Thermally reconfigurable full-space terahertz polarization-converting metasurface with physics-constrained forward prediction and attention-based inverse design.

Abstract

Reconfigurable terahertz polarization control is of growing importance for sixth-generation (6G) wireless communications, imaging, and sensing, yet simultaneously achieving broadband full-space operation, compact device architectures, and efficient design workflows remains a significant challenge. Here, we propose a thermally reconfigurable full-space terahertz polarization-converting metasurface composed of a 45°-rotated wrench-shaped gold resonator, a SiO2 spacer, and an Au-VO2 hybrid bottom grating. By thermally reconfiguring the bottom boundary via the VO2 phase transition, the same compact platform switches between transmission-mode and reflection-mode polarization conversion without altering the unit-cell topology. The metasurface delivers a transmission polarization-conversion ratio above 90% from 0.30 to 6.66 THz and above 99% from 0.75 to 5.69 THz when VO2 is in the insulating state, while maintaining a reflection polarization-conversion ratio above 90% from 2.44 to 4.56 THz when VO2 becomes metallic. To accelerate electromagnetic modeling while ensuring physical interpretability and enabling on-demand design, we further develop a physics-constrained forward predictor, PhysiLSTM-Net, and an attention-enhanced inverse model, BiCAM-Net. By incorporating the polarization-conversion ratio into the loss function, PhysiLSTM-Net achieves an R2 of 0.9802 with a mean-squared error (MSE) of 0.0051, whereas BiCAM-Net attains a mean R2 of 0.951 across the target structural parameters and an RMSE of 0.269 in structural-parameter inversion. Together, the proposed device and learning framework establish a practical pathway toward broadband full-space terahertz polarization control and physics-guided intelligent metasurface design.