Multi-wavelength diffractive optical neural network integrated with 2D photonic crystals for joint optical...

Area of Science:

• Photonics
• Machine Learning
• Optical Computing

Background:

• Optical neural networks (ONNs) offer advantages over electronic computing due to high parallelism, bandwidth, and low power consumption.
• On-chip diffractive optical neural networks (DONNs) are key for integrated, energy-efficient machine learning.
• Current DONNs are limited by single-wavelength operation, restricting computational parallelism.

Purpose of the Study:

• To propose and demonstrate a multi-wavelength visual classification architecture, PhC-DONN, for enhanced computational throughput.
• To leverage wavelength as a degree of freedom for multidimensional multiplexing in diffractive computing.
• To establish a novel optical classification paradigm for multi-wavelength optical neural networks.

Main Methods:

• Integration of two-dimensional photonic crystal (PhC) components with diffractive computing units.
• Development of a PhC convolutional layer for multi-wavelength feature extraction.
• Implementation of a three-stage diffraction layer for parallel optical field modulation and a PhC nonlinear activation layer for wavelength nonlinear computation.

Main Results:

• PhC-DONN achieved high classification accuracies: 99.09% on MNIST, 66.41% on CIFAR-10, and 92.25% on KTH.
• The architecture demonstrated a 32-fold enhancement in computational throughput compared to conventional DONNs.
• Multi-channel inference was achieved in a single light propagation pass via wavelength-parallel classification.

Conclusions:

• The proposed PhC-DONN successfully implements a multi-wavelength classification mechanism, significantly boosting computational throughput and accuracy.
• This work presents a viable pathway for constructing large-scale photonic intelligence parallel processors.
• The novel architecture advances optical classification paradigms for multi-wavelength optical neural networks.