Review for metasurface holographic imaging enabled by artificial intelligence technology

Abstract

Metasurface holography provides a compact route for controlling amplitude, phase, polarization, and other degrees of freedom of electromagnetic waves, enabling highly integrated, multiplexed, and reconfigurable holographic systems. As device functionality expands toward broadband, multichannel, dynamically tunable, and large-area aperiodic operation, however, design increasingly becomes a high-dimensional, multi-constraint problem that couples meta-atom geometry, electromagnetic response, diffractive propagation, and final image quality. This review develops a function-driven framework for artificial-intelligence-assisted metasurface holography that connects target optical functionality, physical realization, forward modeling, inverse design, full-device modeling, and system-level reconstruction. We first compare visible, terahertz, and microwave implementations in terms of material platforms, fabrication constraints, data acquisition, and validation requirements, and then review phase, amplitude, complex-amplitude, and nonlinear modulation, multidimensional multiplexing, and dynamic reconfiguration. Particular attention is given to when AI provides practical value relative to conventional electromagnetic simulation and numerical optimization, and to the transition from isolated or periodic unit-cell models toward neighborhood-aware and full-device descriptions. Recent advances in forward surrogate modeling and inverse design are examined, with deterministic approaches distinguished from generative multi-solution methods, including tandem networks, variational autoencoders, generative adversarial networks, normalizing flows, diffusion models, and physics-driven generative frameworks. Representative studies are quantitatively compared in terms of dataset size, accuracy, computational cost, device performance, and experimental validation. We further discuss system-level co-design and engineering trade-offs among efficiency, information capacity, crosstalk, response speed, fabrication tolerance, computational cost, and experimental complexity. Finally, key opportunities are identified in cross-platform foundation models, multimodal inverse design, scalable full-device modeling, fabrication-aware and experiment-in-the-loop optimization, and end-to-end system-level co-design.