Antenna modeling and optimization based on 1D GP-HetCNN surrogates and intelligent methods

This paper proposes a lightweight antenna surrogate model based on a one-dimensional grouped-and-pointwise heterogeneous convolutional neural network (1D GP-HetCNN) for the prediction of the scattering parameter ($$\text {S}_{\text {11}}$$) curve of antennas. Using antenna physical parameters as the input, the surrogate model can predict the $$\text {S}_{\text {11}}$$ curve efficiently and accurately. Then, in order to study complicated frequency features of curves, the 1D GP-HetCNN model is further combined with a multi-task learning method, constructing 1D GP-HetCNN-MTL. Based on the two trained surrogate models, two optimization methods employing the whale optimization algorithm and the coati optimization algorithm are developed respectively. On the test set, the 1D GP-HetCNN model achieves the root mean square error (RMSE) of 0.0305 and the mean absolute error (MAE) of 0.0189; the 1D GP-HetCNN-MTL model achieves the RMSE of 0.0426 and the MAE of 0.0282. Both surrogate models perform better than the two baseline models. Moreover, results from the two optimization methods show that they can identify improved physical parameters with a wider operating bandwidth from 19.3 GHz to averagely 20.0 GHz more efficiently than conventional electromagnetic (EM) simulation, which increases the fractional bandwidth from 155.02% to averagely higher than 156%. These findings indicate that surrogate models combined with intelligent optimization algorithms can accelerate antenna modeling and further improve the antenna’s EM performance. Overall, the proposed approach facilitates more efficient antenna design and contributes to the development of intelligent antenna engineering.