Logo do repositório

Métodos de machine learning aplicados em estruturas de água e gelo

Carregando...
Imagem de Miniatura

Orientador

Rocha, Alexandre Reily

Coorientador

Pós-graduação

Física - IFT

Curso de graduação

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Estadual Paulista (Unesp)

Tipo

Tese de doutorado

Direito de acesso

Acesso abertoAcesso Aberto

Resumo

Resumo (português)

Neste trabalho, utilizamos potenciais obtidos por meio de machine learning para investigar a influência da inclusão de efeitos quânticos nucleares nas interações interatômicas, bem como na formação e difusão de defeitos na estrutura cristalina do gelo Ih. Os potenciais foram gerados a partir de energias e forças obtidas em cálculos baseados na teoria do funcional da densidade, empregando diferentes funcionais de correlação e troca: PBE, vdW-cx, SCAN e optB88. Inicialmente, apresentamos a estratégia para a construção de um modelo confiável capaz de reproduzir com precisão as propriedades de sistemas aquosos. Verificamos que um modelo treinado a partir de um sistema contendo 64 moléculas de água é suficiente para reproduzir suas propriedades estruturais na fase líquida. Além disso, para redes neurais de grafos, observamos que a quantidade de configurações utilizadas no treinamento pode ser reduzida para apenas 6% do dataset original, sem prejuízo das propriedades estruturais e dinâmicas obtidas com o conjunto completo de dados. Cálculos de dinâmica molecular realizados com potenciais obtidos por machine learning (DP-MD) mostram que a densidade do gelo Ih está em bom acordo com os resultados obtidos por dinâmica molecular ab initio. Ademais, a inclusão de efeitos quânticos nucleares leva a um leve aumento no valor médio do comprimento da ligação covalente O–H, promovendo um maior compartilhamento de prótons. Simultaneamente, esse efeito compete com a redução das distâncias intermoleculares decorrente do fortalecimento das ligações de hidrogênio. Como consequência, observa-se um aumento geral da densidade de equilíbrio para todos os funcionais, afastando os resultados dos valores experimentais. No estudo dos defeitos, verificamos que a inclusão de termos não locais reduz a energia de formação dos defeitos intersticiais, alterando o tipo de defeito predominante em todo o intervalo de temperatura analisado. Determinamos também a energia necessária para que um defeito migre de um sítio para outro. Observamos que defeitos intersticiais apresentam barreiras de migração menores que as dos defeitos de vacância. Em contrapartida, a frequência tentativa de migração dos defeitos de vacância é superior à dos intersticiais. Como resultado, os coeficientes de difusão dos diferentes defeitos assumem valores semelhantes. Além disso, constatamos que os defeitos intersticiais apresentam maior tendência a se difundir na direção perpendicular ao eixo c do que na direção paralela. Por fim, mostramos que o aumento da temperatura afeta de forma distinta a energia de ligação: intensifica a interação entre a molécula intersticial e o cristal, mas reduz a intensidade das ligações das moléculas no interior da rede cristalina.

Resumo (inglês)

In this work, we employ machine learning potentials to investigate the influence of nuclear quantum effects on interatomic interactions, as well as on the formation and diffusion of defects in the crystalline structure of ice Ih. The potentials were generated from energies and forces obtained through density functional theory calculations using different exchange–correlation functionals: PBE, vdW-cx, SCAN, and optB88. Initially, we present the strategy for constructing a reliable model capable of accurately reproducing the properties of aqueous systems. We found that a model trained on a system containing 64 water molecules is sufficient to reproduce the structural properties of liquid water. Furthermore, for graph neural networks, we observed that the number of configurations used for training can be reduced to only 6% of the original dataset without compromising the structural and dynamical properties obtained with the full dataset. Molecular dynamics simulations performed with machine learning potentials (DP-MD) show that the density of ice Ih is in good agreement with results from ab initio molecular dynamics. Moreover, the inclusion of nuclear quantum effects leads to a slight increase in the average O–H covalent bond length, enhancing proton sharing. At the same time, this effect competes with the reduction of intermolecular distances resulting from the strengthening of hydrogen bonds. Consequently, there is an overall increase in the equilibrium density for all functionals, further deviating the results from experimental values. In the study of defects, we found that the inclusion of nonlocal terms lowers the formation energy of interstitial defects, changing the predominant defect type over the entire temperature range analyzed. We also determined the energy required for a defect to migrate from one site to another. Our results show that interstitial defects exhibit lower migration barriers compared to vacancy defects. In contrast, the attempt frequency for vacancy defect migration is higher than that for interstitial defects. As a result, the diffusion coefficients of the different defects display similar values. Additionally, we observed that interstitial defects have a greater tendency to diffuse in the direction perpendicular to the c axis than parallel to it. Finally, we show that increasing temperature affects the binding energy in different ways: it strengthens the interaction between the interstitial molecule and the crystal, while simultaneously weakening the bonding between molecules in the interior of the crystalline lattice.

Descrição

Palavras-chave

Dinâmica Molecular, Gelo, Aprendizado do computador

Idioma

Português

Citação

MIRANDA, Lucas Thiago Siqueira de. Métodos de machine learning aplicados a estruturas de água e gelo. 2026. Tese (Doutorado em Física) - Instituto de Física Teórica, Universidade Estadual Paulista (Unesp), São Paulo, 2025

Itens relacionados

Unidades

Departamentos

Cursos de graduação

Programas de pós-graduação