Enabling smart vision with metasurfaces

Conventional optical components rely on gradual phase shifts accumulated during light propagation to shape light beams. New degrees of freedom are attained by introducing abrupt phase changes over the scale of the wavelength. A two-dimensional array of optical resonators with spatially varying phase response and subwavelength separation can imprint such phase discontinuities on propagating light as it traverses the interface between two media. Anomalous reflection and refraction phenomena are observed in this regime in optically thin arrays of metallic antennas on silicon with a linear phase variation along the interface, which are in excellent agreement with generalized laws derived from Fermat's principle. Phase discontinuities provide great flexibility in the design of light beams, as illustrated by the generation of optical vortices through use of planar designer metallic interfaces.

Metasurfaces are planar structures that locally modify the polarization, phase and amplitude of light in reflection or transmission, thus enabling lithographically patterned flat optical components with functionalities controlled by design. Transmissive metasurfaces are especially important, as most optical systems used in practice operate in transmission. Several types of transmissive metasurface have been realized, but with either low transmission efficiencies or limited control over polarization and phase. Here, we show a metasurface platform based on high-contrast dielectric elliptical nanoposts that provides complete control of polarization and phase with subwavelength spatial resolution and an experimentally measured efficiency ranging from 72% to 97%, depending on the exact design. Such complete control enables the realization of most free-space transmissive optical elements such as lenses, phase plates, wave plates, polarizers, beamsplitters, as well as polarization-switchable phase holograms and arbitrary vector beam generators using the same metamaterial platform.

Deep learning has been transforming our ability to execute advanced inference tasks using computers. We introduce a physical mechanism to perform machine learning by demonstrating an all-optical Diffractive Deep Neural Network (D2NN) architecture that can implement various functions following the deep learning-based design of passive diffractive layers that work collectively. We create 3D-printed D2NNs that implement classification of images of handwritten digits and fashion products as well as the function of an imaging lens at terahertz spectrum. Our all-optical deep learning framework can perform, at the speed of light, various complex functions that computer-based neural networks can implement, and will find applications in all-optical image analysis, feature detection and object classification, also enabling new camera designs and optical components that perform unique tasks using D2NNs.