Fusion Neural Networks for High-Precision Design and Ultrawideband Shielding in Frequency Selective Surfaces

Abstract

This article advances a multi-input multi-output regression approach for continually optimizing the physical design of frequency selective surfaces (FSSs) to enhance shielding effectiveness (SE) over an ultrawide bandwidth. The classical models of neural networks (NNs) are frequently used for such regression tasks due to their capacity to extract features from average datasets; however, they struggle with complex design parameters and broad responses. To accomplish this, a hybrid NN technique is developed, which combines convolutional neural networks (CNNs) for deep feature capture with long short-term memory (LSTM) networks and an attention mechanism for processing. This method makes efficient use of linear information in design dimensions while also consistently computing S-parameters. The FSS design employs a double square loop that resonates at two frequencies, resulting in a wideband response; by adding four stubs between the loops, it becomes an ultrawideband (UWB) response. The equivalent lumped circuit (ELC) model is applied for estimating capacitance (C) and inductance (L), which are then transformed to ABCD and S-parameters. The model obtained an $r^{2}$ value of 99.3% and a mean squared error (mse) of $1 \times 10^{-3}$ after optimizing the design parameters in 124.25 s. The design, implemented on Rogers 5880LZ with a thickness of 1.27 mm and unit-cell dimensions of $11 \times 11$ mm, is stable across varied incidence angles and polarization insensitive, allowing for downsizing. Comparative research demonstrates that the CNN-LSTM hybrid model surpasses traditional techniques, illustrating higher SE.