IMPROVEMENT IN UNSTEADY WAKE PREDICTION THROUGH MACHINE LEARNING BASED RANS MODEL TRAINING

The inability of RANS to correctly capture profile wake dynamics and decay prevents the accurate prediction of unsteady losses and poses additional challenges to the aeromechanic verification of both turbines and compressors. This paper addresses this problem by introducing a novel technique applied to improve wake prediction through a RANS based calculation. In particular, for the first time, a data-driven approach is used to develop bespoke turbulence closures to be used in the context of unsteady RANS (URANS) calculations. The new closure is obtained from an evolutionary machine learning algorithm. The algorithm requires a high-fidelity dataset which, in this study, is provided through a DNS for the wake downstream of a normal flat plate with a Reynolds number = 2,000, based on the freestream velocity and plate height. URANS are conducted with the developed closure and the mean flow statistics are compared with the high-fidelity data. The results from the developed closure show excellent agreement with the reference data. Furthermore, the generalisability of the developed closure is evaluated by considering other flat plate wake data, which differ in both the aspect ratio and the Reynolds number from the case used in this study to develop the closures. The results once again show remarkable improvement compared with the standard URANS