Towards a generalised artificial neural network for sub-grid filtered density function closure in turbulent combustion

Abstract

An artificial neural network (ANN) is trained on moderate or intense low-oxygen dilution (MILD) combustion to predict the sub-grid filtered density function (FDF) in large eddy simulation (LES). For wide usability, a new quantity ψ is calculated by scaling the filtered mixture fraction using the stoichiometric value and a logarithmic function, which is used as an input feature of the ANN. Self-testing on MILD combustion is first conducted for validation and determining the best layout of ANN. An ANN with 4 hidden layers and the activation function of rectified linear unit (ReLU) has the highest accuracy. This ANN is subsequently tested on five different premixed combustion cases. Overall predictions of the progress variable FDF are satisfactory, and the filtered chemical source term modelled by the FDFs is in good agreement with the DNS data for each case. However, an underprediction of the filtered reaction rate occurs for the hydrogen–air flames, and the ANN accuracy is lower than that of a presumed β-FDF approach. DNS data of one premixed flame are then added to the training datasets, and newly trained ANNs are tested on remaining premixed flame cases. When the testing cases are similar to the added training dataset, improved predictions of marginal FDFs are observed, and the modelled progress variable source terms show good comparisons with the corresponding DNS data. Compared to the presumed β-FDF method, the new ANN shows an improved accuracy and serves as a viable alternative for FDF-based combustion modelling in the LES of turbulent flames.