GPR Antenna Modelling Based on DML Exploiting Cosine Similarity Metric

Abstract

Traditional full-wave electromagnetic simulation software, such as HFSS or CST, requires substantial computational resources and time consumption, as does the construction of surrogate models when acquiring new samples. To overcome this data efficiency challenge, we propose a Data Augmentation (DA) modelling approach based on Deep Metric Learning (DML) that integrates a Siamese Neural Network (SNN) with K-Nearest Neighbours (KNN) to generate augmented samples for training surrogate models. First, SNN is employed to extract embeddings of antenna parameters. The performance between loss functions constructed using various distance metrics, including cosine similarity and others, is systematically compared. Subsequently, the KNN algorithm generates new samples based on the distance to neighbours. These augmented samples are combined with the original dataset and fed into the surrogate model using a Gaussian Process Regressor (GPR) and comparative analysis. The proposed method is validated using a WLAN Dual-Band Monopole Antenna (WLAN-DBMA) and a Log-Periodic Folded Dipole Array (LPFDA) antenna. Results demonstrate that, compared with surrogate models without sample augmentation, the proposed method efficiently reduces the mean absolute percentage error (MAPE) by over 20%, decreases the mean squared error (MSE) by more than 35%, and improves the coefficient of determination (R²) by over 4%.