Machine Learning-Based Antenna Optimization with Multi-Fidelity EM Analysis and Dimensionality Reduction by Global Sensitivity Analysis

This work introduces a novel technique for cost-efficient antenna optimization. Our approach employs computational intelligence tools, including machine learning (ML) and bio-inspired algorithms. The ML process is conducted within a dimensionality-constrained parameter space established through global sensitivity analysis (GSA). The underlying surrogate model is kriging interpolation, whereas candidate designs are generated using the particle swarm optimizer (PSO). Further enhancement of the computational efficacy is obtained by employing multi-fidelity electromagnetic (EM) models, with majority of operations executed at low-fidelity antenna representation and final (gradient-based) tuning performed using the high-fidelity model. Extensive numerical experiments demonstrate the superiority of the introduced approach over several benchmark procedures regarding reliability, computational efficiency, and repeatability of results.