Robust Inverse Design of Metasurfaces Under Geometric Damage via Hyperlearning

The accelerated development of intelligent meta-surfaces is reshaping electromagnetic manipulation technologies, spanning applications from wireless communications to invisibility cloaks. Central to these advances lies the precise and efficient inverse design of metasurface global patterns. Yet, unpredictable accidents such as structural deformation exacerbate the inherently ill-posed and non-unique nature of inverse design, severely restricting reliable deployment of metasurfaces in harsh realistic environments and causing customized functions to fail. To overcome this, we propose a geometry-robust global design framework driven by a hyper-learning strategy, which integrates a generation-optimization model. The generator is mainly built upon a conditional variational autoencoder (CVAE), which is augmented with explicit geometric constraints and adversarial loss to reduce mapping ambiguity and produce high-fidelity candidate patterns. The dynamic hidden layer of the generator incorporates a hyper-learning updating approach that modulates weight and bias parameters based on geometry features, embedding geometry-awareness directly into the network inference process and thereby enhancing design diversity and output quality. A pixel-level optimization network is cascaded afterwards, enabling rapid refinement of candidate patterns in a few iterations within millisecond-level latency. Both simulation and experimental results show that the similarity of the far-field is improved by up to around 40% compared with conventional deep learning models, while maintaining a competitive generating time of less than 1 ms. Our approach maintains reliable metasurface functionality under breakdowns and supports electromagnetic restoration in harsh conditions (e.g., satellite communication), meanwhile reducing maintenance costs and extending device longevity.