Metasurfaces have become a promising means for manipulating optical wavefronts in flat and high-performance optical devices. Conventional metasurface device design relies on trial-and-error methods to obtain target electromagnetic (EM) responses, an approach that demands significant efforts to investigate the enormous number of possible meta-atom structures. In this paper, a deep learning modeling approach is introduced that significantly improves on both speed and accuracy compared to techniques currently used to characterize the subwavelength optical structures. Our neural network approach overcomes two key challenges that have limited previous neural-network-based design schemes: input/output vector dimensional mismatch and accurate EM-wave phase prediction. Additionally, this is the first neural network to characterize 3-D dielectric structures. By combining with optimization algorithms or neural networks, this approach can be generically applied to a wide variety of metasurface device designs across the entire electromagnetic spectrum. Using this new methodology, examples of neural networks capable of producing on-demand designs for meta-atoms, metasurface filters, and phase-change reconfigurable metasurfaces are demonstrated.

中文翻译：

---

目标驱动全介电超表面设计的深度学习方法

超表面已经成为在平面和高性能光学设备中操纵光波前的一种有前途的手段。常规的超表面装置设计依靠反复试验的方法来获得目标电磁（EM）响应，该方法需要大量的努力来研究大量可能的超原子结构。本文介绍了一种深度学习建模方法，与目前用于表征亚波长光学结构的技术相比，该方法可显着提高速度和准确性。我们的神经网络方法克服了两个关键挑战，这些挑战已限制了以前基于神经网络的设计方案：输入/输出矢量维失配和准确的EM波相位预测。此外，这是第一个表征3-D介电结构的神经网络。通过与优化算法或神经网络结合，该方法可以普遍应用于整个电磁频谱中的各种超表面设备设计。使用这种新方法，展示了能够按需设计超原子，超颖表面滤波器和相变可重构超颖表面的神经网络示例。