Hardware-efficient photonic neural network processor by mode-division multiplexing

In this study, we propose and demonstrate a hardware-efficient photonic neural network processor utilizing low-rank approximation (LRA) in conjunction with mode-division multiplexing (MDM) techniques. By exploiting MDM, the architecture requires only a single optical source, substantially reducing the hardware complexity associated with multi-source systems. Through LRA, the proposed architecture integrates two single-mode Mach-Zehnder interferometers (MZIs) and four mode-insensitive MZIs to construct a 4 × 2 matrix with the rank of 1. Compared to previously photonic neural network processors employing MZIs or microring resonators (MRRs), our approach achieves a 25% reduction in the number of optical components through LRA for the 4 × 2 matrix configuration. Fabricated on an 800 nm silicon nitride platform, the MDM-enabled LRA photonic neural network processor was experimentally validated on a facial expression recognition task and achieved a prediction accuracy of 96% on the MNIST dataset. Our work provides a scalable and hardware-efficient pathway toward photonic matrix operations, particularly well-suited for resource-constrained applications.