Automated Stopping Criterion for Sparse Near-Field Scanning With Active Machine Learning

Active learning has been demonstrated to accelerate near-field scanning (NFS) measurement by adaptively sampling the significant points to reconstruct the complete electromagnetic field. However, an automated stopping criterion for active-learning sampling is still lacking, and the decision to stop sampling is often made manually. Since the ground-truth error using sparse samples compared to the full scanning result is unknown, we propose a novel stopping criterion based on a statistically estimated worst-case predicted error. The active learning-based NFS process will be terminated when the maximum predicted error at the unsampled positions falls below a user-defined threshold, which can flexibly and intuitively satisfy the accuracy requirements of users in various experimental scenarios. The proposed stopping criterion provides sufficient accuracy and appropriate sample size for sparse NFS measurement, which is validated through both simulation and measurement cases.