CDS&E: Reinforcement learning for robust wall models in large-eddy simulations

California Institute Of Technology, Pasadena CA

Investigators

Abstract

Simulations of wall-bounded turbulent flows have become a key element in the design cycle of wind farms and aircraft, and a major factor in the predictive capabilities of simulations of atmospheric flows. Due to the high Reynolds numbers associated with these flows, simulations resolving all scales of motion are not attainable with current computing capabilities. Specifically, wall models are necessary to overcome the prohibitive grid resolution requirements in the near-wall region. The abundance of data from experiments and simulations and the advent of machine learning have provided a boost to turbulence modeling efforts. However, simulations of turbulent flows remain hindered by the inability of heuristics and supervised learning to accurately model the near-wall dynamics. The principal aim of this project is to develop a robust wall model that can accurately predict the near-wall dynamics. The project will also encompass significant educational activities, including a multi-year undergraduate summer research program for the under-represented minority groups. The goal of the project is to develop a robust wall model for large-eddy simulations through reinforcement learning. Presently, the development of the state-of-the-art wall models relies on Reynolds-averaged Navier-Stokes parametrizations with an explicit or implicit assumption of a particular flow state close to the wall. These assumptions limit the robustness and applicability of the model and often lead to erroneous predictions of separation and laminar-to-turbulent transition, both of which are crucial components in external aerodynamics. By utilizing reinforcement learning methods, the project will allow the development of novel wall models that can adapt to various flow configurations based on the instantaneous flow input. The wall modeling problem will be cast as a control problem, where the discovered model is optimized to accurately reproduce the quantities of interest by automating the exploration of the relevant flow physics. The development of the proposed wall model will advance the state-of-the-art in the simulation of high-Reynolds-number turbulent flows in complex external aerodynamic applications. This will provide a means to obtain cheap and reliable simulations of complex flows such as flow over an aircraft. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.