Direct field-to-pattern monolithic design of holographic metasurface via residual encoder-decoder convolutional neural network

Complex-amplitude holographic metasurfaces (CAHMs) with the flexibility in modulating phase and amplitude profiles have been used to manipulate propagation of wavefront an unprecedented level, leading higher image-reconstruction quality compared their natural counterparts. However, prevailing design methods CAHMs are based on Huygens-Fresnel theory, meta-atom optimization, numerical simulation experimental verification, which results a consumption computing resources. Here, we applied residual encoder-decoder convolutional neural network directly map electric field distributions input images for monolithic metasurface design. A pretrained is firstly trained by calculated diffraction subsequently migrated as transfer learning framework simulated images. The training show that normalized mean pixel error about 3% dataset. As prototypes fabricated, measured. reconstructed reverse-engineered exhibits high similarity target field, demonstrates effectiveness our Encouragingly, this work provides field-to-pattern method CAHMs, paves new route direct reconstruction metasurfaces.