Adaptive Sampling for Spherical Near-Field Antenna Measurements via Uncertainty-Driven Gaussian Process Regression

This paper proposes an adaptive sampling technique for spherical near-field (SNF) antenna measurements that leverages the predictive uncertainty of Gaussian process regression (GPR) to guide measurement point selection. This framework bridges Bayesian machine learning and antenna metrology, enabling real-time adaptive sampling. Unlike conventional uniform or clustering-based methods, the proposed approach iteratively refines the sampling grid by prioritizing regions with high uncertainty, particularly within the high-intensity near-field or other defined regions of interest. This uncertainty-driven GPR strategy enables high-fidelity far-field (FF) reconstruction with substantially fewer measurements. In conventional low-cost SNF systems, the probe must stop at every measurement point to reduce vibration, making the stop-and-go process time-consuming. In contrast, the proposed method minimizes unnecessary stops by targeting only high-uncertainty regions, offering significant advantages in research and development and production-line environments. In such cases, the optimized sampling points identified during the initial adaptive process can be reused, enabling rapid and efficient FF characterization without repeating the full measurement cycle. The method is validated through both full-wave simulations and experimental measurements involving various antenna types, including a large commercial base station antenna. Results demonstrate that the proposed approach achieves up to a 77% reduction in sampling points compared to uniform sampling, while maintaining FF pattern correlation exceeding 99.35%. Overall, the technique offers a statistically grounded and time-efficient solution for high-fidelity antenna characterization with reduced measurement effort.