Diffractive neural networks (D2NNs) have been transformative in many fields, motivating the development of various optical computing components. However, these optical computing components, achieving in diffractive optical elements (DOEs) fabricated using commercial 3D printers, are bulky, and the size of individual neurons is only comparable to the wavelength scale. They simultaneously suffer from challenges posed by high-order diffraction and low spatial utilization. Here, we present a design of D2NNs based on all-dielectric metasurfaces, which reduces the size of individual neurons of optical D2NNs to scales significantly smaller than the wavelength. In order to demonstrate that our metasurface-based D2NNs are smart and sensitive enough comparable to state-of-the-art DOE-based D2NNs, we numerically simulated the design of two D2NNs including an optical 1-bit half-adder and a full-adder. Our experiment on aforementioned half-adder proofs the advantages of our architecture. Metasurface-based optical computational elements can offer higher spatial neuron density, while completely eliminate high-order diffraction to demonstrate the feasibility. Our metasurface-based D2NNs can facilitate miniaturization and integration on all-optical information processing especially optical computing components, and will find applications in numerous space-constrained fields such as 6G communication, optical smart sensors, and robotics.
|