DEEP REINFORCEMENT LEARNING FOR LARGE-EDDY SIMULATION OF WALL-BOUNDED TURBULENCE

Large-eddy simulation (LES) is actively carried out for many scientific and engineering problems. Recently supervised learning approaches using high-fidelity flow data in training process are applied to LES modeling, however, there are significant disadvantages. For example, the trained model is not effective in actual LES, and such high-fidelity data is usually not available in real-world problems. To overcome these, we employed deep reinforcement learning (DRL) where actual LES and training of subgrid-scale (SGS) model are carried out simultaneously using only target statistics as given information. We additionally applied physical constraints such as reflectional invariance and wall boundary conditions on DRL for reducing the training cost. Through this, we are challenging to find a reliable SGS model for three-dimensional LES of wall-bounded turbulence. The DRL model that produces the local SGS stress based on the local velocity gradient were trained, as a result, we found that in various training environments DRL could discover models that make mean velocity and mean Reynolds shear stress of actual LES be consistent with the target, while the conventional SGS models usually mispredict them. We conclude that DRL would be a effective tool for turbulence modeling in practical problems.