Prior Knowledge-Guided Generative Deep Learning for Multi-Objective Synthesis of Miniaturized Transmissive Metasurface Elements

A generative deep learning-enabled method has been proposed to efficiently synthesize quasi-freeform miniaturized metacells for dual-polarized wide-angle transmission. A four-layer pixelated metacell with metal vias is developed under a constrained small period, wherein the enriched degrees of freedom (DoF) enhance the potential for high transmission amplitude and wide-range transmission phase shifts under dualpolarized wide-angle incidence. However, the high DoF of these quasi-freeform metacells also poses the time-consuming challenges of searching in a vast solution space for multi-objective synthesis. To address this, multiple types of prior knowledge (PK) are extracted from the design targets to guide the data construction for efficient training and effective synthesizing. A condition generative neural network is proposed to learn the PK-guided dataset and realize a controllable generation for miniaturized metacells using the trained network. Finally, the generated metacells achieve a transmission phase shift range exceeding 297° and a transmission amplitude above 0.75 within a compact period of 0.24 wavelengths under the dual-polarized incoming waves at the incident angles from 0° to 45°.