Influência do tratamento da base de dados DNS na aplicação de técnicas de aprendizagem de máquina para melhorar acurácia de simulações RANS

• Ano de defesa: 2019
• Autor(a) principal: Rangel, Victor Bitencour
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Mecânica; UFRJ
• Palavras-chave em Português: Turbulência; Aprendizado de máquina; Simulação numérica; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA
• Link de acesso: http://hdl.handle.net/11422/13776

Among numerical solution techniques for turbulent flow, Reynolds Average Navier-Stokes (RANS) presents low computational and time costs, but also lower accuracy. Large Eddy Simulation (LES) and Direct Numerical Simulation (DNS) present greater accuracy, but both require much more computational effort and simulation time, making them impracticable in daily industry problem solution. The utilization of machine learning (ML) to help improve RANS results is receiving some attention lately. Such approach makes RANS simulations achieve better results by learning them from more noble ones, as DNS simulation for example. The main entity applied in RANS correction is the Reynolds stress tensor (R). Although, recent studies show that R does not present well converged results from available DNS data base when compared to other fields convergence. Another entity is introduced in ML techniques in order to bypass uncertainties generated by the correction of R, it is the modified divergence of R which is called here as vector t. The correction of t, when compared to R methodology, has shown promising results, even presenting better results than the ones achieved by correcting R. The present work studies how DNS data base quality influences on neural network (NN) prediction for both aforementioned methodologies: correction of R and t, for square duct turbulent flow. DNS data are manipulated in order to simulate better convergence. Results show that better converged DNS data leads to better NN predictions, which has presented its best results when total data symmetry was imposed for both correction methodologies. It was also observed that correction of t presents lower global errors for the main direction flow when compared to R correction, but not for the secondary direction flows, which puts such methodology in doubt about its advantages in relation to R correction.