Large Eddy Simulation of the Premixed Flame-Turbulence Interactions using Artificial Neural Network in Chemical Kinetics Tabulation

A large eddy simulation (LES) of premixed flame-turbulence interaction is performed with special emphasis on computing the instantaneous chemical species reaction rates with the recently developed approach of artificial neural networks (ANNs) for chemical kinetics. Training of the neural network is based on an independent flame study using linear eddy mixing technique. An analysis of computational performance, considering CPU time and a comparison between the performance of artificial neural network technique and other conventional methods is used to represent the chemical kinetics such as direct integration (DI) - and the ability of neural networks to model the highly non-linear and stiff chemistry ODEs is illustrated. The sub-grid combustion model of the LES is based on a linear eddy mixing model while a skeletal multi-species, multi-step chemical kinetic mechanism is applied for the combustion. A feed-forward, multi-layer architecture is chosen for the neural network and the training algorithm is based on a back-propagation gradient descent rule with adaptive learning rate and individual momentum factors for the weight coefficients. The flow field distribution and the flame characteristics obtained by LES with neural network based chemical kinetics tabulation, are in reasonable agreement with previous direct numerical simulation (DNS) study of the flame. The results show if the neural network is trained accurately, it can predict the instantaneous chemical species reaction rates in LES framework.