Metasurface Vision Transformer: A Generic AI Model for Metasurface Inverse Design

Metasurfaces enable diverse applications by controlling light's amplitude, phase, and polarization. Although deep learning-based inverse design has revolutionized metasurface design, current models are limited by fixed operating conditions and lack universality, often requiring retraining for new wavelengths, polarizations, or application scenarios. To address this, we introduce MetasurfaceViT (Metasurface Vision Transformer), a generic AI model for inverse design. Our solution leverages a large dataset of Jones matrices, significantly expanded via physics-informed data augmentation. By pretraining through masking wavelengths and polarization channels, MetasurfaceViT can reconstruct full-wavelength Jones matrices, which are then used by a fine-tuning model for inverse design. This versatility allows one-shot structure design for arbitrary wavelength, polarization, and application requirements. We demonstrate MetasurfaceViT's capabilities in designing multiplexed printings and holograms and broadband achromatic metalenses. Prediction accuracy exceeds 99% for physically realistic designs, showcasing a significant step toward a universal optical inverse design paradigm.