Modelling of turbulent combustion using the Rate-Controlled Constrained Equilibrium (RCCE) - Arti cial Neural Networks (ANNs) approach

Abstract

The objective of this work is the formulation, development and implementation of Artificial Neural Networks (ANNs) to turbulent combustion problems, for the representation of reduced chemical kinetics. Although ANNs are general and robust tools for simulating dynamical systems within reasonable computational times, their employment in combustion has been limited. In previous studies, ANNs were trained with data collected from either the test case of interest or from a similar problem. To overcome this training drawback, in this work, ANNs are trained with samples generated from an abstract problem; the laminar flamelet equation, allowing the simulation of a wide range of problems. To achieve this, the first step is to reduce a detailed chemical mechanism to a manageable number of variables. This task is performed by the Rate-Controlled Constrained Equilibrium (RCCE) reduction method. The training data sets consist of the composition of points with random mixture fraction, recorded from flamelets with random strain rates. The training, testing and simulation of the ANNs is carried out via the Self-Organising Map - Multilayer Perceptrons (SOM-MLPs) approach. The SOM-MLPs combination takes advantage of a reference map and splits the chemical space into domains of chemical similarity, allowing the employment of a separate MLP for each sub-domain. The RCCE-ANNs tabulation is used to replace conventional chemistry integration methods in RANS computations and LES of real turbulent flames. In the context of RANS the interaction of turbulence and combustion is described by using a PDF method utilising stochastic Lagrangian particles. In LES the sub-grid PDF is represented by an ensemble of Eulerian stochastic fields. Test cases include non-premixed and partially premixed turbulent flames in both non-piloted and piloted burner configurations. The comparison between RCCE-ANNs, real-time RCCE and experimental measurements shows good overall agreement in reproducing the overall flame structure and a significant speed-up of CPU time by the RCCE-ANN method.