Large eddy simulation of turbulent premixed combustion with a data-driven filtered density function model

Recent data-driven approaches have demonstrated promising predictive capabilities in many a priori assessments. However, their performance evaluation in a posteriori large eddy simulation (LES) assessments is limited. This study implements an artificial neural network (ANN)-based sub-grid filtered density function (FDF) model in LES of turbulent premixed combustion. The ANN model is trained on DNS data from moderate or intense low-oxygen dilution (MILD) combustion and a premixed twin-V flame, predicting sub-grid marginal FDFs of the progress variable to model the filter reaction rate through on-the-fly inference. Simulations with the ANN models are 2 to 3 times longer than those for the presumed FDF-based tabulation method because of its complex architecture and reduced parallel scalability. However, its over 300-fold lower memory usage offsets the computational expenses. The ANN model is evaluated using two cases, a bluff-body stabilised methane–air flame and the well-known Volvo afterburner propane–air flame. The simulations demonstrate improved prediction of key flow and flame characteristics, including recirculation zone length, velocity recovery, and downstream temperature distributions, with lower sensitivity to mesh resolution. These findings highlight the potentials of the ANN-based sub-grid FDF modelling approach as a better alternative to conventional tabulation methods. • The filtered density function is predicted through the on-the-fly inference. • The data-driven model shows much lower memory usage than the look-up table. • Simulations with the data-driven model demonstrate improved accuracy. • Leveraging the model decreases the sensitivity to the mesh resolution.