Integrated photonic neural network based on silicon metalines

Area of Science:

• Photonics
• Artificial Intelligence
• Materials Science

Background:

• Digital neural networks face limitations in speed and energy consumption.
• Integrated photonics offers a potential solution for high-performance computing.
• Metasurfaces provide a novel platform for optical computation.

Purpose of the Study:

• To propose and demonstrate an on-chip photonic neural network.
• To leverage cascaded one-dimensional metasurfaces (metalines) for neural network functions.
• To achieve compact, low-power, and high-speed optical computing.

Main Methods:

• Designing and fabricating high-contrast transmitarray metalines on a silicon-on-insulator substrate.
• Utilizing the intrinsic parallelism and low-loss properties of silicon metalines.
• Implementing matrix-vector multiplications optically for neural network operations.

Main Results:

• Demonstrated a compact, whole-passive, fully-optical meta-neural-network.
• Achieved parallel matrix-vector multiplication with low energy consumption.
• Attained classification accuracy comparable to state-of-the-art on the MNIST dataset at 1.55 µm wavelength.

Conclusions:

• The proposed on-chip photonic neural network offers a promising alternative to digital neural networks.
• This technology enables high-speed, low-energy optical computation for complex functions.
• Further development could lead to significant advancements in artificial intelligence hardware.