Abstract
In this article, a deep learning model of precise electromagnetic (EM) response prediction for frequency selective surface (FSS) is proposed, incorporating both a new regularized learning strategy and a concatenated model structure. The preliminary prediction network (PPN) employs nine fully connected hidden layers and the elastic weight consolidation (EWC) regularization strategy is introduced during the training process. In the EWC regularization strategy, the new and old tasks are defined, effectively addressing the overfitting problem, thus improving the overall accuracy of EM response prediction. The residual compensation network (RCN) comprises three layers of fully connected networks, which are concatenated with the pre-trained PPN. The RCN can effectively compensate for the initial prediction errors of FSS, further improving the EM response prediction accuracy. Compared to artificial neural networks, deep neural networks, and convolutional neural networks, the proposed model demonstrates the lowest mean square errors in predicting the EM response of a typical single-band bandstop FSS and a dual-band bandpass FSS, with values of 2.96 and 4.39, respectively. Then, a dual-band bandpass FSS prototype is fabricated and measured. The measured results further validate the effectiveness of the concatenated structure with the EWC regularization strategy for predicting the EM response of FSS.
Funding Information
  • National Natural Science Foundation of China (52075069)
  • Fundamental Research Funds for the Central Universities (DUT21RC(3)069)

This publication has 27 references indexed in Scilit: