Turbulent premixed flame modeling using artificial neural networks based chemical kinetics

Abstract The applicability of Artificial Neural Networks (ANN) approach as a chemistry integrator for Large Eddy Simulations (LES) of reactive flows is evaluated with special emphasis on generating training tables independent of the computation of interest. An ANN code based on back-propagation algorithm is developed with a new approach for self-determining the model coefficients adaptively with respect to the error surface topology. The training table is constructed with an independent flame study, and the trained networks are used in LES of turbulent Flame–Vortex Interaction (FVI) studies at different equivalence ratios and turbulence levels. It is shown that once the ANN is well trained, it can successfully predict the reaction rates in both memory and time efficient manner compared to traditional look-up table approach and stiff ODE solvers, respectively.