Metasurface-enabled on-chip multiplexed diffractive neural networks in the visible

Area of Science:

• Photonics
• Artificial Intelligence
• Metasurface Technology

Background:

• Optical computing offers advantages in speed and power efficiency over electronic computing.
• Diffractive networks perform machine learning via optical transformations but lack multitasking capabilities.
• Existing architectures can be bulky and inefficient for complex tasks.

Purpose of the Study:

• To develop a multi-skilled diffractive neural network capable of on-chip multitasking and multi-channel sensing.
• To overcome the limitations of existing diffractive networks in terms of size and functionality.
• To demonstrate a compact and efficient optical computing architecture for AI tasks.

Main Methods:

• Designed a metasurface device with subwavelength nanostructures for polarization multiplexing.
• Constructed a multi-channel classifier framework for simultaneous recognition.
• Integrated the metasurface with a complementary metal-oxide semiconductor imaging sensor.

Main Results:

• Achieved on-chip multi-channel sensing and multitasking in the visible spectrum.
• Demonstrated simultaneous recognition of diverse items (digital and fashionable).
• Reported high areal density of artificial neurons (up to 6.25 × 10^6 mm^-2 per channel).

Conclusions:

• The developed metasurface-based diffractive neural network enables efficient, chip-scale AI processing.
• This technology facilitates energy-efficient, ultra-fast image processing for applications like machine vision and autonomous driving.
• The multi-skilled optical computing approach mimics brain-like multitasking for advanced AI capabilities.