Artificial-Neural-Network-Based Wall-Modeled Large-Eddy Simulation with an Eddy-Viscosity Correction for the Fully-Developed Turbulent Channel Flow

Abstract

One significant challenge in employing large-eddy simulation (LES) for engineering applications, where geometry is complex and Reynolds number is high, is the prohibitive computational cost. Such excessive computational cost comes from the inner layer region of the boundary layer, where the smallest dissipative turbulence scale becomes small with Reynolds number increase. Wall-modeled large-eddy simulation (WMLES) reduces grid requirements by modeling the inner-layer turbulence. The commonly used WMLES is an equilibrium model (i.e., log-law based model) that predicts wall- shear stress using the law of the wall derived from the momentum conservation equation with equilibrium assumption. Despite the reasonable accuracy in the outer layer flow, log-law based model showed a log-layer mismatch (LLM) problem because of inaccurate wall-shear stress prediction, and inaccurate inner layer turbulent statistics. In the present study, we suggest artificial-neural-network- based WMLES (i.e., ANN-based WMLES) and eddy-viscosity correction method to enhance the accuracy in both inner and outer layer. ANN-based WMLES predicts wall-shear stress using the outer layer flow and compensates the excluded inner layer turbulence using ANN. ANN takes FIK identity and its contributions as an input, and outputs the corrected skin friction coefficient. Although ANN- based WMLES accurately predicts wall-shear stress, the inner layer turbulent statistics were not improved significantly. We found that the sub-grid scale (SGS) stress is not modeled properly because of the coarse grid resolution in the wall-adjacent region. So, we developed eddy-viscosity correction method to improve the SGS stress modeling. The proposed models improved accuracy in both inner and outer layer turbulent statistics.