Consistently Enforced Wall Models by Reinforcement Learning for Wall-Modeled Large-Eddy Simulation

A reinforcement-learning-based wall-modeled large-eddy simulation (RL-WMLES) framework is proposed to improve the physical consistency of near-wall turbulence predictions. In this approach, a reinforcement learning agent is coupled with the WMLES solver to dynamically adjust a compensating stress term, with the objective of enforcing agreement between the LES solution and the law of the wall. The agent is trained using the proximal policy optimization (PPO) algorithm, where the state is defined as the discrepancy between the near-wall LES velocity and the wall-model prediction, and the action corresponds to modifying a parameterized support viscosity distribution. The proposed method is implemented within a high-performance CFD solver and trained on turbulent channel flow. Numerical results demonstrate that the trained agent effectively reduces the log-layer mismatch and significantly improves the accuracy of near-wall velocity predictions. Furthermore, the RL-WMLES framework exhibits a degree of generalization capability: the trained agent performs robustly with varying levels of numerical dissipation and Reynolds numbers. By introducing a simple interpolation strategy, the same agent can be successfully applied to configurations with different matching locations. Overall, the RL-WMLES framework provides a flexible and data-driven approach for enforcing physical constraints in turbulence modeling. The method shows strong potential for extension to more complex flows.