Improved all photonics diffraction neural network based on multi-channel integrated optical fibers.

Area of Science:

• Photonics and optical engineering
• Artificial intelligence hardware
• Machine learning infrastructure

Background:

• Artificial intelligence (AI) models face escalating computational demands, outstripping current hardware advancements.
• High power consumption and extended training durations of large-scale AI models necessitate alternative computing paradigms.
• Existing optical computing solutions, including diffractive deep neural networks (D²NNs), present deployment challenges and dependency on electronic networks.

Purpose of the Study:

• To propose and validate a novel fiber-based optical network architecture for AI computation.
• To overcome the limitations of free-space propagation in conventional optical networks.
• To enhance the efficiency, robustness, and deployability of optical AI computing.

Main Methods:

• Replaced free-space interlayer propagation with fiber-based propagation in optical networks.
• Experimental validation of the fiber-based optical network's nonlinear mapping capabilities.
• Assessment of environmental and noise influences on target recognition performance.

Main Results:

• The fiber-based architecture demonstrates robust nonlinear mapping capabilities across various configurations.
• Distinct outputs were achieved for three input targets, particularly at 1550 nm wavelength.
• Environmental factors and noise minimally impacted target recognition, with an approximate 1% influence.

Conclusions:

• The proposed fiber-based optical network offers a practical and robust solution for AI computation.
• The architecture leverages spectral discrimination for effective multidimensional target identification.
• This approach has significant implications for complex target classification and advanced multidimensional sensing applications.