Application of hybrid feature selection methods to improve turbulence models using machine learning

Abstract One of the effective approaches to improve the accuracy of RANS modelling in the case of certain classes of flows is the development of corrections which are terms or multipliers in equations of the turbulence model and depend on local characteristics of the flow. The effectiveness of such corrections is largely determined by the choice of arguments for the correction function. This paper is focused on the development of a methodology for optimal selection of such arguments based on machine learning methods. In this paper, we use a 2-stage hybrid method of feature selection that combines filter and wrapper methods. Three multivariate filter methods (RRreliefF, MIMIC, and mRMR) in combination with sequential forward selection method were considered. Using the results of training a neural network model to improve the accuracy of the Spalart–Allmaras (SA) model in the Bump Flow problem as an example, it was shown that a hybrid method using any of the considered filter methods successfully generates a set of arguments for the correction function from a predefined expanded set. Within the framework of this paper, the hybrid method with MIMIC provided the best results.