Electromagnetic metamaterials and artificial intelligence: from intelligent enabling to autonomous intelligence

The interaction between metamaterials and artificial intelligence (AI) has undergone a significant and remarkable evolution, advancing steadily from its initial stages of auxiliary design support to the current frontier of autonomous intelligent systems. This progressive journey began with the fundamental establishment of precise “structure–response” mapping for design assistance, which laid the essential groundwork. It then moved into the realm of transformative hardware innovation through the advent of wave‑based computing paradigms. This phase was followed by a further, distinct development into the domain of real‑time electromagnetic wave control via the integration of active components. Today, this trajectory is decisively converging toward a novel and promising paradigm of intelligent collaboration between meta‑agents and agent-enabled metamaterials. This article provides a systematic and comprehensive review of this deepening fusion, meticulously encompassing its foundational frameworks, recent groundbreaking breakthroughs, and compelling future prospects. Regarding the foundational computational architecture, forward design methodologies effectively leverage a suite of advanced techniques, including gradient‑based optimization, sophisticated heuristic algorithms, and powerful deep‑learning surrogate models, to achieve efficient and accurate structure‑to‑response mapping. For the more challenging and inherently complex inverse design problem, which involves navigating intricate “many‑to‑many” mappings with non‑unique solutions, discriminative and generative AI models offer particularly powerful and flexible solutions. Critically, the integration of cutting‑edge techniques such as physics‑informed learning, strategic data augmentation, and transfer learning has substantially improved the generalization capability, predictive accuracy, and overall reliability of these computational models. Recent breakthroughs vividly highlight several transformative directions shaping the field. First, the development of low‑power consumption wave‑based computing hardware, now being realized through innovative platforms like mechanically programmable metamaterials and highly integrated optical computing architectures, presents a profoundly promising path for next‑generation, energy‑efficient information processing. Second, the strategic incorporation of active components such as varactors, phase‑change materials, and micro‑electromechanical systems enables the real‑time, intelligent, and dynamic reconfiguration of metamaterial properties, allowing for sophisticated adaptation to changing electromagnetic environments and application demands. Third, the cognitive capabilities of agent-enabled metamaterials are being rapidly enhanced. Emerging frameworks that utilize large language models and other AI agents are empowering more advanced “perception–reasoning–execution” closed‑loop systems. Concurrently, dedicated meta‑agents are being developed to automate, optimize, and oversee entire design and optimization workflows, thereby paving a clear way toward the ambitious goal of unsupervised, autonomous metamaterial discovery. Looking ahead, three critical and interconnected dimensions demand focused research to fully realize the immense potential of this synergy: enhancing the robustness, reliability, and interpretability of agent‑based reasoning processes; improving the cross‑scenario transferability and adaptability of learned models and control strategies; and achieving the practical miniaturization and seamless system‑level integration of intelligent metamaterial systems for real‑world deployment. By deliberately delineating the clear technical trajectory from “foundational frameworks” to “frontier breakthroughs” and finally to a reasoned “future outlook”, this review aims to chart a detailed course for the ongoing evolution of the metamaterial‑AI field. Its ultimate goal is to guide the field’s critical transition from being merely augmented by external intelligence toward embodying truly embedded, autonomous, and adaptive intelligence within material systems themselves.