Reviews of algorithm-driven terahertz metamaterials: From intelligent design to multidisciplinary applications

Terahertz metamaterials, composed of subwavelength artificial structures, exhibit strongly nonlinear and tightly coupled electromagnetic responses governed by geometry, material properties, and resonance modes. Driven by growing demands in high-sensitivity sensing, terahertz communication, and functional imaging, the design space of these devices has rapidly expanded, rendering conventional empirical and parameter-sweeping approaches inefficient for global optimization. Algorithm-driven strategies, particularly those leveraging machine learning and deep learning, have emerged as powerful surrogates for electromagnetic simulation, enabling automated multi-objective optimization and rapid inverse design. Beyond predictive capabilities, these algorithms uncover latent physical mechanisms and support adaptive, closed-loop experimental implementations. Within a unified framework, this review systematically categorizes and summarizes advances in algorithm-driven design of terahertz metamaterials, including traditional optimization methods, machine learning, deep learning, and reinforcement learning, highlighting their roles in multi-layer structural design, multimodal coupling control, and dynamic multi-target detection. Finally, we conduct a detailed analysis of current challenges, including data quality, model generalization, and physical interpretability, and clearly identify future research directions. These directions are expected to lead to practical applications in high-sensitivity chemical and biological detection, terahertz wireless communication, and wave-based functional imaging.