Antenna Modeling With Data Augmentation Based on TimeGAN

In the field of antenna design, direct optimization using full‐wave electromagnetic simulation software combined with swarm intelligence algorithms usually consumes substantial computational time and expensive hardware resources. To improve optimization efficiency, surrogate models for antennas are typically constructed, and the trained surrogate models serve as the fitness evaluation function for global optimization algorithms. However, the training accuracy of surrogate models usually depends on a large number of training samples, and obtaining such samples itself is time‐consuming. To address this issue, this study proposes a data augmentation method based on generative adversarial networks (GANs), aiming to generate a large number of reliable samples using a small amount of full‐wave electromagnetic simulation data, thereby meeting the training requirements of high‐precision surrogate models. Specifically, based on limited full‐wave simulation data, three generative models, including TimeGAN, Wasserstein‐GP‐TimeGAN (WG‐TimeGAN), and Wasserstein‐TimeGAN (W‐TimeGAN), are employed separately for data augmentation. Subsequently, three types of surrogate models, covering Gaussian process regression (GPR), radial basis function network (RBFN), and ensemble learning (EL), are constructed to verify the effectiveness of the generated data. Experimental results on a dual‐frequency slotted patch antenna and an ultra‐wideband (UWB) microstrip antenna show that the generated data can be effectively used for surrogate model training. In the scenario of the dual‐frequency slotted patch antenna, the GPR model trained on the WG‐TimeGAN‐generated data exhibits the best performance, with a mean absolute percentage error (MAPE) of 4.80%, representing a 1.70% improvement compared to the results based on TimeGAN‐generated data. In the application of the UWB microstrip antenna, the XGBOOST model within the EL framework achieves excellent performance on both WG‐TimeGAN and W‐TimeGAN generated data, achieving MAPEs of 7.53% and 8.69%, respectively. These results verify the applicability and effectiveness of the proposed method across different antenna types.