Field inversion for data-augmented RANS modelling in turbomachinery flows. (15th April 2020)

• Record Type: Journal Article
• Authors: Ferrero, Andrea; Iollo, Angelo; Larocca, Francesco
• Is Part Of: Computers & fluids. Volume 201(2020)
• Publication Date: 2020-04-15
• Subjects: Field inversion -- Machine learning -- Turbulence modelling -- Turbomachinery; Fluid dynamics -- Data processing -- Periodicals; 532.050285
• DOI: 10.1016/j.compfluid.2020.104474
• Languages: English
• ISSNs: 0045-7930

Abstract

Highlights: The field inversion approach is investigated for improving RANS models in turbomachinery flows. Working conditions characterised by transition and separation are considered. Some approaches to improve the robustness of the method are proposed. The augmented RANS model includes an Artificial Neural Network which acts as an intermittency term. The predictive ability of the method is investigated for several working conditions on different geometries. Graphical abstract: Abstract: Turbulence modelling in turbomachinery flows remains a challenge, especially when transition and separation phenomena occur. Recently, several research efforts have been devoted to the improvement of closure models for Reynolds-averaged Navier-Stokes (RANS) equations by means of machine learning approaches which make it possible to extract the knowledge hidden inside the available high-fidelity data (from experiments or from scale-resolving simulations). In this work the use of the field inversion approach is investigated for the augmentation of the Spalart–Allmaras RANS model applied to the flow in low pressure gas turbine cascades. As a first step, the field inversion method is applied to the T106c cascade at two different values of Reynolds number (80000-250000): An adjoint-based gradient method is employed in order to minimise the prediction error on the wall isentropic Mach number distribution. The data obtained by the correction field are then analysed by means of an Artificial Neural Network (ANN) which makes it possible to generalise the correction by finding correlations which depend on physical variables. A study on the definition of the input variables and on the architecture of the ANN is performed. Different kind of corrections are evaluated and a particularly robust correction factor is obtained by limiting the range of the correction in the spirit of intermittency models. Finally, the ANN is introduced in an augmented version of the Spalart–Allmaras model which is tested on the T106c cascade (for values of the Reynolds number not considered during the training) and for the T2 cascade. The prediction ability of the method is investigated by comparing the numerical predictions with the available experimental data not only in terms of wall isentropic Mach number distribution (which was used as goal function during the field inversion) but also in terms of mass-averaged exit angle and kinetic losses.