Kang Chanik, Seo Joonhyuk, Jang Ikbeom, Chung Haejun
Department of Artificial Intelligence, Hanyang University, Seoul 04763, South Korea.
Department of Computer Engineering, Hankuk University of Foreign Studies, Yongin 17035, South Korea.
iScience. 2024 Dec 6;28(1):111545. doi: 10.1016/j.isci.2024.111545. eCollection 2025 Jan 17.
We present a Fourier neural operator (FNO)-based surrogate solver for the efficient optimization of wavefronts in tunable metasurface controls. Existing methods, including the Gerchberg-Saxton algorithm and the adjoint optimization, are often computationally demanding due to their iterative processes, which require numerical simulations at each step. Our surrogate solver overcomes this limitation by providing highly accurate gradient estimations with respect to changes in tunable meta-atoms without the need for direct simulations. This approach substantially reduces both computational time and cost in wavefront shaping applications. The proposed solver demonstrates a residual of 0.02 when compared to the normalized figure of merit achieved by the optimized structure obtained through the adjoint method, and its inference time is 887.5 times faster than conventional simulation-based methods. This advancement enables ultra-fast wavefront shaping across a range of applications, including optical wavefront shaping, reconfigurable intelligent metasurfaces, and biomedical imaging.
我们提出了一种基于傅里叶神经算子（FNO）的替代求解器，用于在可调谐超表面控制中高效优化波前。包括格尔奇伯格 - 萨克斯顿算法和伴随优化在内的现有方法，由于其迭代过程，通常计算量很大，每个步骤都需要进行数值模拟。我们的替代求解器通过提供关于可调谐超原子变化的高精度梯度估计来克服这一限制，而无需直接模拟。这种方法在波前整形应用中大幅减少了计算时间和成本。与通过伴随方法获得的优化结构所实现的归一化品质因数相比，所提出的求解器的残差为0.02，其推理时间比传统的基于模拟的方法快887.5倍。这一进展使得在包括光波前整形、可重构智能超表面和生物医学成像在内的一系列应用中实现超快速波前整形成为可能。
