Artificial Neural Networks Prediction of Wall-Pressure Spectrum: A Boundary-Layer Insight

An artificial neural network (ANN) is presented for modeling the wall-pressure spectrum (WPS) beneath turbulent boundary layers (BLs). This machine-learning-based method offers an alternative to traditional semi-empirical models or to previous works using ANNs, which rely on identifying relationships between specific BL parameters and the WPS. In contrast to previous approaches, the present method employs the complete BL velocity profile as input to predict the WPS. To do so, two separate ANNs are trained. On the one hand, an autoencoder compresses the information of the BL velocity profile and extracts its relevant features, known as latent space. On the other hand, a second network predicts the WPS based on the latent code of the velocity profile. The ANN is trained on numerical large-eddy simulation and direct numerical simulation data that cover a wide range of flow conditions. This exploitation of the database using autoencoders has demonstrated that only three parameters are sufficient to reconstruct all velocity profiles, regardless of their nature (forward, zero, or adverse pressure gradient). In addition, the WPS ANN-based prediction captures the effect of the main midfrequency tonal peaks, as well as the overall noise trends in the low-, mid-, and high-frequency content.