Compact and voltage-tunable surface plasmon polariton-based optical neural networks.

Abstract

Optical neural networks (ONNs) offer advantages in parallel processing, low power consumption, and high-speed operation. However, existing ONN designs face challenges in miniaturization, stability, tunability, and integration. This study proposes a graphene surface plasmon polariton (GSPP) waveguide switch array for all-optical neural networks. The design features a compact structure with a lateral area of only 0.045 $\boldsymbol{\mathrm{\mu}}{{\mathbf m}^2}$. Numerical simulations show that within the 30.2 to 49.4 THz range, the transmission rate is tunable from 0 to 0.875, accurately simulating synaptic weights in ONNs. The compact switch array achieves a recognition accuracy of 93.83% on the CIFAR-10 dataset, demonstrating its potential for high-speed, low-power, and highly integrated neural network computing platforms.