Microdisk resonator crossbar array using phase change material for an optical neural network.

Area of Science:

• Photonics
• Artificial Intelligence
• Materials Science

Background:

• Increasing computational demands of AI necessitate faster, more energy-efficient hardware.
• Electronic accelerators face limitations in speed and power consumption.
• Photonic neural networks (PNNs) offer a promising alternative due to high speed, low latency, and energy efficiency.

Purpose of the Study:

• To present a novel silicon-based microdisk resonator (MDR) crossbar array for optical neural network applications.
• To integrate phase change material (PCM) Ge2Sb2Se4Te1 (GSST) for nonvolatile and programmable weight control.
• To demonstrate a scalable and efficient architecture for on-chip photonic computing.

Main Methods:

• Design of a low-loss, compact multimode interference (MMI) crossing waveguide with a 70-degree offset.
• Integration of GSST onto silicon-based MDRs for programmable optical weights.
• Demonstration of a 3x3 MDR array for photonic matrix-vector multiplication and image convolution tasks.

Main Results:

• The MMI crossing waveguide achieved a compact footprint (1 µm²) and reduced device area by 2.03x while maintaining low insertion loss and crosstalk.
• Nonvolatile and programmable weight control was achieved by modulating the crystallization degree of GSST.
• The 3x3 MDR array demonstrated a high computational efficiency of 1.2x10¹² MAC/s/mm² and scalability for image convolution and fall detection tasks.

Conclusions:

• The developed GSST-integrated silicon MDR crossbar array offers a scalable and efficient architecture for on-chip photonic computing.
• This approach addresses the computational demands of AI by leveraging the advantages of photonic neural networks.
• The demonstrated system shows significant potential for future optical computing applications.