Multi-wavelength diffractive neural network with the weighting method.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computational Science

Background:

• Diffractive deep neural networks (D²NNs) offer high speed, low power, and scalability for computational tasks.
• Traditional D²NNs are typically designed for monochromatic light sources.
• The need for D²NNs capable of operating under multi-wavelength illumination is growing.

Purpose of the Study:

• To propose and demonstrate a framework for a multi-wavelength diffractive deep neural network (MW-D²NN).
• To enable D²NNs to perform computations using multiple wavelengths of light simultaneously.
• To adapt D²NNs for applications requiring broadband or multi-color light sources.

Main Methods:

• Developed a novel MW-D²NN framework utilizing a method of weight coefficients.
• Implemented a training strategy where each wavelength is assigned a specific weighting.
• Constructed a wavelength weighting loss function based on the output planes of different wavelengths.

Main Results:

• The trained MW-D²NN successfully classified handwritten digit images under multi-wavelength incident beams.
• A 3-layer MW-D²NN achieved a simulation classification accuracy of 83.3%.
• A 1-layer MW-D²NN demonstrated simulation and experimental classification accuracies of 71.4% and 67.5% respectively at RGB wavelengths.

Conclusions:

• The proposed MW-D²NN framework effectively extends diffractive deep neural network capabilities to multi-wavelength operations.
• The MW-D²NN shows potential for integration into intelligent machine vision systems operating under multi-wavelength and incoherent illumination.
• This research paves the way for more versatile and robust diffractive computing applications.