Predicting the sound refraction on shear layers with deep neural networks

A novel methodology to create a surrogate model able to propagate acoustic waves in complex flows is presented. Here, the problem of mean shear flow is considered, represented by the academic configuration of acoustic propagation of an oblique plane wave in a thick planar shear layer. The methodology justification as well as a detailed explanation of the database generated with the LES solver AVBP consisting of 59 runs with different incident wave angles has been presented. Depending on this angle, full transmission, partial transmission and reflection, full reflections are obtained. The reflection and transmission coefficients also compare favorably with an asymptotic analytical model for propagation through a thick shear layer. The trained neural network is used to make predictions for the acoustic propagation. First the capacity to predict in the same range as the database is evaluated (is the network able to learn the training data?), then, and more importantly, the capacity of the network to extrapolate outside its training range (e.g. in terms of shear layer parameters M, δ) is assessed for the major different propagation regimes highlighted above. Good results are obtained for the full transmission case, whereas larger errors are observed for the full reflection case. Intermediate results are found for the partial transmission and reflection, with again most of the errors located on the reflection side.