Inverse design of free-form metasurfaces: from adjoint-based optimization to physics-informed deep learning

Abstract

Recent advances in nanofabrication enable the manufacture of arbitrary-shaped optical metasurfaces. The carefully designed free-form metasurfaces display advanced photonic performance which cannot be achieved by traditional optics. However, the design of free-form metasurfaces incurs an expensive cost. Here, we demonstrate that free-form metasurfaces can be optimized rapidly by employing both adjoint-based optimization and physics-informed deep learning. In particular, the design method of metamask for proximity field nanopatterning enables the generation of unachieved patterns. In addition, we propose symmetry-encoded convolutional neural networks to reflect physical characteristics of periodic metasurfaces. The proposed deep learning models can replace optical simulators for measurements of physical properties. Combining adjoint-based optimization and physics-informed deep learning, the inverse design process can be significantly accelerated.