Screening molecular crystals with machine learning potentials

Autor: Ruža, Marko
Přispěvatelé: Lončarić, Ivor, Batistić, Ivo
Jazyk: chorvatština
Rok vydání: 2020
Předmět:
Popis: Teorija funkcionala gustoće (DFT) najraširenija je metoda za računanje svojstava materijala, no za sustave s velikim brojem atoma ili molekula izračuni postaju računalno zahtjevni. Primjer takvog sustava predstavljaju molekularni kristali čije se ćelije često sastoje od desetaka ili stotina atoma. Uzmimo u obzir situaciju u kojoj je na raspolaganju velik broj poznatih materijala. Tražimo li jedno specifično svojstvo koje bi samo dio tih materijala mogao imati, provedba eksperimenata za sve materijale bila bi dugotrajna i skupa, te na kraju i neisplativa. Praktično rješenje u takvim slučajevima su računalne simulacije, no već spomenuti DFT izračuni također su neisplativi za kompleksnije strukture. Stoga se za simulacije traže alternativne metode. U zadnje vrijeme strojno naučeni potencijali se pokazuju kao dobra alternativna metoda. Takvi računi neusporedivo su brži u odnosu na DFT, a mogu pokazivati sličnu točnost. S obzirom da kompleksnost tih potencijala jako brzo rase s brojem različitih elemenata, oni su primijenjivi na klase materijala sastavljenje od nekoliko istih elemenata. Konkretno u ovom diplomskom radu proučavani su organski kristali sastavljeni od atoma C, H, N i O. Time je omogućena obrada velikog broja organskih molekularnih kristala sastavljenih od spomenutih atoma u puno kraćem vremenu. Za te kristale napravljeni su računi elastičnih tenzora kako bi se identificirati oni kristali koji pokazuju neuobičajeno svojstvo negativne linearne kompresibilnosti (NLC). Za izabrane materijale predviđanja su provjerena DFT računom. Density functional theory (DFT) is the most widely used accurate method for predicting material properties, but its practicality is limited when used on systems with a large number of atoms or molecules. An example of such systems are molecular crystals as their unit cells usually have tens or hundreds of individual atoms. Let us consider a situation in which we have many known materials. If we want to identify certain materials which exhibit only one specific property, experimental research with such a goal often represents a technical problem. A large amount of experiments required to identify such materials can be time consuming and costly, in the end even unprofitable. Practical solution for such a problem are computer simulations, but DFT calculations of complex structures could also turn out to be impractical which is the main reason for developing alternative methods, like machine learning potentials. Lately, they have proven to be a good substitute. Calculations with machine learning potentials are much faster than DFT and their accuracy can be comparable. Certain potentials are constructed with a focus on a certain classes of materials, because complexity of such potentials grows with the number of different elements in a system. In this thesis, our focus is centered on organic crystals consisting only of C, H, N and O atoms. Such constraints make material property screening much faster and cost effective. Specifically, the main goal is to calculate elastic tensors for such crystals and identify those which could exhibit the property of negative linear compressibility (NLC). At the very end, several such results are verified by a DFT calculation.
Databáze: OpenAIRE
Externí odkaz: