Magneto-optical diffractive deep neural network.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Materials Science

Background:

• Diffractive deep neural networks (D²NNs) offer a promising avenue for optical computing.
• Existing D²NN implementations lack reconfigurability in their hidden layers.
• Magneto-optical effects provide a tunable mechanism for optical information processing.

Purpose of the Study:

• To propose and simulate a novel magneto-optical diffractive deep neural network (MO-D²NN).
• To evaluate the classification performance of MO-D²NNs using the MNIST dataset.
• To investigate the impact of Faraday rotation on classification accuracy.

Main Methods:

• Simulation of MO-D²NNs with five hidden layers composed of magnetic materials.
• Utilizing 100 × 100 magnetic domains with a 1 µm domain width and 0.7 mm interlayer distance.
• Employing light intensity and polarization angle as classification measures.

Main Results:

• Achieved > 90% classification accuracy on the MNIST dataset using light intensity.
• Obtained > 80% accuracy even with a small Faraday rotation angle (π/100 rad) using polarization angle.
• Demonstrated the reconfigurability of hidden layers, a key advantage over previous D²NNs.

Conclusions:

• MO-D²NNs represent a significant advancement in reconfigurable optical neural networks.
• The proposed architecture shows high potential for efficient and tunable optical classification tasks.
• The ability to rewrite hidden layers enhances the versatility and applicability of diffractive deep neural networks.