A data-driven subgrid scale model in Large Eddy Simulation of turbulent premixed combustion

Abstract

We derive a data-driven model of a subgrid scale (SGS) closure for turbulent premixed combustion in the context of Large Eddy Simulation (LES) using deep learning. We validate the model through LES of the direct numerical simulation (DNS) flame configuration and compare it to other subgrid models from the literature. The filtered DNS used as training data was provided by Lapeyre et al. (2019). The neural network developed in this study was designed to estimate the SGS flame surface density, using only local progress variable values as a basis. A priori tests show that the results inferred from the frozen neural network were comparable to results obtained from the convolutional neural networks (CNNs) using the full nonlocal set of variables, and were in good agreement with the filtered DNS. A model-agnostic method for interpreting machine learning was employed to investigate the behavior of the trained neural network. A posteriori evaluation using the network as an LES subgrid model demonstrates that the proposed data-driven modeling is more accurate than classical algebraic models in terms of the integrated flame area in the axial direction. This illustrates that the proposed data-driven subgrid model to represent the non-linear unresolved terms is a successful approximation from both an a priori and an a posteriori perspective and that only fully local values in the filtered domain suffice to yield good agreement with DNS results. This is in contrast to earlier attempts, which use the full LES domain dataset as input to the network.