Holograms which reconstruct the transverse profile of light with complex-amplitude information have demonstrated more excellent performances with an improved signal-to-noise ratio compared with those containing amplitude-only and phase-only information. Metasurfaces have been widely utilized for complex-amplitude holograms owing to its capability of arbitrary light modulation at a subwavelength scale which conventional holographic devices cannot achieve. However, existing methods for metasurface-based complex-amplitude hologram design employ single back-diffraction propagation and rely on the artificial blocks which are able to independently and completely control both amplitude and phase. Here, we propose an unsupervised physics-driven deep neural network for the design of metasurface-based complex-amplitude holograms using artificial blocks with incomplete light modulation. This method integrates a neural network module with a forward physical propagation module and directly maps geometric parameters of the blocks to holographic images for end-to-end design. The perfect reconstruction of holographic images verified by numerical simulations has demonstrated that compared with the complete blocks, an efficient utilization, association and cooperation of the limited artificial blocks can achieve reconstruction performance as well. Furthermore, more restricted controls of the incident light are adopted for robustness test. The proposed method offers a real-time and robust way towards large-scale ideal holographic displays with subwavelength resolution.

中文翻译：

---

通过物理驱动的深度神经网络对基于超表面的复振幅全息图进行端到端设计

与仅包含幅度和仅相位信息的全息图相比，使用复幅度信息重建光的横向轮廓的全息图表现出更出色的性能，并提高了信噪比。超表面已被广泛用于复振幅全息图，因为它具有在亚波长范围内进行任意光调制的能力，这是传统全息设备无法实现的。然而，现有的基于超表面的复振幅全息图设计方法采用单一的背向衍射传播，并且依赖于能够独立和完全控制振幅和相位的人工块。这里，我们提出了一种无监督的物理驱动的深度神经网络，用于使用具有不完全光调制的人工块设计基于超表面的复振幅全息图。该方法将神经网络模块与前向物理传播模块集成在一起，并将块的几何参数直接映射到全息图像以进行端到端设计。数值模拟验证的全息图像完美重建表明，与完整块相比，有限人工块的有效利用、关联和协作也可以实现重建性能。此外，鲁棒性测试采用了对入射光的更多限制控制。