Tese: Assessment of reduced order models applied to steady-state bi-dimensional laminar methane air diffusion flame
Aluno(a) : Nicole Lopes Monteiro de Barros JunqueiraOrientador(a): Igor de Paula , Luís Fernando Silva e Louise Ramos
Área de Concentração: Termociências
Data: 17/02/2022
Link para tese/dissertação: https://doi.org/10.17771/PUCRio.acad.58795
Resumo: Computational fluid dynamics (CFD) is often applied to the study of combustion, enabling to optimize the process and control the emission of pollutants. However, reproducing the behavior observed in engineering systems has a high computational burden. To overcome this cost, machine learning techniques, such as reduced order models (ROM), have been applied to several engineering applications aiming to create models for complex systems with reduced computational cost. Here, the ROM is created using CFD laminar non premixed flame simulation data, decomposing it, and then applying a machine learning algorithm, creating a static ROM. This work analyzes the effect of five different data pre-processing approaches on the ROM, these being: (1) the properties treated as an uncoupled system or as a coupled system, (2) without normalization, (3) with temperature and velocity normalized, (4) all properties normalized, and (5) the logarithm of the chemical species. For all ROM constructed are analyzed the energy of the reduction process and the reconstruction of the flame properties fields. Regarding the reduction energy analysis, the coupled ROM, except the ROM (4), and the logarithm ROM converges faster, similarly to the uncoupled temperature ROM, whereas the uncoupled minor chemical species ROM exhibits a slower convergence, as does the coupled ROM with all properties normalized. So, the learning is achieved with a smaller number of modes for the ROM (2), (3) and (5). As for the reconstruction of the property fields, it is noted that there are regions of negative mass fraction, which suggest that the ROM methodology does not preserve the monocity or the boundedness of the properties. The logarithm approach shows that these problems are overcome and reproduce the original data.