The diamond mesh, a phase-error- and loss-tolerant field-programmable MZI-based optical processor for optical neural...

Area of Science:

• Photonics
• Optical Computing
• Machine Learning Hardware

Background:

• Integrated photonic devices are crucial for optical neural networks (ONNs).
• Minimizing performance degradation due to experimental imperfections like phase errors and insertion loss is a key challenge.
• Existing mesh architectures may not offer optimal robustness against these imperfections.

Purpose of the Study:

• To present a performance analysis of a novel phase error- and loss-tolerant multiport MZI-based structure for ONNs.
• To compare the proposed diamond mesh topology with the conventional triangular (Reck) mesh.
• To evaluate the robustness and scalability of the diamond mesh for implementing ONNs.

Main Methods:

• Development and theoretical analysis of a diamond mesh architecture using Mach-Zehnder Interferometers (MZIs).
• Comparison of the diamond mesh with the triangular (Reck) mesh in terms of MZI count, topology, and degrees of freedom.
• Analytical evaluation of performance under simulated phase errors and insertion loss.
• Assessment of classification accuracy for different ONN sizes.

Main Results:

• The diamond mesh utilizes more MZIs, resulting in a symmetric topology and enhanced optimization capabilities for weight matrices.
• Additional MZIs in the diamond mesh effectively eliminate excess light intensity via tapered waveguides.
• The diamond topology exhibits significantly higher robustness against insertion loss and phase errors compared to the triangular mesh.
• Improved classification accuracy was observed for the diamond mesh in the presence of experimental imperfections.

Conclusions:

• The proposed diamond mesh architecture offers superior tolerance to phase errors and insertion loss in MZI-based ONNs.
• This enhanced robustness translates to better classification accuracy, making it suitable for practical implementations.
• The diamond mesh demonstrates practical performance and scalability for various sizes of optical neural networks.