Accelerating Full-Wave Antenna Optimization: An Adaptive Surrogate-Assisted Differential Evolution Framework

Full-wave electromagnetic (EM) simulation, particularly in environments such as CST Studio Suite, makes large-scale antenna optimization computationally prohibitive. We introduce an adaptive surrogate-assisted differential evolution (DE) framework, implemented via a unified CST–Python workflow, designed to accelerate design-on-demand antenna optimization. The workflow integrates target-frequency-driven design requests, multimetric antenna performance targeting, cross-validated surrogate selection, selective CST validation, and iterative CST-verified dataset updating. It begins with Latin hypercube sampling (LHS) to create a CST-simulated training set, selects a regressor (KNN, RF, SVR, GB, XGBoost) via five-fold cross-validation based on mean squared error, and then uses the surrogate to guide the DE search. The core adaptive mechanism involves mandatory full-wave validation of the best design candidate from each optimization cycle, appending the verified result to the training dataset to enable targeted model refinement. Optimization is governed by a multiobjective penalized aggregate function that minimizes the resonant-frequency error while maximizing the design performance metrics of bandwidth, return loss, and gain. We evaluated this approach on three antenna families—dipole (2.00 GHz), microstrip patch (2.55 GHz), and Yagi–Uda (2.50 GHz)—and met targets with only 10–12 full-wave validations per run. Our method achieved a verified design in 9–16 min, whereas pure DE took 21–113 min with 28–55 full-wave solves, and pure PSO took 18–217 min with 28–106 full-wave solves. This corresponds to speedups of 2.38–8.06× and 2.04–13.68×, respectively. This work demonstrates that integrating an adaptively selected surrogate model into the optimization strategy substantially reduces the computational cost of full-wave analysis, establishing a highly efficient and robust methodology for diverse EM design applications.