Metasurface-Based Diffractive Optical Networks With Dual-Channel Complex Amplitude Modulation

Abstract

The diffractive optical network(DON) shows promise in addressing the limitations of traditional electronic computation with its high parallelism, ultra-fast speed, and low energy consumption. However, it currently lacks of integrated methods to enhance neural network completeness, making it less suitable for complex tasks. In this paper, we propose a DON architecture with dual polarization channel and complex amplitude modulation. This neural network (named DC CA DON) makes full use of the metasurface's high degrees of freedom, by independent control of the amplitude and phase on orthogonal optical polarization channels, and achieves feature fusion on these channels using linear polarizers, significantly increasing network completeness and prediction accuracy without adding light paths or diffraction layers. DC CA DON with fewer number of diffraction layers can achieve better results. For example, the 3-layer DC CA DON shows an improvement in accuracy of 2.70% and 4.73% over the 11-layer DON when trained on the MNIST and FashionMNIST datasets, respectively. Guided by this principle, we also develope a DON which combines modulation and residual channels into a single light path (named CA Res DON). This network exhibits enhanced predictive capability and enables deeper training.