Inverse design of an integrated-nanophotonics optical neural network

Area of Science:

• Photonics
• Machine Learning
• Computer Engineering

Background:

• Artificial neural networks (ANNs) excel in machine learning tasks like image recognition.
• Electronic hardware for ANNs faces performance limitations due to Moore's Law slowing.
• Current hardware struggles with speed, power consumption, and size for complex deep learning.

Purpose of the Study:

• To propose a novel optical neural network (ONN) architecture.
• To overcome the limitations of electronic hardware for deep learning.
• To achieve high-speed, low-power, and compact deep learning implementations.

Main Methods:

• Developed an ONN architecture utilizing optical scattering units.
• Optimized optical scattering units using an inverse design method.
• Implemented a "Kernel Matrix" framework for the ONN.

Main Results:

• Optical scattering units achieved high-precision stochastic matrix multiplication (MSE <10⁻⁴).
• Each unit has a compact footprint (4 × 4 μm²).
• The ONN framework achieved 97.1% accuracy on the MNIST dataset.

Conclusions:

• The proposed ONN architecture offers a promising alternative to electronic hardware.
• Optical scattering units enable efficient and accurate deep learning computations.
• This approach paves the way for next-generation AI hardware.