Artificial neural networks based subgrid chemistry model for turbulent reactive flow simulations

Computational analysis of turbulent reactive flow applications requires resolution of the wide range of scales both in time and space from a flow modeling perspective. From a thermo-chemistry point of view, information regarding the radical chemical species is needed in order to capture flame-turbulence interactions accurately. A detailed investigation of all of these processes is time consuming. Thus, there is a need for speeding-up the computations by using the state-of-the art modeling capabilities. This study seeks to answer this problem and focuses in particular on the chemical kinetics calculations. The new approach proposed here is based on incorporating the artificial neural network (ANN) based modeling of the chemical kinetics into the large eddy simulation (LES) of reactive flows.;Two separate and new ANN based modeling approaches relevant to the LES are proposed within the thesis work. Here, the first approach depends on employing ANN to predict the species instantaneous reaction rates as a function of the thermochemical state vector ( w&d2;i = ANN(Yk, T)). The second one is based on using ANN specifically to predict the spatially filtered chemical source terms in the LES modeling as a function of the filtered thermo-chemical state vector and flow quantities ( w&d2;&d1; i = ANN(Y˜k, T˜, ReDelta, 6Y&d5;i 6x )).;First part of the thesis work dealt with testing different thermo-chemical tabulation techniques that can be used in connection with the ANN approach for the LES. Basically, three distinct methods (and tools) are developed here: thermo-chemical tables based on (i) laminar flames, (ii) laminar flame-vortex interactions (FVI) and (iii) laminar flame-turbulence interactions (FTI). Results based on premixed flame-vortex-turbulence interaction simulations showed that the tables generated based on the second and third approaches are capable of representing the actual thermo-chemical state-space accessed by the LES.;Once the tabulation procedure and the ANN training is achieved, ANN is first validated against direct numerical simulation (DNS) of a temporally evolving planar jet flame, which is reported to exhibit complex extinction and re-ignition type of physics. The flame physics are also investigated for this flame, and it is observed that the scalar dissipation rate has a weak sensitivity to extinction and re-ignition, than it is reported in the literature. As a second test case, comparisons against experimental and previous computational data are provided for a practical combustor simulation. Overall for all simulations considered here, the results indicated that the ANN works as a combustion mode independent model and provides fairly accurate results with considerable amount of speed-up and memory savings.