Subgrid-scale Flux Modeling of a Passive Scalar in Turbulent Channel Flow using Artificial Neural Network

A deep neural network (DNN) has been developed to model the subgrid-scale (SGS) flux associated with a passive scalar in incompressible turbulent channel flow. To construct the training dataset for the DNN, a direct numerical simulation (DNS) was performed for a channel flow at the friction Reynolds number  encompassing a passive scalar transport with Prandtl number  using a pseudo-spectral in-house code. The DNS data of velocity and scalar fields was filtered to obtain the SGS scalar flux vector, , filtered scalar gradient, and filtered strain-rate tensor, which were subsequently used to train the DNN, enabling it to predict  for large-eddy simulation. A priori evaluation of the DNN’s performance in predicting  revealed a close match with filtered DNS data, demonstrating correlations of up to 98%, 79% and 85% for the three components of . Additionally, analysis of the mean SGS dissipation and its probability density function indicated promising predictions by the DNN. Notably, this study extends the applications of DNNs for predicting  to the case of turbulent channel flow.