Programmable Three-dimensional Photonic Neural Network Chip.

Area of Science:

• Photonics and Artificial Intelligence
• Integrated Optics and Neuromorphic Computing

Background:

• Photonic neural networks offer advantages in speed and energy efficiency but are limited by 1D input interfaces and crosstalk.
• Serialization of 2D image data into 1D streams creates I/O bottlenecks, hindering scalability and spatial parallelism.

Purpose of the Study:

• To develop a programmable 3D photonic neural network chip capable of directly processing 2D images.
• To overcome the limitations of existing planar photonic platforms and I/O bottlenecks.

Main Methods:

• Fabrication of a 3D photonic neural network chip using femtosecond laser direct writing (FLDW) in glass.
• Implementation of a cascaded architecture with alternating photonic-lantern waveguide arrays and phase-shifter arrays for matrix operations.

Main Results:

• Demonstration of an 8-layer 8x8 device achieving 6554 TOPS computing throughput.
• Achieved 93% accuracy on MNIST classification and 94% fidelity in optical pattern generation.
• Outperformed leading planar photonic platforms in computing throughput.

Conclusions:

• The developed 3D photonic chip directly processes 2D images, enabling true spatial parallelism.
• The combination of 3D architecture and programmability offers a scalable paradigm for reconfigurable photonic computing.
• This approach addresses key challenges in photonic neural network scalability and performance for complex inference tasks.