Generative-AI-assisted knowledge-based augmented exploration in nanophotonics

It is generally difficult to obtain valuable scientific findings without sufficient background knowledge or experience. Exploiting generative AI (genAI) and reliable computational method(s) strategically is expected to reduce the difficulty and attain scientific findings beyond human knowledge and experience. Here, we introduce a new approach that initially relies on a substantial amount of scientific literature to eliminate hallucinations, prepares metadata from the literature using genAIs, conducts retrieval augmented generation using a genAI, implements genAI-driven generations of concrete candidates of nanostructures, and evaluates the candidates quantitatively, which contain extended sets of structural parameters, by implementing a scientifically established simulation method. This genAI-assisted nonempirical approach has been applied in the field of nanophotonics, enabling the exploration of single-layer circular dichroic (CD) all-dielectric metasurfaces in telecom bands. We successfully revealed single-layer perfect CD metasurfaces, which have never been attained so far. The perfect CD metasurfaces are thin with subwavelength thickness in the near-infrared bands, are feasible in a semiconductor nanofabrication process, and significantly enlarge degrees of freedom in light-wave manipulations for telecom applications.