Airfoil self-noise prediction using deep neural networks

Abstract

This study explores the prediction of airfoil self-noise through Deep Learning whilst focusing more specifically on the so-called turbulent boundary layer trailing edge (TBL-TE) noise. To this end, a predictive model relying on a Deep Neural Network (DNN) is developed, being then trained using an experimental database of TBL-TE noise signatures previously acquired by NASA. The DNN is favorably benchmarked against the test results, demonstrating its superiority over a popular semi-empirical prediction tool, i.e., the BPM model from NASA. Special attention is paid to the sensitivity of the DNN towards its architecture and/or its training extent. All in all, the DNN proves robust and accurate, reproducing faithfully the TBL-TE noise signatures with an average error of about 1.5 ∼ 2.5 dB in terms of Sound Pressure Level. Special attention is also paid to sensitivity of the DNN towards the composition of its training dataset, whose consistency is enforced by clustering all datapoints belonging to an identical test configuration. This allows evaluating how far the model constitutes a true prediction tool, i.e., can extrapolate the experimental database instead of merely interpolating it through overfitting. Finally, the sensitivity of the DNN towards its training data is further explored to tentatively discriminate which quantities constituting the experimental database may contribute more significantly to the correct prediction of the noise signatures and − by extension – to their underlying physical mechanisms.