Characterizing and Improving Predictive Accuracy in Shock-Turbulent Boundary Layer Interactions Using Data-driven Models

摘要

A data-driven framework is applied to enhance Reynolds-averaged Navier-Stokes (RANS) predictions of flows involving shock-boundary layer interactions. The methodology involves solving inverse problems to infer spatial discrepancies in the Spalart Allmaras (SA) model and projecting these discrepancies to locally non-dimensional flow features using machine learning. The machine-learned reconstruction of the discrepancy is then used within the RANS partial differential equation solver for predictions. The methodology is applied to problems involving transonic flow over an axisymmetric bump, oblique shock-boundary layer interactions and shock train flows. The ability of the model to assimilate data (surface pressure and field velocities) while predicting other quantities (Reynolds stresses) is studied. Different approaches to infer discrepancies are compared, including a form that preserves a log-layer constraint in the SA model.