Freeform metasurface design with a conditional generative adversarial network.

As a two-dimensional artificial composite electromagnetic material, metasurfaces have achieved complex control of inherent properties such as amplitude, phase, polarization, and frequency of electromagnetic waves. However, the design of metasurfaces requires deep expertise and intensive iterative calculations. Deep learning, as a machine learning method, has been applied to design metasurfaces in recent years. The traditional neural network architecture solves the reverse design of nanophotonic metasurfaces as a regression problem, thereby achieving the mapping of optical responses to structural design space. This approach suffers from a lack of flexibility in the design of nanophotonic structures, as they typically limit the optimization process to a predefined set of design candidates, making it difficult to create entirely new metasurface designs. Here, we propose a deep learning method based on conditional generative adversarial networks (CGAN), which is capable of generating new graphic patterns. Compared with traditional neural networks, CGAN can achieve higher accuracy and generalization performance in reverse design. [ABSTRACT FROM AUTHOR]