Design of Artificial Electromagnetic Materials Using ResNet-Based Deep Learning Method

The design of artificial electromagnetic materials (AEMMs) depends highly on full-wave numerical simulations or equivalent circuit model (ECM)-assisted analysis. This work proposes an intelligent design method using a deep learning (DL) technique based on the residual neural network (ResNet) to improve its efficiency. Firstly, adopting pixeled matrix modelling methods enhances the freedom of design. Next, the staircase approximation is utilised for the S-parameter curve, which also describes the required electromagnetic (EM) property to be used in the training process. These processed samples, along with their corresponding labels, are transformed and fed into ResNet for training. After these procedures, the structural matrix of the desired curve can be predicted through well-trained networks. To validate the effectiveness of the method, typical notched-band frequency selective absorbers (FSAs) are designed, while the reflective band can easily be adjusted. Compared with conventional methods and other deep neural network (DNN)-based methods, this method performs more efficiently and accurately. Finally, an illustrative sample is fabricated to validate the prediction result.