Deep learning-assisted focus engineering for metalens design with high numerical aperture cylindrical vector beams

Deep learning, in comparison to traditional optimization algorithms, offers significant advantages in addressing complex problems involving multi-dimensional design parameters for the customization of three-dimensional focal fields. In this paper, we present a design method that combines Richards-Wolf vector diffraction theory and neural network techniques for achieving the customization of different focal fields. We utilize a Physics-Connected Neural Network (PCNN) to devise a discrete filter with 25 rings for guiding the structure of the all-dielectric metalens, thereby facilitating the inverse design of optical needles, optical tubes, and flat-top light fields with an extended focal depth (>10λ). The results demonstrate that when combined with a physical model, the neural network effectively adapts to diverse design objectives, reduces design complexity, and improves the efficiency of the design process.