A recursive neural-network-based subgrid-scale model for large eddy simulation of turbulent channel flow

Cho et al. (JFM 2024) proposed a recursive neural-network-based subgrid-scale (SGS) model (RNM) for large-eddy simulation (LES) of three-dimensional homogeneous isotropic turbulence (HIT). In their framework, a neural network (NN) was initially trained on filtered direct numerical simulation (fDNS) data at a low Reynolds number, applied to LES at a higher Reynolds number, and then retrained using an augmented dataset comprising both fDNS and fLES data. This retrained NN was applied to LES at further higher Reynolds numbers. The recursive process continued until a target Reynolds number was reached. The RNM outperformed traditional SGS models in HIT across a range of Reynolds numbers. However, its generalization to other untrained flows remains limited, as the training data consisted solely of HIT. In the present study, we extend the RNM framework to turbulent channel flow (TCF), which introduces new challenges because the near-wall region in TCF exhibits significantly lower SGS stress magnitudes, and the inhomogeneous wall-normal direction introduces strong non-zero mean components. To address these TCF-specific characteristics, we propose methodological modifications tailored for TCF. The NN is initially trained at the frictional Reynolds number of Re_τ=395 and recursively updated up to Re_τ=2000 without relying on high-Reynolds-number fDNS data. The performance of the RNM is evaluated and compared against traditional SGS models in LES of TCF at Re_τ=2000.