The influence of the learning data on the reduced order model of laminar non-premixed flames

Abstract

Computational fluid dynamics (CFD) is often applied to the study of combustion, enabling to optimize the process and control the emission of pollutants. This numerical methodology enables the analysis of different flame properties, such as the components of velocity, temperature, and mass fractions of chemical species. However, reproducing the behavior observed in engineering problems requires a high computational cost associated with memory and simulation time. Reduced order model (ROM) is a machine learning technique that has been applied to several engineering applications, aiming to develop models for complex systems with reduced computational cost. In this way, a high-fidelity model of complex systems is created from available data to learn its behavior and its main characteristics. In this work, different ROMs are created using CFD simulation data. The CFD model solves the mass, species, energy, and momentum conservation equations for a methane/air laminar diffusion flame, stabilized on the Gülder burner. Chemistry is modeled using a 19-species skeletal chemical kinetic mechanism. The static reduced order model uses the singular value decomposition (SVD) algorithm to decompose the CFD data and obtain the system's modes. Then, genetic aggregation response surface interpolation is applied on the higher SVD modes, creating the static ROM. This work analyzes the effect of different data preprocessing approaches on the ROM. The first analysis is the impact of reducing the number of learning data points, showing that this decrease does not directly impact the energy of the SVD modes, but, in the reconstruction field is possible to notice a degradation of the reconstruction. The second analysis is related to the effect of creating a ROM for each uncoupled flame property or treating the properties as a coupled system. The results of the coupled and uncoupled reduced order models are quite similar in terms of properties field reconstruction. However, in the energy analysis the coupled ROM converges rapidly, similarly to the uncoupled temperature ROM, while the uncoupled chemical species ROMs have a slower convergence.

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

“…In this work, a ROM of a laminar diffusion flame based on several flame properties obtained through CFD is created, since difficulties linked to the learning such different flame scales have been seen on previous studies. For instance, an issue related to the monotonicity of the predicted combustion properties has been observed [28]. To overcome such monotonicity problem [28], different methodologies to preprocess the available data are proposed.…”

“…For instance, an issue related to the monotonicity of the predicted combustion properties has been observed [28]. To overcome such monotonicity problem [28], different methodologies to preprocess the available data are proposed. Firstly, a simple methodology regarding the properties being treated as an uncoupled or as a coupled system in the construction of the ROM is proposed.…”

“…Figure 1b presents an example of an ethylene/air laminar diffusion flame stabilized on a Gülder burner [2]. Based on the burner characteristics [28], one can define the non-premixed flame experimental setup and geometry and boundary conditions, as presented on the following section, which are necessary to the development of the combustion modeling of interest.…”

“…Aiming to overcome the monotonicity problem seen in [28], the logarithm is applied to the learning data is applied, as given by Eq. (2).…”

“…Furthermore, to ensure that algorithms learn with prior knowledge about the underlying physics of a problem or its restrictions, it has been proposed to combine ML with first principles models of engineering system, e.g., the physics informed machine learning used to predict the critical heat flux, with superior performance over standalone approaches [16][17][18]. Here, the focus is on using data preprocessing to solve the monotonicity problem found in ROM applied to laminar flames using known combustion principles, such as the exponential decay of chemical species with distance [27][28][29][30].…”