Driving deep-learning-based metasurface design with Kramers-Kronig relations

Abstract

Deep-learning-enabled design is an essential part of common metamaterial applications such as diverse metasurfaces, plasmonic nanostructures, and photonic crystals. Further progress in this field has become stultified, though, due to the lack of physical insight in ``black box'' design approaches. The authors propose a neural-network framework that includes an adversary channel based on the familiar Kramers-Kronig relations, yielding high-precision output that conforms to the internal physics even in cases of incomplete physical representations and equations. This framework is also widely applicable, beyond just metasurface design.