Attention-Based Deep Neural Network for High-Dimensional Microwave Modeling of Non-Reciprocal Bandpass Filters

This paper presents attention-based deep neural networks for high-dimensional microwave modeling to predict behavior of spatio-temporal modulated (STM) non-reciprocal bandpass filters (NR-BPFs). The proposed method integrates convolutional neural networks (CNN), joint channel-spatial attention, and residual connections to enhance prediction accuracy and model generalization. The combined channel-spatial attention approach allows the model to concentrate on essential features, while residual connections improve learning by reducing overfitting and preventing vanishing gradients, ensuring efficient gradient flow. For experimental validation of the proposed model, high-dimensional microwave modeling of four types of NR-BPF (example 1: third-order single-band NR-BPF, example 2: fourth-order single-band NR-BPF, example 3: third-order dual-band NR-BPF, example 4: microstrip line third-order single-band NR-BPF) are performed. Experimental results show that the proposed model exhibits superior performances for four types of NR-BPFs, achieving lower root mean square error (RMSE) and higher R-squared as compared to conventional deep multilayer perceptron (MLP) and standard CNN models. The results indicate that the proposed approach provides a robust solution for accurately predicting complex input-output relationships in STM-based NR-BPFs, outperforming conventional deep MLP and CNN models in both prediction accuracy and model robustness.