Acoustic metamaterial design using Conditional Wasserstein Generative Adversarial Networks

Abstract

This talk presents a method for generating planar configurations of scatterers with a reduced total scattering cross section (TSCS) by means of generative modeling and deep learning. The TSCS minimization via repeated forward modeling techniques, trial-error methods, and traditional optimization methods requires considerable computer resources and time. However, similar or better results can be achieved more efficiently by training a deep learning model to generate such optimized configurations producing low scattering effect. In this work, the Conditional Wasserstein Generative Adversarial Networks (cWGAN) is combined with Convolutional Neural Networks (CNN) to create the generative modeling architecture [1]. The generative model is implemented with a conditional proponent to allow the TSCS targeted design generation and is enhanced with the coordinate convolution (CordConv) layer to improve the model’s spatial recognition of cylinder configurations. The cWGAN model [1] is capable of generating images of 2D configurations of scatterers that exhibit low scattering. The method is demonstrated by giving examples of generating 2-cylinder and 4-cylinder planar configurations with minimal TSCS. [1]. P. Lai, F. Amirkulova, and P. Gerstoft. “Conditional Wasserstein generative adversarial networks applied to acoustic metamaterial design,” J. Acoust. Soc. Am. 150(6), 4362–4374 (2021).