TOWARDS EXPLAINABLE MACHINE-LEARNING-ASSISTED TURBULENCE MODELING FOR TRANSONIC FLOWS

Abstract

A machine-learning-assisted turbulence modeling framework is proposed to improve the prediction accuracy of the Spalart-Allmaras turbulence model. The case studied is the transonic bump flow, which partially resembles the flow physics of a transonic compressor. A random forest model is trained, cross-validated and tested to construct a mapping between the input features and the eddy viscosity discrepancy. These input features concern the physical effects of pressure gradient, strain versus vorticity, flow misalignment, wall proximity and viscosity ratio. Results show that the proposed approach predicts an interpolation and an extrapolation test case with L1-type errors of 11.1% and 16.5%, respectively. The Shapley additive explanations method is employed to investigate the global and local sensitivities of each input feature. The capability of these input features in identifying specific flow features is discussed. The methods and results of this work provide useful guidance for turbulence model developers.