Programmable surface plasmonic neural networks for microwave detection and processing

Abstract

A range of alternative approaches to traditional digital hardware have been explored for the implementation of artificial neural networks, including optical neural networks and diffractive deep neural networks. Spoof surface plasmon polariton waveguides, which operate at microwave and terahertz frequencies, can offer low crosstalk, low radiation loss and easy integration, and are of potential use in the development of an alternative technology for artificial neural networks. Here, we report a programmable surface plasmonic neural network that is based on a spoof surface plasmon polariton platform and can detect and process microwaves. The approach uses a parallel coupled spoof surface plasmon polariton cell integrated with varactors. The weight coefficients of the cell can be adjusted by tuning the voltages of the varactors, and the activation function of the neural network can be programmed by detecting the input intensity and feeding back the threshold to an amplifier. We show that a two-layer fully connected surface plasmonic neural network consisting of four input cells and four output cells can perform a vector classification task. The surface plasmonic neural network can also be used to create a wireless communication system to decode and recover images. In addition, we show that partially connected surface plasmonic neural networks can classify handwritten digits with high accuracy. A spoof surface plasmon polariton platform can be used to create a surface plasmonic neural network with programmable weights and activation functions.