Machine learning methodology for thermochemistry in turbulent combustion

Abstract

Many turbulent combustion simulation methods require the direct numeric integration of ordinary differential equations (ODEs) to compute the chemical source term. To obtain accurate simulation results, a complex chemical mechanism with multiple species and a large number of reactions is mandatory, which leads to a large and usually stiff set of ODEs. The direct integration process using such mechanisms is highly time-demanding and has become the most expensive element of a simulation, which hampers the large-scale industrial applications of these simulation methods. In the present work, a new machine learning methodology based on artificial neural networks (ANNs) is proposed for speeding up thermochemistry computations in turbulent combustion simulations. ANNs have been applied for chemistry tabulation in many works, but their employment in complex combustion problems is still limited mainly because of two factors: generalisation and accuracy. The ANN methodology in this work is developed to tackle these two issues. Regarding the first issue, a hybrid flamelet/random data (HFRD) method is proposed for generating the training dataset. The random element can endow the resulting ANNs with increased capacity for generalisation. Regarding the issue of accuracy, a multiple multilayer perceptron (MMLP) approach is developed where several multilayer perceptrons are trained for each species to predict composition states with different magnitudes of concentration changes or initial concentrations, and it is shown that this method can significantly reduce the prediction errors. The proposed approach is used to simulate a wide range of combustion problems, including 1-D laminar flames and different types of turbulent diffusion flames. The methodology is also applied to different fuels/mechanisms, including methane, methane/hydrogen blended fuel, and dimethyl ether. The range of problems simulated indicates that the approach has a great capacity for generalisation. The excellent agreement between ANN simulation results and direct integration simulation results for both major species and minor species fully manifests the high accuracy feature of the methodology. Finally, great savings in computational costs are observed. A speed-up ratio of more than 10 was attained for the reaction step. Such time reduction means that real-time thermochemistry computation is longer an obstacle to turbulent combustion simulations.