Diffractive Deep Neural Networks based on Programmable Nonlinear Metasurfaces

Neural morphic computing, including all-optical diffractive deep neural networks (D2NNs) and photonic circuit neural networks, has gained popularity for their fast and efficient computing architecture. Microwave-based D2NNs have also been proposed, such as reconfigurable metasurface-based D2NNs that allows real-time adjustment of transmittance weight using programmable amplifiers [1]. Planar neural networks constructed using artificial surface plasmon polaritons (SSPPs) have also been developed, where the programmable weight matrix can be modulated in real-time by controlling the voltage applied to a varactor diode coupled to the SSPP coupler [2]. However, these designs lack reconfigurable and self-contained non-linear activation functions that do not rely on external digital control circuits.