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

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Materials Science

Background:

• Metasurfaces offer precise control over light properties (amplitude, phase, polarization) for advanced optical applications.
• Current deep learning models for metasurface inverse design lack universality and require retraining for new parameters like wavelength or polarization.

Purpose of the Study:

• To develop a universal, AI-driven inverse design model for metasurfaces that operates across various conditions.
• To enable efficient, one-shot design of metasurface structures for arbitrary optical requirements.

Main Methods:

• Introduced MetasurfaceViT, a Vision Transformer-based AI model for metasurface inverse design.
• Utilized a large dataset of Jones matrices, augmented with physics-informed data.
• Employed pretraining with masked wavelength and polarization channels to enable reconstruction of full-wavelength Jones matrices.

Main Results:

• MetasurfaceViT achieved over 99% prediction accuracy for physically realistic designs.
• Demonstrated successful one-shot design of multiplexed holograms, printings, and broadband achromatic metalenses.
• Showcased the model's versatility for arbitrary wavelength, polarization, and application requirements.

Conclusions:

• MetasurfaceViT represents a significant advancement towards a universal paradigm for optical inverse design.
• The developed AI model overcomes the limitations of fixed-condition designs, paving the way for more adaptable and efficient metasurface engineering.