Low-phase quantization error Mach–Zehnder interferometers for high-precision optical neural network training

Abstract

A Mach–Zehnder interferometer is a basic building block for linear transformations that has been widely applied in optical neural networks. However, its sinusoidal transfer function leads to the inevitable dynamic phase quantization error, which is hard to eliminate through pre-calibration. Here, a strongly overcoupled ring is introduced to compensate for the phase change without adding perceptible loss. Two full-scale linearized Mach–Zehnder interferometers are proposed and experimentally validated to improve the bit precision from 4-bit to 6- and 7-bit, providing ∼3.5× to 6.1× lower phase quantization errors while maintaining the same scalability. The corresponding optical neural networks demonstrate higher training accuracy.