Polarization‐Multiplexing Microwave Diffractive Deep Neural Network Based on Cascaded Metasurfaces

Abstract Recently, electromagnetic wave‐based computational architectures, such as photonic circuit neural networks and all‐optical diffractive deep neural networks (D 2 NNs), have attracted considerable attention due to their potential for high‐speed processing and low power consumption. However, many existing implementations face challenges related to functional limitations and system complexity, which hinder their broader applicability. In this work, a polarization‐multiplexed D 2 NN architecture operating in the microwave frequency range, capable of simultaneously recognizing handwritten letters and digits under orthogonal polarization states, is proposed. For each polarization channel, distinct sets of metallic input patterns are fabricated and tested, achieving 100% recognition accuracy in experimental validation. Furthermore, the potential of this architecture for parallel computing is investigated and its feasibility is demonstrated through numerical simulations. This study presents a new direction for implementing parallel computing in microwave D 2 NNs and offers a foundation for future exploration of multi‐dimensional multiplexing, with the potential to enhance computational speed and reduce energy consumption.