MetasurfaceViT: A generic AI model for metasurface inverse design

Metasurfaces, sub-wavelength artificial structures, can control light's amplitude, phase, and polar ization, enabling applications in efficient imaging, holograms, and sensing. Recent years, AI has witnessed remarkable progress and spurred scientific discovery. In metasurface design, optical inverse design has recently emerged as a revolutionary approach. It uses deep learning to create a nonlinear mapping between optical structures and functions, bypassing time-consuming traditional design and attaining higher accuracy. Yet, current deep-learning models for optical design face limitations. They often work only for fixed wavelengths and polarizations, and lack universality as input-output vector size changes may require retraining. There's also a lack of compatibility across different application scenarios. This paper introduces MetasurfaceViT, a revolutionary generic AI model. It leverages a large amount of data using Jones matrices and physics-informed data augmentation. By pre-training through masking wavelengths and polarization channels, it can reconstruct full-wavelength Jones matrices, which will be utilized by fine-tuning model to enable inverse design. Finally, a tandem workflow appended by a forward prediction network is introduced to evaluate performance. The versatility of MetasurfaceViT with high prediction accuracy will open a new paradigm for optical inverse design.