Application of Artificial Neural Networks for the Simulation of a Perfectly Premixed Chemical Reactor

Abstract

Finite-rate chemistry combustion simulations require the computationally expensive direct integration of the chemical source term. In order to reduce the calculation time of such simulations, an efficient artificial neural network approach is developed. The artificial neural network consists of a classification methodology to subdivide the thermochemical state space into smaller clusters and a subsequent multi-layer perceptron for each of the clusters. Due to the subdivision approach, the multi-layer perceptron size can be kept small, which guarantees a computationally efficient artificial neural network. For the classification two different approaches are investigated, namely a self-organizing map and a k-means binary decision tree. Both methods are tested with respect to their clustering quality and their performance. Chemical reactors are used to generate training and validation data. The multi-layer perceptron prediction includes all species of the original chemical mechanism. A priori comparison with direct integration proved the ability of both methods to give accurate results even for minor species. A posteriori calculations of ignition delay are conducted over an initial temperature range of 1100 to 1700 K and an equivalence ratio range of φ = 0.7 to φ = 1.4. While species concentrations and temperature profiles are reproduced well for most of the initial conditions, the prediction quality of the artificial neural networks decreases for a few calculations starting at low temperatures. Performance benchmarks confirmed that the artificial neural network approaches are superior to direct source term integration in terms of computational costs. The benchmarks also revealed that the k-means binary decision tree-based approach is three times faster than the self-organizing map approach.