NAZCA: a machine learning based framework for performance prediction and configuration recommendation of multiscale numerical simulations

• Ano de defesa: 2022
• Autor(a) principal: Fabian, Juan Humberto Leonardo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Laboratório Nacional de Computação Científica; Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA); Brasil; LNCC; Programa de Pós-Graduação em Modelagem Computacional
• Palavras-chave em Português: Aprendizado por computador; Análise numérica; Computação de alto desempenho; Métodos dos elementos finitos; Modelagem multiescala; CNPQ::CIENCIAS EXATAS E DA TERRA::MATEMATICA::ANALISE
• Link de acesso: https://tede.lncc.br/handle/tede/368

Multiscale phenomena are observed in nature, which has increasingly attracted the attention of researchers from different areas. Simulations that intend to represent such phenomena should use computationally robust methods. The use of these multiscale methods is often limited to a small group of users who understand the inherent complexity of each method. The overarching objective of this thesis is to help users unfamiliar with the complexity of these methods to configure and run multiscale simulations. With this objective in mind, we propose a methodology called NAZCA. This methodology is based on machine learning, thereby using a dataset from previous simulations to help users. We have defined several scenarios in which help is needed for users of multiscale simulations. Each of these scenarios is associated with a task, which will be solved with a specific machine learning technique. In this thesis, we consider the multiscale hybrid mixed (MHM) finite element method as our case study. We have used NAZCA to analyze three tasks from three different scenarios involving the MHM method. In the first task, we estimate the execution time and numerical error of an MHM simulation. For this task, we developed a specific tree-based learning technique that explores specific knowledge about the numerical method. We show that this technique obtains smaller errors than other state-of-the-art techniques, and a high level of interpretability. In the second task, we recommend numerical parameters for an MHM simulation, based on the execution time and numerical error targeted by the user. For this task, we employed a distance-based technique on a performance metric space that is formed by predictions obtained by the first task. We show that this technique obtains an accuracy of more than 80% in determining the numerical parameters that present execution time and numerical error closest to the user’s target. In the third task, we recommend computational configurations for an MHM simulation, based on the numerical parameters desired by the user. For this task, we employed a ranking-based technique that orders the available computational configurations according to the estimated time to run a numerical simulation on them. We show that this technique obtains configuration rankings that approximate the actual ranking obtained with the dataset of previous simulations. We expect it will be possible to use this methodology with other numerical methods with similar computational characteristics.