Exploit Both Patterns and Structural Parameters by Multimodal Learning to Predict and Design Metasurfaces

A multimodal (MM) deep learning (DL) model is proposed that seamlessly combines the fully connected (FC) and convolutional layers to enhance the prediction accuracy and the design efficiency of metasurfaces. The convolutional layers in the model compensate for the inability of FC models to extract features from topology patterns in the metasurface. Specific regions instead of the whole metasurface are binarized by the equivalent-circuit (EC)-model-based region selection method to avoid heavy computational cost for the convolutional layers, which makes the proposed model applicable to multilayer metasurfaces with complex patterns. The proposed model is validated numerically by predicting and designing a complex frequency-selective rasorber (FSR). Numeric results demonstrate that the proposed model achieves 30.1% more prediction accuracy and 61% more design efficiency when compared with FC models to design the rasorber to satisfy specific performance requirements.