Broadband suppression of total multiple scattering cross section using neural networks

Abstract

We will demonstrate a novel method to simulate acoustic multiple scattering by a configuration of cylinders and solve inverse problems using artificial neural networks (NN) and deep learning. We will research how to apply deep learning to solve inverse design problems efficiently. Solving inverse design problems using optimization requires an iterative process of function evaluations and determining gradients, which are computationally expensive. In this work, forward multiple scattering problems are solved first by means of multiple scattering theory to provide training data for NN. Then, NN are trained to approximate the total scattering cross section (TSCS) function using backpropagation algorithm; the input of NN is positions of the cylinders, and the output is the TSCS evaluated at discrete values of wavenumber. Finally, trained NN are employed to solve inverse problems. Specifically, the TSCS by a plane configuration of cylinders is minimized over a range of wavenumbers using trained NN. A suppression of the TSCS can lead to the efficient design of broadband acoustic cloak. This method will be illustrated giving examples for a plane configuration of rigid cylinders.