Improved all photonics diffraction neural network based on multi-channel integrated optical fibers

Abstract

AI’s exponentially growing computational demands conflict with slow hardware advances. The high-power consumption and long training times of large-scale models call for alternative solutions. Optical computing-based traditional optical networks and diffractive deep neural network (D²NN) still face deployment challenges and reliance on electronic networks. To address these issues, we replace the free-space interlayer propagation in conventional optical networks with fiber-based propagation. This preserves the advantages of traditional optical networks while providing additional benefits such as ease of deployment, reduced dependence on electronic networks, and enhanced robustness. Experimental results demonstrate that this untrained structure exhibits strong nonlinear mapping capabilities across different configurations, yielding distinct outputs for three input targets, especially at 1550 nm. Furthermore, the influence of environment and noise is around 1% in target recognition. Leveraging inherent spectral discrimination, this architecture enables multidimensional target identification with important implications for complex target classification and multidimensional sensing.