Multiobjective Optimization of Antenna Inverse Design With Data Augmentation Based on K-Means-NN

To enhance the accuracy of the inverse model with a limited number of antenna sample points, a data augmentation method based on K-means and forward neural networks (FNNs) is proposed. In this method, K-means is used to partition the initial imbalanced dataset, while a perturbation factor and FNNs are introduced to oversample the minority samples, ensuring a balanced performance distribution. A weighted loss combining soft dynamic time warping and mean squared error is embedded in FNNs to predict antenna performance more accurately. The convolutional inverse neural network integrates multiple performance features stacked along a specific dimension and is iteratively invoked until the optimized antenna meets the desired objectives. Taking a wideband circularly polarized antenna as a numerical example, the proposed method requires fewer electromagnetic simulations than other advanced optimization techniques.