Systematic method of applying ANN for chemical kinetics reduction in turbulent premixed combustion modeling

Abstract

A novel method to apply artificial neural network (ANN) for both chemical kinetics reduction and source term evaluation is introduced and tested in direct numerical simulation (DNS) and large eddy simulation (LES) of reactive flows. To gather turbulence affected flame data for ANN training, a new computation-economical method, called 1D pseudo-velocity disturbed flame (PVDF), is developed and used to generate thermo-chemical states independent of the modeled flame. Then a back-propagation ANN is trained using scaled conjugate gradient algorithm to memorize the sample states with reduced orders. The new method is employed in DNS and LES modeling of H2/air and C3H8/air premixed flames experiencing various levels of turbulence. The test result shows that compared to traditional computation with full mechanism and direct integration, this method can obtain quite large speed-ups with adequate prediction accuracy.