Deep Learning‐Assisted Inverse Design of Terahertz Bilayer Split‐Ring Metasurface Wave Plates

ABSTRACT Metasurface quarter‐wave plates (QWPs) are used to manipulate the polarization of terahertz (THz) waves. However, traditional trial‐and‐error designs for multilayer achromatic QWPs are time‐consuming and challenging. To address this, a deep probabilistic prediction framework (RMCV‐MDN) is constructed by integrating residual blocks, multi‐head attention, conditional variational autoencoders, and conditional normalizing flows with a mixture density network. While effectively handling the intrinsic “one‐to‐many” problem and mapping uncertainty in QWP inverse design, RMCV‐MDN demonstrates strong small‐sample learning, achieving a coefficient of determination () exceeding 0.96 and a mean absolute percentage error (MAPE) as low as 0.25%. It can also predict QWP structural parameters with the highest probability of meeting performance metrics, effectively overcoming the insufficient generalization ability of traditional deep learning. Utilizing RMCV‐MDN, the inverse design of a reflective bilayer split‐ring (SR) THz metasurface QWP is successfully achieved as a benchmark model. The simulation and experimental results of the inverse designed QWPs demonstrate a high degree of correlation with the target performances, achieving high‐efficiency polarization conversion across the 0.6–1.4 THz and 0.85–1.8 THz ranges, respectively. This work provides a faster, more accurate approach for designing multilayer metasurface broadband achromatic QWPs, paving the way for the development of other metasurface devices.