Metasurface-empowered snapshot hyperspectral imaging with convex/deep (CODE) small-data learning theory.

Area of Science:

• Optics and Photonics
• Material Science
• Computer Vision

Background:

• Hyperspectral imaging is crucial for material identification but limited by bulky traditional systems.
• Existing metasurface solutions reduce volume but face fabrication complexity and large footprints.
• Compact and efficient hyperspectral imaging systems are needed for broader applications.

Purpose of the Study:

• To develop a compact snapshot hyperspectral imager.
• To integrate meta-optics with small-data deep learning theory.
• To demonstrate high-fidelity hyperspectral data generation with minimal training.

Main Methods:

• Developed a snapshot hyperspectral imager using a single multi-wavelength metasurface chip (500-650 nm).
• Incorporated the meta-optics with small-data convex/deep (CODE) deep learning theory.
• Trained the system using a 4-band multispectral dataset and only 18 training data points.

Main Results:

• Achieved a significantly reduced device area compared to traditional systems.
• Efficiently generated an 18-band hyperspectral data cube with high fidelity.
• Demonstrated the effectiveness of CODE deep learning for hyperspectral reconstruction with limited data.

Conclusions:

• The integration of multi-resonant metasurfaces and small-data learning theory enables compact hyperspectral imagers.
• This approach overcomes limitations of previous metasurface designs.
• The developed system holds promise for low-profile advanced instruments in science and industry.