LLM-enabled cognitive metasurfaces

Metasurface is an ultrathin engineered structure that can flexibly manipulate electromagnetic (EM) waves and is utilized to design advanced device for multifaceted EM functions. Recently, it has shown a fascinating evolving trend toward intelligent EM platform as it is integrated with advanced algorithms. However, from a cognitive perspective, existing metasurface systems still lack the autonomous intelligence in the sense that they rely on human-prespecified instructions and workflows, making them difficult in autonomously responding to changing environments and dynamic user demands. To address this formidable difficulty, we here propose cognitive metasurfaces (CM in short) by exploring modern large language models (LLMs) for their digital-domain cognitive capability and integrating them into programmable metasurfaces for physical-domain cognitive tasks. Thereby, the CM can infer user’s intents, adaptively organize task-specific cognitive modes, and coordinate multiple metasurface-based skills across wireless communication, sensing and countermeasures. Moreover, the CM can autonomously update its control coding strategies through real-time interaction with the physical environment, enabling the self-correction when the predefined coding sequences is mismatched with environmental boundary conditions. Selected experimental results have been provided to show that our CM can complete long-term tasks by understanding natural-language requests, self-organizing long-horizon workflows, and adaptively controlling multi-metasurface system. These capabilities allow the CM to achieve a high level of autonomy, flexibility, and robustness in complex EM scenarios. We faithfully expect that the proposed CM could open a new avenue for designing next-generation intelligent information systems.