Deep Learning Surrogate for the Temporal Propagation and Scattering of Acoustic Waves

Abstract

A deep learning surrogate for the direct numerical temporal prediction of two-dimensional acoustic waves propagation and scattering with obstacles is developed through an autoregressive spatiotemporal convolutional neural network. A single database of high-fidelity lattice Boltzmann simulations is employed in the training of the network, achieving accurate predictions for long simulation times for a variety of test cases representative of bounded and unbounded configurations. The capacity of the network to extrapolate outside the manifold of examples seen during the training phase is demonstrated by obtaining accurate acoustic predictions for relevant applications, such as the scattering of acoustic waves on an airfoil trailing edge, an engine nacelle, or an in-duct propagation. The study focuses on the influence of three main design parameters that allow rolling out accurate and stable long-term predictions: 1) the choice of a dataset-related characteristic time, 2) the normalization of the input data, and 3) the number of input temporal frames into the neural network. The results show that for the optimum choice of design parameters, the presented data-driven model is able to systematically obtain low-error prediction at a lower computational cost than the reference high-fidelity computational code.