Braided interferometer mesh for robust photonic matrix-vector multiplications with non-ideal components.

Area of Science:

• Photonics and optical engineering
• Quantum computing and machine learning hardware
• Interferometric device design

Background:

• Matrix-vector multiplications (MVMs) are fundamental operations in machine learning and quantum computing.
• Photonic architectures offer high speed, low latency, and minimal loss for linear operations.
• Existing designs like Clements and Fldzhyan face scalability challenges in stability and robustness.

Purpose of the Study:

• Introduce and evaluate a novel photonic braid interferometer architecture for MVMs.
• Compare the braid architecture's performance against Clements and Fldzhyan designs.
• Assess robustness and scalability under realistic non-ideal conditions.

Main Methods:

• Numerical simulations to evaluate photonic interferometer architectures.
• Systematic introduction of non-idealities: insertion losses, beam splitter imbalances, crosstalk.
• Performance, footprint, and insertion loss analysis of braid, Clements, and Fldzhyan designs.

Main Results:

• The photonic braid architecture demonstrates superior robustness and performance over Clements and Fldzhyan designs.
• Braid architecture shows enhanced scalability and better performance with increasing interferometer size.
• Despite minor increases in footprint and loss due to crossings, recent technological advances mitigate these effects.

Conclusions:

• The photonic braid interferometer is a robust and high-fidelity solution for large-scale photonic neuromorphic computing.
• Its symmetrical design and reduced layer count contribute to superior performance in realistic, imperfect conditions.
• This architecture represents a significant advancement for photonic linear operations in demanding applications.