“…Therefore, the use of full-wave simulations for executing design tasks such as parametric optimization or tolerance-aware design may be impractical. Fast replacement antenna models (or surrogates) [1], [2] can be obtained by approximating EM simulation data, using, e.g., neural networks [3]- [5], radial basis functions [6], support-vector regression [7]- [9], fuzzy systems [10], or Gaussian process regression (GPR) [11]- [13]. A bottleneck of approximation-based models is high cost of acquiring the training data, with typically a few thousands samples required to ensure reasonable predictive power.…”