Reconfigurable metamaterial neuromorphic computing

Abstract

Neuromorphic computing was originally introduced in electronic circuits to mimic neuro-biological architectures. In these systems, a physical agent (e.g., an electromagnetic or acoustic wave) propagates through multiple layers of metasurfaces which are trained to perform a computational task (e.g., classification). Despite their potential, current neuromorphic metasurfaces rely on passive designs which limits their computational power to a single task. Furthermore, attempts to realize these systems in the context of mechanical wave propagation have been very scarce. This work presents a neuromorphic metasurface which is designed to exploit elastic wave scattering to realize a physical computing environment. Owing to the reconfigurable design of the chosen unit cell, the neuromorphic metasurface can be tuned to conduct multiple distinct classification tasks without the need for remanufacturing. The designed subwavelength metasurface cell will be used to train a customized neural network with constant weights (representing the elastic wave propagation between different layers of metasurface) and trainable activation functions (representing the phase modulation at each layer of metasurface). To perform distinct classification tasks, the trainable parameter in the activation function will be tuned accordingly. This work opens up new avenues in high performance mechanical computing.