Sharpness‐Aware Minimization for Robust Inverse Design of Nonlocal Thin Films

ABSTRACT

Multilayer thin films can facilitate on‐demand optical response engineering in angle, wavelength, and polarization, which underpins emerging technologies in nonlocal flat optics. Nonetheless, their optimization in high‐dimensional target and structural parameter spaces is challenging, as minor fabrication errors may lead to significant performance degradation. This work presents a sharpness‐aware gradient descent (SAGD) method for the robust inverse design of nonlocal multilayer thin films. The SAGD method integrates the sharpness‐aware minimization strategy with a differentiable transfer matrix model to navigate the solution space toward perturbation‐tolerant minima. The superior robustness of our method against thickness and refractive‐index variations is demonstrated when compared to conventional techniques. Experimentally, a dual‐channel spatial‐frequency filter is designed and fabricated, which enables switchable bright‐field and dark‐field imaging. This work provides a practical pathway for designing high‐performance, perturbation‐tolerant nonlocal flat optics.