An integrated PDF/neural network approach for simulating turbulent reacting systems

In this paper, we introduce an integrated PDF/neutral network approach for the simulation of turbulent flames. The use of artificial neural network (ANN) to represent chemical reaction in turbulent flames offers a significant reduction in the computer memory and run time demands over the classical methods, namely, look-up tables and direct integration. The adequacy of the neural network model strongly depends on the selection of the training set, which should be representative of the most accessible composition space. This is essential for the network to give an accurate model of the chemistry. A novel method known as statistical mapping is used to generate a training set for the neural network. This is a small-scale presimulation of the flame to produce a set of samples that are representative of the most accessible compositions during actual flame computations. In case of difficulties in achieving convergence of the network, “histogram redistribution” technique is used to smooth the PDF of the input samples. This technique is found to improve the convergence of network; however, its generality is yet to be determined. The integrated PDF/ANN approach is demonstrated here for a piloted flame with simple chemistry.