Programmable diffractive deep neural networks enabled by integrated rewritable metasurfaces.

Area of Science:

• Photonics
• Materials Science
• Artificial Intelligence

Background:

• Photonic neural networks require rapid programming for complex functionalities.
• Phase change technology offers nonvolatile programmability for photonic devices.
• Direct laser writing integration with phase change materials (PCM) enables on-chip computing.

Purpose of the Study:

• To demonstrate an ultra-compact, on-chip programmable diffractive deep neural network.
• To utilize Sb2Se3, an ultralow-loss phase change material, for photonic neural networks.
• To achieve high accuracy in machine learning tasks using novel photonic hardware.

Main Methods:

• Cascading multiple layers of on-chip phase-change metasurfaces.
• Utilizing direct laser writing on Sb2Se3 thin films to create programmable metasurfaces.
• Theoretical demonstration at a 1.55 μm wavelength.

Main Results:

• Creation of compact, low-loss, rewritable, and nonvolatile on-chip phase-change metasurfaces.
• Theoretical demonstration of an ultra-compact programmable diffractive deep neural network.
• Achieved accuracies comparable to state-of-the-art on pattern recognition and MNIST classification tasks.

Conclusions:

• Direct laser writing on Sb2Se3 enables advanced on-chip photonic neural networks.
• Cascaded phase-change metasurfaces offer a pathway to powerful in-memory computing.
• This approach achieves high performance for machine learning applications.