics and statistical physics

master's thesis

• Title (croatian): Strojno učenje u fizici čvrstog stanja i statističkoj fizici
• Author: Lovro Vrček
• Mentor: Vinko Zlatić (mentor)
• Mentor: Ivo Batistić (mentor)
• Committee member: Ivo Batistić (predsjednik povjerenstva)
• Committee member: Vinko Zlatić (član povjerenstva)
• Committee member: Damir Pajić (član povjerenstva)
• Committee member: Davor Horvatić (član povjerenstva)
• Committee member: Emil Tafra (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2018-07-05, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

In this work, we study an alternative approach to simulating molecular dynamics of large systems over long time periods; the one using machine learning instead of DFT. Studied system is a ruthenium surface which is interacting with hydrogen atoms. We use neural networks to obtain the regression model for the studied system, and train several different architectures of neural networks in order to find the optimal one. For input we use Gaussian descriptors which take Cartesian coordinates of the atoms and translate them in a more suitable form. In this case, there are 20 descriptors for each type of atoms, meaning that input layer of neural network has in total 40 nodes. Optimal architecture was found to be the one with three hidden layers with 50, 30, and 10 nodes, respectively. It was shown how our regression model behaves depending on number of training steps, importance of used descriptors was analyzed, and it was shown how model behaves if Zernike descriptors are used instead of Gaussian, or if cutoff radius is altered.

Abstract (croatian)

U ovom radu proučavamo alternativni pristup simuliranju molekularne dinamike velikih sustava preko dugih vremenskih perioda; korištenje strojnog učenja umjesto DFT-a. Proučavani sustav sastoji se od površine rutenija koja međudjeluje s atomima vodika. Za konstruiranje regresijskog modela koristimo neuronske mreže te treniramo nekoliko različitih arhitektura mreža kako bismo našli optimalnu. Kao ulaz koristimo Gaussove deskriptore koji kartezijeve koordinate atoma pretvore u oblik pogodniji za opis sustava. U ovom slučaju, postoji 20 deskriptora za svaku vrstu atoma, što znači da u ulaznom sloju neuronske mreže imamo 40 čvorova. Pokazano je da optimalna arhitektura neuronske mreže sadrži tri skrivena sloja, od kojih prvi ima 50 čvorova, drugi 30, a treći 10. Osim toga, pokazano je kako se taj regresijski model ponaša ovisno o broju koraka prilikom treniranja, analizirana je važnost korištenih deskriptora te je proučeno ponašanje modela u slučaju korištenja Zernike deskriptora umjesto Gaussovih, ili mijenjanja polumjera obuhvaćanja atoma.

• Keywords: DFT; molecular dynamics; neural networks; descriptors
• Keywords (croatian): DFT; molekularna dinamika; neuronske mreže; deskriptori
• Language: english
• URN:NBN: urn:nbn:hr:217:394430
• 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: 2018-07-13 12:58:00