Artificial intelligence-enabled metasurface design: a functional perspective

AI has accelerated metasurface design, but its practical impact hinges on ensuring data quality, physical consistency, and agreement between simulation and experiment. This challenge is amplified by the complex interplay of geometry, resonance, coupling, and fabrication constraints, while key objectives such as efficiency, bandwidth, tolerance, and multifunctionality often compete. This review organizes AI-assisted metasurface design by the role of learning in the design loop. We cover forward modeling, inverse and generative design for non-unique solution spaces, hybrid optimization with electromagnetic solvers, end-to-end optical-computational co-design, and physics-embedded learning across these workflows. Representative applications include metalenses, holography, spectral and polarization devices, reconfigurable metasurfaces, and computational imaging. We further highlight critical challenges, including data quality, limited generalization, simulation–experiment mismatch, fabrication tolerance, and inconsistent evaluation. The resulting framework clarifies where learning can accelerate metasurface design and where physical verification remains essential for practical deployment.