Pushing the Limits of Metasurface Cloak Using Global Inverse Design

Abstract The breakthroughs of transformation optics and metamaterials have kick‐started the study of modern invisibility cloak since the beginning of this century. Many cloaking methodologies have been progressively proposed for specific application scenarios, among which metasurface cloak is largely welcomed owing to its salient features of negligible thickness, easy fabrication, and low loss. Similar to other cloaking methodologies, however, metasurface cloak suffers from inherent limits that impair it to a convex shape, narrow bandwidth, and small incident angle. Here, a global inverse design is reported to push the limits of metasurface cloak to free form, complementary to conventional physics‐informed approaches. A tandem neural network to build up a bidirectional channel between the metasurface cloak and its electromagnetic response is formulated, in which an ineluctable nonuniqueness issue is mitigated to improve the output accuracy >93%. Compared with conventional metasurface cloak, the foveated cloak underscores intricate coupling and nonlocal effect and widens the bandwidth to 8.5–10.5 GHz and the incident angle to ±45°. These results provide an important step forward to generalizing metasurface cloak and enable a high‐speed surrogate solver required in emerging intelligent meta‐devices.