Frequency-multiplexing Microwave Diffractive Neural Network Based on Metasurface

Microwave diffractive deep neural networks (D²NNs) are often constrained by limited network capacity and large physical dimensions due to millimeter-scale wavelength. To address the bottlenecks, a dual-frequency D²NN architecture with complex-amplitude modulation is proposed in this work. By exploiting spectral parallelism and independent complex-amplitude modulation at two distinct frequencies, the proposed architecture significantly expands the learnable degrees of freedom without increasing the system depth. A frequency-space joint multiplexing strategy and a physics-aware cross-talk suppression training method are introduced to enable high-performance cross-frequency training and inference capability. Numerical evaluation reports an accuracy of 83.43% based on a single layer 32×32 diffractive network on MNIST classification task. Further full-wave simulations are implemented through a transmissive dual-frequency complex-amplitude metasurface to validate the physical feasibility of proposed network architecture, offering a compact and efficient solution for intelligent electromagnetic perception.