15th European Conference on Turbomachinery Fluid dynamics & Thermodynamics
Authors
Abstract
Corner separation in axial compressors is a complex phenomenon, which is hardly well captured by traditional steady RANS calculations, primarily because of the deficiencies of the Reynolds stress tensor formulations. In this work a Machine Learning (ML) framework is applied to Large-Eddy Simulation (LES) data to develop non-linear turbulence stress closures. Two linear compressor cascade LES calculations are run with the Rolls-Royce solver HYDRA: one at nominal incidence condition, for which no corner separation is observed, and one at high incidence, where more complex secondary flow structures and recirculation appear. The two cases are validated against experiments performed in the linear cascade facility at LMFA, Lyon. Wilcox’s k −ω SST is used as the baseline turbulence model and is found to perform well at nominal conditions, but poorly at higher incidence, as it strongly overpredicts losses and secondary flows. The coefficients of an Explicit Algebraic Reynolds Stress Model (EARSM) are trained using Artificial Neural Networks (ANN) on the high incidence dataset. When tested a posteriori in HYDRA, improvements are observed in the prediction of endwall losses, as well as vorticity and Reynolds stress contours downstream of the separation region. The model is also proven to not worsen the baseline predictions at nominal incidence. Flow transition and coefficient interactions in the near wall region are reviewed. Finally, endwall loss polars are computed and the effects of moving from a linear to a non-linear constitutive relationship are discussed.
ETC2023-248
Paper is available for download