Aplicação de redes neurais do tipo função de base radial simples com poucos neurônios na previsão da temperatura de transição vítrea de boratos de metais alcalinos

• Ano de defesa: 2022
• Autor(a) principal: Vitória, Leonardo dos Santos
• Orientador(a): Lalic, Susana de Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Programa de Pós-Graduação: Pós-Graduação em Física
• Palavras-chave em Português: Vidro; Temperatura de transição vítrea; Alcalino borato; Função de base radial; Aprendizado de máquina; Inteligência artificial
• Palavras-chave em Inglês: Glass; Glass transition temperature; Alkaline borate; Radial base function; Machine learning; Artificial intelligence
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::FISICA
• Link de acesso: https://ri.ufs.br/jspui/handle/riufs/15460

Interest in glass applications and uses has existed since the beginning of civilization. Such materials are appreciated for several functions and properties given the extensive possibility of making different compositions. It is estimated that there are about 1052 glasses compositions using 80 elements of the periodic table considering minimum concentrations up to 1%, and only 105 have been explored, which allows many discoveries. In view of the many possibilities associated, the objective of this work was to define properties in terms of compositions, and among them, the characteristic glass transition temperature. Artificial intelligence proposed by Alan Turing in 1950, which today branched out into computational processes known as machine and deep learnings, as well as artificial neural networks, that are considered highly predictive tools and can be used in the description of several physical systems. Among different types of artificial neural networks, stands out radial basis functions. This neural network is characterized by performing an effective and fast training due to its particular learning mechanism, able to transform complex systems into a simple linear algebra problem. Complex neural networks can be created to describe different phenomena; however, the more complex the network, more overfitted to the data it will tend to be, and a way to avoid this can be via the use of a certain number of neurons in its construction. Given these aspects, radial basis functions networks with only two neurons (or poles) were applied to describe the glass transition temperature of alkali metal borate systems. The results show that the technique it’s highly predictive, as data reached R² adjustment value over 90%. It was also possible to carry out adjustments and train the network including the known phenomenon of the boron anomaly. The Gaussian activation function proved to be superior to two others, named multiquadratic functions. The neuron location in the networks was a highlight, as the tests showed an improvement in the performance of the adjustments by up to 5% when manipulated.The Tg behavior of alkaline borate systems as a function of molar concentration x of oxides is similar and corresponds to the topological model by Mauro, Gupta and Loucks (2009), established only for lithium and sodium borate systems. Such work suggests extending this study to other physical properties using a few neurons and associated with the vitreous state.