Machine Learning Based Turbulence Model Calibration for Aeroacoustic Prediction of Transonic Cavity Flows

Transonic cavity noise prediction presents significant challenges due to factors such as edge separation, recirculation, shear layer oscillation, vortex shedding, and turbulent structures, all of which contribute to strong acoustic effects. In this work, we aim to develop a deep learning surrogate model to optimize the coefficients of the Shear Stress Transport (SST) turbulence model, enabling more accurate aero-acoustic predictions of transonic flow over a cavity. Unsteady Reynolds Averaged Navier-Stokes (RANS) simulations are performed using the OpenFOAM finite volume solver, employing various sets of model coefficients to obtain pressure measurements at probes positioned along the cavity floor. A sensitivity analysis, based on variance-based decomposition, is conducted to identify the most influential turbulence model parameters. A hybrid deep learning architecture, combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, is developed as a surrogate model to predict unsteady pressure signals and overall sound pressure levels (OASPL) at the probe locations. This surrogate model is integrated into an optimization framework, utilizing genetic algorithms to calibrate the turbulence model coefficients. The optimized coefficients lead to improvements in the prediction of pressure fluctuations and OASPL distributions along the cavity floor.