Reduced order model of diffusion flames based on multi-scale data from detailed CFD: the impact of preprocessing

• Published in: Journal of the Brazilian Society of Mechanical Sciences and Engineering 2024-04, Vol.46 (4), Article 215
• Main Authors: Lopes Junqueira, Nicole, da Costa Ramos, Louise, Figueira da Silva, Luís Fernando
• Format: Article
• Language: English
• Subjects: Algorithms; Complex systems; Computational fluid dynamics; Diffusion flames; Engineering; Engineering Sciences; Flames; Fuels; Logarithms; Machine learning; Mathematical models; Mechanical Engineering; Model accuracy; Preprocessing; Reactive fluid environment; Reduced order models; Singular value decomposition; Technical Paper
• ISSN: 1678-5878; 1806-3691
• DOI: 10.1007/s40430-024-04749-6

Machine learning techniques, such as reduced order models (ROM), have demonstrated low cost when creating models of complex systems while aiming at the same accuracy as high-fidelity models, such as computational fluid dynamics (CFD). However, reduced models must preserve some properties tied to the studied system. For a combustion problem, those are in particular monotonicity, positivity, and boundedness. Here, ROM are created using data from CFD simulations of non-premixed laminar flames with detailed chemistry and transport. The data obtained for variable fuel velocity is reduced using singular value decomposition (SVD), and then a genetic aggregation response surface algorithm is applied to predict the properties fields for an arbitrary fuel inlet velocity. This work analyzes the effect of different data preprocessing approaches on the ROM, i.e., (1) the properties treated as an uncoupled or as a coupled system; (2) normalization of different properties; (3) the logarithm of the chemical species. For all constructed ROM, the energy content of the reduction process and the reconstructed fields of the flame properties evidence the slow convergence of SVD modes for the uncoupled ROM, while a faster one is seen when the logarithm preprocessing is applied. Also, the learning is shown to be achieved with a smaller number of modes for two of the coupled ROM and for the ROM using the logarithm. The reconstruction of the mass fraction fields is characterized by regions of negative values, underscoring that the baseline ROM methodology does not preserve the properties of monotonicity, positivity, and boundedness. The proposed logarithm preprocessing enables to overcome the positivity problem and to accurately reproduce the original data.