Sound manipulation through multi-scattering, gradient-based optimization, deep learning and reinforcement learning

Abstract

This talk discusses recent advances and current challenges encountered during the application of gradient-based optimization, deep learning, reinforcement learning, and generative modeling in an inverse design of acoustic metamaterials. We first summarize and group existing approaches in the inverse design of metamaterials then discuss current algorithmic limitations and open challenges to preview possible future developments in metamaterial design. Specifically, broadband sound focusing, cloaking, steering, localization, and diffusion are the focus of this talk. The objective (loss) functions are evaluated by means of multiple scattering theory, and analytical gradients are evaluated with respect to a design vector, i.e., the positions of each scatterer. Our observations show that the development of hybrid models improves the performance of these algorithms. Specifically, the performance of deep reinforcement learning and gradient-based optimization as well as generative network models are enhanced when gradients of target functions are supplied to the model. In reinforcement learning models, the agent receives a reward proportional to the negative of the target function. The performance of the double-deep Q-learning network (DDQN) and the deep deterministic policy gradient (DDPG) algorithms were improved when gradients were supplied to the reward function. Numerical examples are presented for planar configurations of cylindrical scatterers.