Integrated photonic metasystem for image classifications at telecommunication wavelength

Abstract

Miniaturized image classifiers are potential for revolutionizing their applications in optical communication, autonomous vehicles, and healthcare. With subwavelength structure enabled directional diffraction and dispersion engineering, the light propagation through multi-layer metasurfaces achieves wavelength-selective image recognitions on a silicon photonic platform at telecommunication wavelength. The metasystems implement high-throughput vector-by-matrix multiplications, enabled by near 103 nanoscale phase shifters as weight elements within 0.135 mm2 footprints. The diffraction manifested computing capability incorporates the fabrication and measurement related phase fluctuations, and thus the pre-trained metasystem can handle uncertainties in inputs without post-tuning. Here we demonstrate three functional metasystems: a 15-pixel spatial pattern classifier that reaches near 90% accuracy with femtosecond inputs, a multi-channel wavelength demultiplexer, and a hyperspectral image classifier. The diffractive metasystem provides an alternative machine learning architecture for photonic integrated circuits, with densely integrated phase shifters, spatially multiplexed throughput, and data processing capabilities.

Realization of optical logic gates using on-chip diffractive optical neural networks

“…5 b,c) is dominantly due to the local periodic approximation that is used in the numerical modeling. Locally periodic approximation is alternatively utilized in many numerical modeling of metasurface-optics 22 – 26 , 32 , 33 , because of large memory requirements of full-wave simulators. Also, the assumption of unit-transmission-amplitude for all meta-atom geometries along metalines can cause some small discrepancies between the two results (the reason that we mention “small discrepancies” is that the transmission amplitude of all meta-atoms is very near to 1).…”

Metasurface-enabled on-chip multiplexed diffractive neural networks in the visible

“…The physical network is integrated with CMOS imaging sensors for miniaturized and portable sensing and computing all-in-one chip. Although there were also explorations of on-chip integration 26 , 52 , our architecture can be easily mass-produced because both CMOS chips and metasurfaces can be manufactured based on semiconductor processes. Another huge advantage of MDNN is the ability to fully exploit parallel operations of light by using the multiplexing of the metasurface.…”

Integrated Photonic Neural Networks: Opportunities and Challenges

“…This scheme has a simple structure design, and the manufacturing process is compatible with the CMOS process, which is convenient for large-scale and low-cost manufacturing. Similarly, in 2022, Gu et al experimentally fabricated and demonstrated this 1D multilayer diffractive metasurface system, which was fabricated on SOI substrates by one-step lithography and etching (Figure c). The demonstrated spatial pattern classifier in a two-layer metasystem achieves accuracies of 92% and 89% under narrowband continuous wave excitation and broadband femtosecond pulse excitation, respectively.…”

“…Subwavelength structure manifested diffraction between metasurface layers (right). Reprinted with permission from ref . Copyright 2022 Springer Nature.…”