Coupling Enhanced Diffractive Deep Neural Network with Structural Nonlinearity

Abstract

The increasing complexity of deep learning models poses stringent requirements on electronic computers. Diffractive deep neural networks (D²NNs), as one of the most representative optical computing architectures, have emerged as a significant substitute for electronic-based devices due to the advantages of high speed, low power consumption, and high parallelism. However, the absence of optical nonlinearity constrains the potential advancement of D²NNs. Recent progress in structural nonlinearity has offered a promising avenue for addressing this issue, but it necessitates complex digital data pre-encoding. Herein, structural nonlinearity is introduced into D²NNs by incorporating encoding-free data repetition layers, enabling high-order optical nonlinearity while reducing the system complexity. The effectiveness of different data repetition manners demonstrates the robustness of this approach. Additionally, to enhance the design accuracy of D²NNs, a graph neural network framework is developed to characterize the coupling effects in metasurface layers and integrate it into D²NNs. This work provides a novel approach for the design of optical computing devices and holds significant importance for the development of high-performance and highly integrated all-optical devices.