Development of Artificial Neural Network Chemistry Framework for Turbulent Combustion of Non-premixed Flames

Abstract

Detailed chemical kinetics are required to accurately model pollutant formation in turbulent combustion simulations. However, existing methods for representing detailed turbulent combustion chemistry can have prohibitive computational cost. Machine learning (ML) methods are gaining attention for representing detailed chemistry in a computationally efficient manner, however the difficulties in preparing these models limit their application. In this work, artificial neural networks (ANNs) are used to represent non-premixed turbulent combustion chemistry in the Conditional Source-term Estimation (CSE) model, which accounts for detailed kinetics and turbulence-chemistry interactions (TCI). This work represents the first application of ML to CSE.

The objectives of the present study are to (i) develop a methodology for representing any chemistry dataset with ANNs in the context of the CSE combustion model, and (ii) apply ANN-CSE to the simulation of two turbulent non-premixed methane jet flames in a Reynolds-Averaged Navier-Stokes (RANS) framework.

The ANNs are developed using the MATLAB Deep Learning Toolbox. Two tabulated chemistry datasets are considered for ANN development: Trajectory-Generated Low-Dimensional Manifolds (TGLDM) and samples from direct integration (DI). In addition, pure and diluted methane fuels are considered. Detailed chemistry via GRI-Mech 3.0 and reduced chemistry via Smooke's mechanism is used. A data preparation procedure for ANN development is outlined. The sensitivity of various ANN parameters are also investigated to optimize the ANNs to each dataset. It is found that the ANNs can predict species mass fractions, reaction rates, temperature and heat release rate with good accuracy for each case. The storage requirement is also reduced by over 50\% for each case.

ANN-CSE is applied to turbulent non-premixed jet flames with both sets of data and either pure or diluted methane. The predictions of conditional and Favre averages of temperature, species mass fractions and source terms from ANN-CSE are compared to those from CSE with tabulated chemistry (TGLDM-CSE) or direct chemistry (DI-CSE) to verify ANN-CSE. Identical computational settings for each flame are used to verify the influence of the chemistry implementation on the flame structure. For each case, ANN-CSE is generally able to capture the trends of each conditional and Favre-averaged quantity at various locations. In addition, ANN-CSE requires less memory and is faster than regular CSE for each case.

This study shows that CSE can be effectively coupled with ANNs for chemistry representation of different mechanisms and sources of data. This will enable CSE for simulating more detailed cases. Future work may involve more complex fuels, more sophisticated ANNs and data preparation, or the Large Eddy Simulation (LES) framework.