Semi-supervised Latent Diffusion Model: Reverse Design of Dual-Polarization Frequency Selective Metasurface

To avoid the complex design process for the pixelate frequency selective metasurface, a semi-supervised latent diffusion model (S-LDM) is proposed in this article. S-LDM utilizes polarization-independent TE and TM spectrums to guide the denoising of random Gaussian noise, and a semi-supervised learning strategy is adopted for training the latent diffusion model (LDM). First, the modified National Institute of Standards and Technology (MNIST) Dataset and its extended version are selected for constructing the polarization-independence frequency selective metasurface with asymmetry 28 × 28 pixelates. Then, an auto-encoder (AE) network is adopted to compress the 28 × 28 metasurface pattern into an 8 × 8 latent space, and such compression method has effectively shortened the training time. Next, pseudo labels for unlabeled data are constructed, and their loss values are added as dynamic regularization weights in the total loss function. In doing so, S-LDM generates more precise metasurfaces than supervised training. Full-wave simulation results demonstrate that the dual-polarization frequency-selective metasurfaces can be accurately reverse-generated based on the desired TE-TM spectrum in the test dataset. Reverse design metasurfaces have also been successfully achieved by randomly generating Gaussian spectrums with various frequency selective and polarization capabilities. Finally, a real-world metasurface is fabricated by S-LDM. The measurement results of this metasurface show the obvious dual-polarization frequency selective characteristic.