The paper describes MetaAI, a physics-aware generative framework designed to automate the discovery of advanced metasurface structures. Unlike traditional methods that map desired performance directly to physical geometry, this model utilizes a current-diffusion transformer to estimate electrical current distributions across both spatial and frequency domains. By leveraging the rich physical information found in these currents, the system can generate non-intuitive architectures that surpass the performance limits of existing training data. This approach demonstrates high flexibility through a dynamic input form, allowing it to design single-layer, multilayer, and tunable metasurfaces with superior bandwidth. Ultimately, the framework facilitates out-of-distribution generalization, enabling researchers to locate optimal electromagnetic designs that were previously inaccessible through conventional optimization techniques.

References:

• Li E, Wang Y, Jin L, et al. Current-diffusion model for metasurface structure discoveries with spatial-frequency dynamics[J]. Nature Machine Intelligence, 2025: 1-11.