TUNING OF AN ALGEBRAIC MODEL FOR SEPARATED FLOWS BY MEANS OF BAYESIAN LASSO

In this work, machine learning techniques are exploited to train a new model based on Pope’s tensorial bases, where the common definition of the turbulent viscosity is extended to define the Reynolds stress tensor as an expansion of strain-rate and rotation tensors. Specifically, a Sparse Bayesian approach has been implemented to provide new expressions for the turbulent eddy viscosities appearing in the Pope’s formulation. An experimental dataset acquired with Time-Resolved Particle Image Velocimetry has been used to tune the present model. Data provide the mean flow and the Reynolds stress distributions of different separation bubbles evolving on a flat plate with applied variable adverse pressure gradients, reproducing in some extent the diffusion that characterize the rear suction side of turbine blades. The model has been tuned with different independent conditions of flow Reynolds number and free-stream turbulence intensity. Then, two additional flow conditions have been adopted to cross-validate and test the generated model. The adoption of the Sparse Bayesian approach allows the direct identification of the leading predictors, i.e., flow features, exploiting the main characteristics of the transition scenario (here, the transition process induced by flow separation). Specifically, the correlations here tuned may contribute to improve the capability of RANS solvers in the prediction of transitional flows.