Application of an artificial neural network to sub-filter density function estimation for premixed combustion

Machine learning (ML) is emerging as a well-suited and robust approach for modelling turbulent flames, but its powerful performance is limited by the scope of training cases. The present work aims to develop one ML approach, namely Artificial Neural Network (ANN), to resolve the lack of generalization while maintaining good accuracy. An ANN is trained using a comprehensive direct numerical simulation (DNS) dataset of Moderate or Intense Low-oxygen Dilution (MILD) combustion. A lot of premixed combustion cases, including planar flame, V-flame and swirl flame, with different fuel (hydrogen and methane) have been tested. It is observed the prediction of the Filtered Density Function (FDF) is satisfactory. This observation is further assessed by conditionally averaging FDF on the first and second moments of progress variable. Additional assessment is carried out through comparisons with the reaction rate using DNS data and the values obtained by using a presumed PDF method. It is detected that results from ANN and the presumed PDF approach are comparable, which illustrates the feasibility of replacing traditional look-up table with ANN in Large Eddy Simulation (LES) for cases distinct from the training cases.