Previsão de viscosidade e temperatura liquidus de vidros óxidos via redes neurais artificiais

• Autor(a) principal: Bessa, Graziela Pentean
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/13306
• topic: Artificial neural networks; Glasses; Liquidus temperature; Viscosity; Prediction; Redes neurais artificiais; Vidros; Temperatura liquidus; Viscosidade; Previsão; ENGENHARIAS::ENGENHARIA DE MATERIAIS E METALURGICA

The number of glass compositions that have already been studied (~10^5) is many times smaller than the potential glass compositions (~10^52). The traditional method “cook and look” all of the potential compositions to identify each one has an interesting property is very time-consuming and also expensive. Another possibility is to use machine learning algorithms. This type of algorithm is capable of identifying patterns in large datasets, so they can be used to accelerate the study of glass compositions. There are several types of machine learning algorithms and artificial neural networks (also known as ANN) is one of them and have been used successfully for different types of problems and also used in this thesis. The algorithms used was written to optimize hyper-parameters to find the best topology for each neural network. Viscosity and liquidus temperature were chosen to be studied because they are important parameters for oxide glass production. The ANN was capable to predict liquidus temperature with R² = 0,997 and viscosity temperature T2 and T4 with R² = 0,999 and T3 with R² = 0,998. The mean relative errors for the viscosity ANN are equal to 2,5 at maximum, while for liquidus temperature mean relative error is equal to 3,7%. These errors are considered small and at the same order of magnitude of the scattering of dataset. It is possible to see some problems on the predictions like eutetic regions in phase diagrams for liquidus temperature. Even though the errors, the neural networks were capable of predicting both properties. For future works it is possible to select other properties of atoms to form the dataset besides the composition and the property of interest.