Metasurface-enabled multifunctional diffractive neural networks for image reconstruction

Metasurface-enabled diffractive neural networks are all-optical computing architectures realized at the physical level and have been extensively studied in recent years. However, the immutability of fabricated metasurface diffractive layers necessitates specialized designs to meet multifunctional requirements. Approaches utilizing light's wavelength, polarization, and orbital angular momentum require multidimensional encoding of the input light field, hindering their integration into practical applications. We propose a diffractive neural network based on metasurfaces capable of performing various image reconstruction functions without the need for complex preprocessing of optical information. It achieves optical function reconfiguration simply by adjusting the physical stacking order of multiple diffractive layers. Training results demonstrate that this architecture enables seamless functional switching and possesses scalability. Furthermore, we present a diffractive neural network architecture based on a single-layer silicon substrate with double-sided etched metasurfaces, which achieves bidirectional image processing capabilities. This advances the further application of high-speed, low-power, and multifunctional diffractive neural networks based on metasurfaces.