Adaptive Sampling-Based Surrogate Modeling Approach for Efficient Antenna Characterization

Surrogate models for antennas have been extensively adopted in design optimization and uncertainty quantification. However, developing accurate surrogate models often necessitates large-scale training datasets, which require substantial computational resources and time. Herein, a novel adaptive sampling approach is proposed for more efficient surrogate model construction. Firstly, a uniform Latin hypercube sampling method based on the maximin criterion is proposed to generate initial and candidate sample points with improved the global coverage and sampling diversity. Subsequently, a dual-criterion adaptive sampling strategy, which combines uncertainty expectation and gradient sampling, is introduced to guide new samples towards those regions of high prediction uncertainty and rapid response variation, thereby enhancing global accuracy and local sensitivity. Moreover, a distance-based penalty factor mechanism is also incorporated to prevent oversampling in regions with clustered sample points and thereby enhance global exploration. The proposed method has been verified with a series of nonlinear functions, demonstrating favorable model accuracy and sampling efficiency can be obtained with different surrogate models (Kriging and deep neural network). Further, the effectiveness of the method is demonstrated with two types of antennas: a Ka-band microstrip antenna and a Planar Inverted-F antenna. Results demonstrate that the method reduces the required number of sampling points by up to 38% compared to the existing methods, while maintaining high predictive accuracy (relative error < 4%).