Reconfigurable Magneto-Optoelectronic Devices for Multidimensional Optical Neural Network.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Materials Science

Background:

• Conventional optical neural networks (ONNs) have limitations in handling high-dimensional data, restricting their application to simple tasks like image classification.
• There is a growing demand for reconfigurable ONNs capable of perceiving and processing inherent high-dimensional light information.

Purpose of the Study:

• To theoretically propose a novel magneto-optoelectronic device for constructing ONNs with high-performance multidimensional recognition.
• To leverage polarization sensitivity and switchable magnetic configurations for advanced computational capabilities.

Main Methods:

• Proposed a device composed of 2D magnetic half-metal FeCl2 and 2H-WSe2, utilizing the photogalvanic effect in 2H-WSe2 for polarization sensitivity.
• Employed switchable magnetic configurations of FeCl2 contacts to modulate photoresponse nonvolatilely across UV to near-infrared wavelengths.
• Developed a multidimensional light encoding strategy for the ONN architecture.

Main Results:

• Achieved multidimensional perception under zero power consumption due to broken space-inversion symmetry in 2H-WSe2.
• Demonstrated nonvolatile modulation of photoresponse amplitude and polarity via switchable magnetic configurations.
• Attained up to 93.5% accuracy in complex tasks, including 3D object classification and time-series recognition.

Conclusions:

• The proposed magneto-optoelectronic ONN architecture enables negative value and nonlinear computations in the polarization domain.
• This work highlights the potential of magneto-electronics to significantly extend the real-world applications of ONNs.
• The developed device offers a pathway towards more powerful and versatile optical neural networks.