Agentic AI for inverse design in nanophotonics and meta-optics

Nanophotonic inverse design requires coordinated choices about geometry, optimization, and electromagnetic simulation. We present an agentic AI framework that combines literature-based reproduction, human-approved optimization strategies, adjoint optimization, and separate forward evaluation of binary candidates. Four metagrating strategies tested from a common initialization yield final diffraction efficiencies of 79.4–90.9%. Selecting by peak optimization objective and by final binary efficiency produces different rankings. A separate metagrating redesign increases diffraction efficiency from 81.2% to 94.6%. For a two-dimensional metalens, the minimum focusing efficiency across three visible wavelengths increases from 40.0% to 65.2% relative to a symmetrized reproduced reference under matched lossless evaluation. A Bayer color router improves mean channel routing from 50.3% to 54.7% under a 125 nm minimum-feature criterion. Separate workflow evaluations include five reproducibility-screening examples, seven literature-to-simulation cases, and repair outcomes across five solver errors. Solver-specific retrieval resolves an interface error left unresolved by every tested standalone configuration. These results demonstrate agent-assisted exploration of nanophotonic designs and show why candidate selection must use the optical performance of the final binary structure.