Phase Smoothing for Mutual Coupling Mitigation in Multifunctional Metasurfaces Designed with Diffractive Neural Networks

Abstract Multifunctional metasurfaces hold great promise for next‐generation communication systems, enabling spectral efficiency, dynamic adaptability, and hardware scalability for performing multiple wavefront manipulation tasks. However, traditional designs often neglect critical near‐field interactions, leading to performance degradation in complex aperiodic configurations. In this work, an AI‐driven design model is presented for a single‐layer, passive, transmission‐mode multifunctional metasurface capable of projecting different holographic patterns at 86 GHz (W‐band) through spatial multiplexing of the input excitation. The approach integrates a diffractive neural network with an optimization algorithm and a phase‐smoothing regularization scheme to mitigate mutual coupling between unit cells. The proposed model refines the phase profile to balance target performance and physical constraints, enhancing the validity of the local periodicity approximation. The experimental results closely match the simulations, validating the proposed method. This work demonstrates that an AI‐driven model, enhanced by physical regularization, offers a scalable and efficient path toward high‐performance multifunctional metasurfaces, enabling dynamic‐like functionalities without active components—crucial for next‐generation wireless communication systems.