Acoustic metamaterial-based neural network for task solving

Abstract

Simple and precise sound manipulation is of fundamental interest and practical significance in acoustics and many related fields. The recent emergence of acoustic metamaterials offers acoustic properties unattainable in the nature and opens up possibility for controlling sound, which may also find applications in diverse scenarios. However, acoustic metamaterials still cannot directly solve complicated tasks, such as recognition of objects that needs to analyze scattered wave. Deep learning technique helps to eliminate the conventional dependence on human experts, but is still arouse issues of computation complexity, energy supply, device size and cost. We build a passive neural network by using numerous metamaterial unit cells, whose phase shifts are chosen as learnable parameters for the machine learning training. Such a metamaterial-based neural network interacts with the scattered wave produced by the object and accurately redistributes the output energy on a detection plane, exhibiting the “intelligence” to perform complex machine-learning tasks. Its equivalence with a conventional neural network is analytically proved, and its task-solving performance is numerically and experimentally demonstrated. More importantly, detection and computation are performed simultaneously, and the computational complexity or energy consumption will not increase even if the object becomes more complicated.