Dispersion-Engineered Terahertz Spoof Plasmonic Neural Network for Parallel Computing and On-Chip Communication.

Abstract

Diffractive neural networks offer a novel physical implementation for optical computing to achieve parallelism, low power consumption, and light-speed processing. However, their limited dispersion engineering necessitates increasingly complex architectures for tasks such as spectrum recognition and simultaneous multi-class classification, which in turn leads to increased energy demands. Here, we propose a spoof plasmonic neural network (SPNN) comprising cross-cascaded spoof surface plasmonic waveguides with strong engineered dispersion properties designed for operation in the terahertz regime. This compact platform efficiently separates spectral components from a broadband input signal, achieving a data rate of 22 Gbit/s across two separated spectral channels. We experimentally show that the SPNN can simultaneously classify multiple inputs from Fashion-MNIST+MNIST or Fashion-MNIST+EMNIST datasets, achieving classification accuracies of 98.3% and 97.4% or 97.4% and 93.8%, respectively. For multi-color CIFAR-10 dataset classification, the network architecture incorporating multiple cascaded SPNNs realizes over 10% higher accuracy than single-color-channel methods by leveraging distinct color channels mapped to respective spectrum channels. These findings highlight the potential of SPNNs for machine learning applications and lay the groundwork for future terahertz chip integration.