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

Area of Science:

• Photonics and Optical Computing
• Terahertz Technology
• Machine Learning Hardware

Background:

• Diffractive neural networks offer optical computing advantages but face limitations in dispersion engineering for complex tasks.
• Increased complexity in diffractive networks leads to higher energy consumption for spectrum recognition and multi-class classification.

Purpose of the Study:

• To propose and demonstrate a compact spoof plasmonic neural network (SPNN) for efficient terahertz-speed optical computing.
• To leverage engineered dispersion in SPNNs for enhanced spectral separation and multi-class classification.

Main Methods:

• Utilizing cross-cascaded spoof surface plasmonic waveguides with strong engineered dispersion for terahertz operation.
• Implementing SPNNs for broadband signal spectral component separation and data transmission.
• Experimentally validating SPNN performance on Fashion-MNIST, EMNIST, and multi-color CIFAR-10 datasets.

Main Results:

• Achieved a data rate of 22 Gbit/s across two separated spectral channels.
• Demonstrated high classification accuracies: 98.3% and 97.4% for Fashion-MNIST+MNIST/EMNIST.
• Showcased over 10% accuracy improvement for multi-color CIFAR-10 classification by utilizing distinct color channels.

Conclusions:

• SPNNs offer a promising platform for high-performance machine learning applications in the terahertz regime.
• The engineered dispersion in SPNNs efficiently handles spectral information for complex data processing.
• This work lays the foundation for future terahertz chip integration and advanced optical computing solutions.