3D holographic metasurface design by deep learning with partitioned loss function

Deep learning has significantly accelerated the automation of metasurface design and reduced its dependence on empirical approaches. However, it still has not fully demonstrated its capabilities in the most challenging light field manipulation: 3D holography. In this paper, we present a framework that integrates a fully connected forward prediction network with a 3D convolutional inverse design network to design terahertz 3D holographic metasurfaces. By using a limited number of target image planes as the ground truth, the network learns the concept of true 3D holography, where patterns rotate and translate as they propagate, and designs metasurfaces to realize the corresponding holograms. From a practical perspective, we propose a partitioning and weighting method for calculating the loss function to enhance the recognizability of 3D holographic imaging. Compared to the analytical method and Gerchberg–Saxton algorithm, the proposed network not only exhibits superior imaging performance but also meets the demands of physically unreasonable 3D holography by adopting the approximate solution. This paper provides a novel and effective method for exploring the immense potential of deep learning in the inverse design of metasurfaces, enabling metasurfaces to better meet the requirements of complex light field manipulation across various wavelength bands. The proposed 3D holographic metasurfaces hold significant potential for advancing terahertz devices that generate specific modes requiring strict and continuous control in 3D space.