14th European Conference on Turbomachinery Fluid dynamics & Thermodynamics
Building on the interest around data-driven modelling, in this paper we investigate the feasibility of using DNS data and machine learning algorithms to assist RANS turbulence model development. High-fidelity DNS data are generated with the incompressible Navier-Stokes solver implemented in the spectral/hp element software framework Nektar++. Two test cases are considered: a turbulent channel flow and a stationary serpentine passage, representative of internal turbo-machinery cooling flow, both at a bulk Reynolds number of 5600. The former targets the near wall behaviour of turbulence, whilst the latter aims at improving RANS predictions in the passage curvature, the separation bubbles and the flow reattachment region. The Python framework TensorFlow is chosen to train neural networks (NN) in order to address known limitations from the Boussinesq constitutive relationship between Reynolds stresses and the strain rate tensor and k-omega is chosen as the underlying turbulence model for all RANS calculations. For the serpentine case a clustering based on flow features is run upfront to identify different regions of the flow and enable training on selected areas, including high curvature and separation. A priori analysis is used as a preliminary assessment tool for the new constitutive relationships and the models are then implemented to investigate the a posteriori predictions capability using RANS simulations (within the framework of Rolls-Royce tools and solver HYDRA). As model errors propagate in often unpredictable ways in RANS calculations, the ultimate assessment when comparing new models to baseline RANS is done in terms of resulting velocity field and wall shear stress. The paper presents the fundamental elements of procedure applied, including a brief description of the tools and methods, improvements achieved, theoretical and practical limitations as well as future research efforts.