Integrated photonic neural network based on silicon metalines.

Abstract

An integrated photonic neural network is proposed based on on-chip cascaded one-dimensional (1D) metasurfaces. High-contrast transmitarray metasurfaces, termed as metalines in this paper, are defined sequentially in the silicon-on-insulator substrate with a distance much larger than the operation wavelength. Matrix-vector multiplications can be accomplished in parallel and with low energy consumption due to intrinsic parallelism and low-loss of silicon metalines. The proposed on-chip whole-passive fully-optical meta-neural-network is very compact and works at the speed of light, with very low energy consumption. Various complex functions that are performed by digital neural networks can be implemented by our proposal at the wavelength of 1.55 µm. As an example, the performance of our optical neural network is benchmarked on the prototypical machine learning task of classification of handwritten digits images from the Modified National Institute of Standards and Technology (MNIST) dataset, and an accuracy comparable to the state of the art is achieved.