Field-inversion machine learning for unsteady aerodynamics

Field inversion machine learning (FIML) has been successfully used to improve turbulence and transition models for steady problems. In FIML, a scalar field, β(x), which augments turbulence

production or dissipation, is obtained by solving an optimization problem that minimizes the discrepancy between the trusted solution (obtained via DNS, LES, or experiments) and RANS

prediction. Machine learning (ML) is then used to relate β to flow features, η. This two-step approach of first solving for β and then identifying the relation, β(η) using ML is called the

‘classic FIML’ approach.

In this work, we extend the classic FIML approach to improve the k−ω SST RANS model for predicting unsteady flow over a pitching airfoil at Reynolds number, Re = 200,000. The NACA

0012 airfoil is pitched about its quarter-chord point at a constant rate until it experiences dynamic stall. β(η) is obtained by training an ML model using static airfoil simulations at several angles of attack (α) up to static stall. Two approaches are investigated. In the first, we train across α and obtain one ML model which is used for the entire airfoil maneuver. In the second, we

train different ML models for different α; during the dynamic simulation, β is linearly interpolated between the values predicted by the different ML models. The FIML-improved RANS predictions are compared against LES results.