Neural Optimizer for Inverse Design of Complex‐Modulated Hologram Implemented by Plasmonic Metasurfaces

Inverse design of a metasurface involves searching parameters in a high-dimensional space, which needs huge computational power. To ease the computational burden, neural network, a well-researched computer science stream, has demonstrated its potential usage in the inverse design of a photonic device. Many studies primarily focused on the nanostructure's configuration. However, the near field of a metasurface requires further optimization to achieve a desired holographic image at relieved computational power. Here, a convolutional neural network is developed to optimize the complex field of a computer-generated hologram, which can be fabricated into a metasurface to generate the desired holographic image upon lensless image projection. The neural optimizer can accelerate the design speed 400 times faster than theoretical computation and reduce the time complexity from O(n(4)) to O(n(2)). The neural optimizer has been compared against three other methods, e.g., gradient-based optimization, global genetic algorithm, and coupled-mode theory, to demonstrate a lowered error rate from more than 10% to 1.38% for the benchmark testing and a reduced running time from hours to near 1 s. The neural optimizer is envisioned to play a key role in lensless image projection and real-time metasurface pattern design.