LES using artificial neural networks for chemistry representation

In this work, a large-eddy simulation (LES) was performed using artificial neural networks (ANN) for chemistry representation. The case of Flame D, a turbulent non-premixed piloted methane/air flame, was chosen to validate this new strategy. A second LES utilising a classical structured chemistry table for a steady flamelet model was used for comparison. A Smagorinsky model applying the dynamic procedure by Germano to determine the Smagorinsky parameter was used for the subgrid stresses. It is shown that the new procedure yields approximately three orders of magnitude lower memory requirements, while the required CPU time for the application of the networks increases only little. The results obtained from the two simulations do not differ significantly. Furthermore, the smooth approximation of the chemistry table with the neural networks stabilises the LES of turbulent reactive flows and allows the application of advanced chemistry models with higher dimensionality.