Recently, Professor Wei Zhun’s group and Professor Yin Wenyan’s group published a research paper in Nature Machine Intelligence—the journal’s first paper in the field of AI for electromagnetics—with the work also featured in the journal’s News & Views section. The study introduces an AI-driven approach for directed synthesis of dynamic current distributions in spatial-frequency domain, with experimental validation on multi-functional, high-dimensional artificial electromagnetic structure design.
Artificial electromagnetic structures have important application value in integrated platforms for new stealth components, smart skins, and shared-aperture electromagnetic windows. Traditional design methods rely on extensive full-wave simulations, facing significant challenges particularly in high-dimensional, multi-objective design contexts. While deep learning approaches based on forward surrogate models and generative models have demonstrated strong capabilities in accelerating inverse design, two major challenges remain: the non-uniqueness problem, where different topological structures may yield identical or similar electromagnetic responses that can lead to convergence difficulties, and the difficulty of accurately representing and learning physical features based on topological encoding.
The research team proposed a spatial-frequency dynamic current diffusion modeling framework called MetaAI. By predicting the interaction process between electromagnetic waves and artificial electromagnetic structures—using dynamic current in spatial-frequency domain as an intermediate bridge between electromagnetic response and topological structure—the framework achieves rapid design of both single-layer and multilayer free-form structures for both in-distribution and out-of-distribution performance requirements. In single-layer design cases, MetaAI successfully reconstructed the dynamic current process in spatial-frequency domain under out-of-distribution conditional inputs, with the resulting dynamic processes matching full-wave simulation results and achieving performance beyond the training set bounds under the same constraints. When integrated with active device layers, the framework enabled rapid design of tunable artificial electromagnetic structures and absorption-transmission integrated multilayer structures, with experimental results validating the design framework’s effectiveness.
Unlike topology encoding-based generation approaches, the innovation lies in generating spatial-frequency domain dynamic current processes, where current distribution serves as the physical mechanism directly linking topological structure to electromagnetic response. PhD candidates Li Erji and Wang Yusong are co-first authors; Professor Wei Zhun and Professor Yin Wenyan are co-corresponding authors.
Link: https://www.nature.com/articles/s42256-025-01162-z
