Prediction of wall-pressure fluctuations for separating/reattaching flows applied to space launchers using zonal detached eddy simulation-based convolutional neural networks
Simon Lecler, Pierre-Élie Weiss, Sébastien Deck
Abstract
Convolutional neural networks (CNNs) are used to predict the fluctuating wall-pressure coefficient and associated single-point pressure spectra in the separating/reattaching flow region around a generic space launcher configuration in the transonic regime. The neural networks are trained on a generic axisymmetric afterbody configuration. A Zonal Detached Eddy Simulation of a semi-realistic launcher geometry [NASA (National Aeronautics and Space Administration) model 11 hammerhead] is performed and validated using available experimental results. This configuration is used as a testing case for the trained models. It is shown that the CNNs are able to identify flow features related to physical phenomena of the flow. From this feature identification, the models are able to predict the evolution of fluctuating wall quantities and locate the regions of high pressure fluctuations. A scaling procedure is proposed to retrieve correct levels of the predicted quantities for a given unknown configuration having different free stream conditions. We also demonstrate that the present models perform well applied on Reynolds-Averaged Navier–Stokes mean flow fields, paving the way for a significant reduction in the computational cost for predicting wall-pressure fluctuations around space launchers.