Numerical and experimental generated data during project https://doi.org/10.1109/TAP.2025.3534987

The dataset was generated during a project aimed at developing a novel methodology forhigh-accuracy modeling of antenna characteristics. It is centered around a recurrent neural network (RNN) optimized through Bayesian optimization (BO). The RNN architecture includes—among others—a feature input layer, a long shortterm memory (LSTM) layer, a gated recurrent unit (GRU) layer, and a bidirectional LSTM (Bi-LSTM) layer. BO is employed to fine-tune the network’s hyperparameters, such as the number of units in each layer and the learning rate. The LSTM, GRU, and Bi-LSTM layers are designed to capture frequency-wise dependencies in antenna responses. A distinctive feature of this approach is the utilization of frequency as a sequential parameter processed by the RNN, which enhances reliability and cost efficiency. These advantages are further bolstered by dimensionality reduction techniques based on rapid global sensitivity analysis. The proposed method has undergone extensive validation across four antennas and has been compared against several benchmark methods, including kernel-based regression models and various neural networks. The results demonstrate that our surrogates achieve outstanding accuracy, significantly surpassing the benchmark, and can construct reliable metamodels even with low-cardinality training datasets. Design applications and experimental validation have also been included.