High-capacity directional information processor using all-optical multilayered neural networks

We propose a directional-diffractive deep neural network (D-D²NN) by encoding the wave propagation direction into a D²NN, introducing a new degree of freedom and enable high-capacity information processing. The unprecedented capability of metasurfaces in controlling light propagation provides a new opportunity to realize such a D-D²NN. We propose and demonstrate a versatile D-D²NN using three spin-decoupled metasurfaces. Simultaneous control of geometric and propagation phases is used to realize direction-dependent functionality. High-capacity information processing is demonstrated by controlling the distances between neighboring metasurfaces. The efficacy of our approach is exemplified through the classification of digits and fashion products in two channels and the calculation-like function realized in four channels. Moreover, the proposed D-D²NN can divide rich information into multiple channels, enabling high-volume data encryption. The metasurface-enabled deep learning networks with versatile directional functionalities provide a flexible route toward massively parallel processing, pattern recognition, encryption, and artificial intelligence systems.