An Adaptive Antenna Sensitivity-to-Yield Analysis and Optimization Framework Using Active Learning-Based Gaussian Process Regression

TY - JOUR

T1 - An Adaptive Antenna Sensitivity-to-Yield Analysis and Optimization Framework Using Active Learning-Based Gaussian Process Regression

AU - Zhang, Zhourui

AU - Zhao, Jiaqi

AU - Hu, Weidong

AU - He, Mang

N1 - Publisher Copyright: © 2026 IEEE. All rights reserved.

PY - 2026/3

Y1 - 2026/3

N2 - Geometrical and material tolerances in fabrication can degrade electromagnetic performance and yield of antennas. This letter develops a sensitivity-to-yield (AStY) framework that links sensitivity analysis (SA) with tolerance optimization to enhance antenna yield under fabrication uncertainties. An SA method based on automatic relevance determination Gaussian process regression (ARD-GPR) is developed, where the isotropic Matérn kernel is equipped with dimension-specific length-scales to construct a principled sensitivity metric, enabling direct parameter ranking. Moreover, a variance-guided adaptive sampling with clustering is proposed to reduce simulation cost and ensure a standardized process applicable to different antenna designs. Compared with conventional Latin hypercube sampling, this sampling scheme reduces the number of numerical simulations by 30% and runtime by 39%. Numerical examples on U-slot and broadband Vivaldi antennas show yield improvements from 60.1% to 90.3% and from 76.4% to 92.2%, closely matching full tightening while requiring significantly less fabrication precision. The proposed AStY framework offers a practical route for cost-effective yield enhancement.

AB - Geometrical and material tolerances in fabrication can degrade electromagnetic performance and yield of antennas. This letter develops a sensitivity-to-yield (AStY) framework that links sensitivity analysis (SA) with tolerance optimization to enhance antenna yield under fabrication uncertainties. An SA method based on automatic relevance determination Gaussian process regression (ARD-GPR) is developed, where the isotropic Matérn kernel is equipped with dimension-specific length-scales to construct a principled sensitivity metric, enabling direct parameter ranking. Moreover, a variance-guided adaptive sampling with clustering is proposed to reduce simulation cost and ensure a standardized process applicable to different antenna designs. Compared with conventional Latin hypercube sampling, this sampling scheme reduces the number of numerical simulations by 30% and runtime by 39%. Numerical examples on U-slot and broadband Vivaldi antennas show yield improvements from 60.1% to 90.3% and from 76.4% to 92.2%, closely matching full tightening while requiring significantly less fabrication precision. The proposed AStY framework offers a practical route for cost-effective yield enhancement.

KW - Automatic relevance determination

KW - Gaussian process regression

KW - machine learning

KW - sensitivity analysis

KW - yield optimization

UR - https://www.scopus.com/pages/publications/105027472380

U2 - 10.1109/LAWP.2026.3651907

DO - 10.1109/LAWP.2026.3651907

M3 - Article

AN - SCOPUS:105027472380

SN - 1536-1225

VL - 25

SP - 1244

EP - 1248

JO - IEEE Antennas and Wireless Propagation Letters

JF - IEEE Antennas and Wireless Propagation Letters

IS - 3

ER -