A fast computation method for the performance of meta-atoms based on deep learning

Metasurfaces, with flexible electromagnetic wave control and smaller size than traditional devices, offer new potentials for light manipulation. However, the meta atoms of metasurfaces often require repeated electromagnetic simulations and parameter adjustments for design, which is time-consuming. Recently, deep learning has enabled fast computation and inverse design methods. Fast computation accelerates design speed and aids inverse design by validating data and speeding up the training process. However, current fast computation methods face issues such as a narrow range of applicability and discrepancies compared to electromagnetic simulations. In this work, we propose a fast computation method for optimizing meta-atoms based on deep learning. To validate the method, a deep learning network based on feature fusion was trained using 600,000 datasets, achieving a mean squared error (MSE) of less than 4×10−4. This network can rapidly obtain the optical responses of meta-atoms with almost arbitrary shapes within the 400 to 700nm range for TiO₂ and Si₃N₄ materials, all within milliseconds. This method not only accelerates the design of metasurfaces but also expedites the training of other metasurface networks, offering valuable insights for the rapid design of metasurfaces.