Inverse Design of Arbitrary Large‐Scale Off‐Axis Metalenses With a Physically‐Tailored Deep Learning Model

ABSTRACT Efficient large‐scale off‐axis metalenses with oblique incidence have drawn a lot of attention for designing future augmented and virtual reality devices. Classical methods based on traditional lookup tables for designing such metalenses lacked efficiency and presented several engineering challenges to achieve high‐quality images due to the constant high phase gradient. This study introduces a novel inverse design methodology for design large‐scale metalenses that combines high performance, scalability, and exceptional design flexibility. Our innovative approach assembles multiple optimized beam deflectors that redirect oblique incident light to the focal point. Leveraging deep learning models and incorporating a physics‐aware assembling strategy into the optimization process, the high efficiency of the meta‐deflectors is translated into high focusing efficiency of the final metalens design. Once the base models are trained, the inverse design process can generate large‐scale metalenses with varying sizes, numerical apertures, and focal positions (steering the beam), without the need to re‐initialize the entire optimization process, enabling rapid prototyping and versatility in designing optical devices.