Deep Learning for Metasurfaces and Metasurfaces for Deep Learning

Huygens' metasurfaces have emerged as a powerful means with which to manipulate incident light on a subwavelength scale and thus can be utilized for a broad range of applications in optical devices. Such optical devices have the potential to be more compact, multi-functional, more efficient, and possess novel functionalities as compared to the current state of optics. These metasurfaces are composed of subwavelength dielectric structures (meta-atoms) with a large index of refraction relative to vacuum. The high index allows the meta-atoms to support a variety of magnetic and electric multipole resonances that alter the amplitude and phase of light and can be tuned by changing the shape of the meta-atom. The electromagnetic response of a meta-atom for a given shape is difficult to predict analytically, and it is computationally intensive to optimize them with full-wave simulations, and so sophisticated methods are needed to maximize the design potential of Huygens' surfaces. Deep neural network methods have been demonstrated to be effective for both yielding results much more quickly than full wave simulations and for finding meta-atoms to meet specific design requirements (inverse-design). We present a selection of neural network architectures that have been successful for Huygens' surface design, including fully connected neural networks, convolutional neural networks, recurrent neural networks, and generative adversarial networks. Lastly, we discuss neuromorphic photonics, wherein the meta-atoms can be used to physically construct neural networks for optical computing.