Deep learning in metasurfaces: from automated design to adaptive metadevices

Abstract

Metasurface modeling, designs, and applications using computational approaches are by now well established as an essential pillar in photonics, physics, and materials science. The past years have witnessed tremendous advances in methodologies and technologies to unearth the intricate light-matter interaction and promote adaptive metadevices. They have pushed the studies of metasurfaces from early passive, reconfigurable modalities to the next generation of intelligent metasurfaces. In this review, we elaborate general architecture for intelligent metasurfaces, constructed by the algorithm layer, tunable metasurface layer, and application layer. We first discuss a variety of deep learning models, ranging from the fundamental neural networks inspired by computer science to sophisticated algorithms embedded with physical specialty, highlighting their potential in the forward prediction, inverse design, and spectral correlation of metasurfaces. We then discuss adaptive metadevices in the main applications of invisibility cloaks, smart vision, intelligent sensing, and wireless communication. Finally, we pinpoint current challenges and future perspectives to embrace the coming era of intelligent metasurfaces.