Review maps how large language models enter nanophotonic design as surrogate models and agentic systems

The review organizes LLM-related methods in nanophotonics into two operational modes: surrogate models that treat structure-spectrum mapping as a language task, and agentic systems that have been demonstrated to generate code, orchestrate selected simulation steps, and support closed-loop optimization. Prior deep learning mainly served as task-specific surrogate models constrained by specific architectures and lacking universal reasoning; this review introduces semantic interfaces, code generation, and tool orchestration as a new organizing dimension for established numerical workflows. This is a review-level synthesis based on the surveyed literature rather than a single experiment; the abstract presents the classification framework without listing case counts or performance figures in the visible text.

The review outlines the development from classical neural networks to transformer-based models in nanophotonic design, noting that metasurface design has long been constrained by computationally expensive simulations and complex high-dimensional design spaces. It situates the emergence of LLMs within this evolutionary sequence, indicating that their difference from task-specific surrogate models lies in universal reasoning and semantic interaction. A narrative summary of the field's development, grounded in the text's description of metasurface design bottlenecks and the role of deep learning.

The review briefly examines LLM applications in other research fields such as materials science and wireless communications to identify future cross-disciplinary opportunities, and looks ahead to next-generation multimodal foundation models with physical perception capabilities. It places nanophotonic progress alongside adjacent fields and proposes a vision in which AI evolves from a passive tool into an active collaborator participating in autonomous scientific discovery. A forward-looking outlook stated in the abstract, without specific validation results provided.

The review targets researchers and practitioners in nanophotonics and photonic device design, and suits readers who want to understand how LLMs can be embedded into established numerical simulation workflows; its classification framework is meant to help readers locate surrogate-model and agentic approaches and plan work in code generation, tool orchestration, and closed-loop optimization.

Because the currently visible text is an incomplete abstract, lacking specific cases, method details, performance data, and references from the body, the maturity and degree of validation of work under each category cannot be judged; multimodal foundation models with physical perception and autonomous scientific discovery remain outlooks whose feasibility and applicable conditions await further study.