Flip-flop topology inspired forward deep neural network learning approach for modelling UWB antennas

Debanjali Sarkar, Partha P. Shome, Taimoor Khan, Sembiam R. Rengarajan,
Flip-flop topology inspired forward deep neural network learning approach for modelling UWB antennas,
AEU - International Journal of Electronics and Communications,
Volume 201,
2025,
156007,
ISSN 1434-8411,
https://doi.org/10.1016/j.aeue.2025.156007.
(https://www.sciencedirect.com/science/article/pii/S1434841125003486)
Abstract: Human-engineered systems exhibiting intelligent behavior can reduce the constraints associated with designing complex circuits, computational resources, and processing time. Machine learning (ML) algorithms enable us to imitate and create such intelligent systems. In recent years, ML algorithms have gained recognition for efficiently solving complex electromagnetic (EM) circuit design problems. Modeling high-dimensional multi-parametric structures is a significant problem for the EM research community. To address this issue, a deep neural network (DNN) learning approach based on flip-flop topology is presented in this work for fast and efficient modeling of ultra-wideband (UWB) antennas. The proposed approach comprises two interconnected DNN models (DNN I and DNN II) wherein the output of each model is iteratively fed into the other, enabling it to use both real-time output and past predictions. This bidirectional inter-model feedback enhances the capacity of the model to make precise predictions over time. The effectiveness of the proposed model is demonstrated through its application to fast and accurate modeling of miniaturized high-gain UWB antennas and compact quad-UWB multi-input multi-output (MIMO) antennas. For the high-gain UWB antenna, the FFDNN achieved a training MAPE of 0.35 % and testing MAPE of 1.41 %, representing minimum improvements of up to 78 % in training and 39 % in testing compared to traditional MLP, DNN, and FDDNN models. Similarly, for the UWB-MIMO antenna, the FFDNN achieved a training MAPE of 1.12 % and testing MAPE of 1.20 %, marking minimum improvements of approximately 48 % and 60 %, respectively, over traditional MLP, DNN, and FDDNN models. These results highlight the model’s capability to serve as a fast, data-driven surrogate for EM design tasks, offering significant gains in prediction accuracy and computational efficiency over conventional approaches.
Keywords: Antenna; Deep neural network (DNN); Machine learning (ML); Multi-input multi-output (MIMO); Ultra-wideband (UWB)