Subgrid-scale modelling using deconvolutional artificial neural networks in large eddy simulations of chemically reacting compressible turbulence

Abstract

Based on the framework of deconvolutional artificial neural network (DANN) proposed by Yuan et al.[DOI:https://doi.org/10.1063/5.0027146], we extend the DANN approach to model subgrid-scale (SGS) terms in large eddy simulation (LES) of chemically reacting compressible turbulent flow with evident heat release. In constructing the DANN, the normalized density-weighted filtered variables in the neighbouring stencils are taken as the inputs while the outputs are unfiltered density-weighted variables. The SGS stress, SGS heat flux, SGS scalar flux and chemical reaction source terms are modelled using those unfiltered variables which are recovered by the DANN framework to close the governing equations. The DANN framework is evaluated by two chemically reacting compressible isotropic turbulent flow cases adopting a simple one-step irreversible chemical reaction mechanism at turbulent Mach numbers 0.4 and 0.8. In the a priori study, the DANN method shows better performance compared to the classical approximate-deconvolution method (ADM) and the velocity-gradient model (VGM). In the a posteriori test, the DANN method outperforms the dynamic-Smagorinsky model (DSM), and the dynamic-mixed model (DMM). In addition, the DANN framework can predict flow variables with high accuracy by using limited training samples which are constructed from any single instantaneous flow data during the reaction process.