On-chip photonic diffractive optical neural network based on a spatial domain electromagnetic propagation model

Area of Science:

• Photonics
• Optical Computing
• Machine Learning Hardware

Background:

• Traditional machine learning relies on electronic hardware, facing limitations in speed and power consumption.
• Integrated optical neural networks offer a promising alternative for high-performance, low-power computation.

Purpose of the Study:

• To propose and demonstrate an integrated physical diffractive optical neural network (DONN) on a silicon-on-insulator (SOI) substrate.
• To enable compact, passive, fully-optical machine learning capabilities.
• To ensure manufacturability and scalability for on-chip neural network design.

Main Methods:

• Designing the DONN structure using a spatial domain electromagnetic propagation model.
• Optimizing neuron value mapping for consistency between pre-trained values and SOI implementation.
• Utilizing a standard silicon-on-insulator (SOI) fabrication process.

Main Results:

• A compact, fully-passive, optical neural network architecture was successfully designed and simulated.
• The model ensures manufacturability and scalability for on-chip integration.
• Numerical demonstration on the UCI Heart Disease Dataset achieved accuracy comparable to state-of-the-art methods for coronary heart disease prediction.

Conclusions:

• The proposed integrated DONN on SOI is a viable platform for efficient optical machine learning.
• This approach facilitates the design and manufacturing of scalable on-chip neural networks.
• The DONN shows significant potential for real-world applications, such as medical diagnostics.