Wall-Pressure Spectrum Model Based on Artificial Neural Networks Predictions

Abstract

We propose machine learning approach using Artificial Neural Networks (ANNs) to model the wall-pressure spectra (WPS) beneath turbulent boundary layers. Classical (semi-empirical) wall-pressure models are based on scaling laws according to inner and/or outer parameters of the boundary layer. In this approach, the complete boundary layer profile (i.e. tangent velocity as a function to the wall-normal distance) is provided as an input into the ANN. The aim of this methodology is to obtain more insight on the relationships that may arise between the turbulent boundary layer and its corresponding WPS and that have not been assessed in the literature.The analysis and training of the ANN are performed on data from Large Eddy Simulations (LES) produced by the European SCONE project. The database consists on a set of LES simulations with mach number varying in between 0.3 and 0.7, as well as Reynolds numbers in between 8.3e5 and 2.4e6 and angles of attack from 1º to 7º. The set of data includes zero and adverse pressure gradient effects, including flows experiencing strong adverse pressure gradients. In order to produce the noise prediction, the approach that has been used is the following one. An autoencoder, composed by a encoder and decoder, is trained in order to compress the boundary layer profile into a minimal amount of parameters (reduced latent space) that permit to retrieve back the shape of the profile. This latent space is used together with the flow conditions to train the ANN to produce a prediction of the WPS. It has been found that the predictions on the WPS in the high-frequency content are more reliable than on the low frequency range since more data is available for training. Also, there is large effect of the recirculating regions on the boundary layer plays an important role in the noise prediction.