Abstract
Surrogate-based optimization (SBO) exploiting variable-fidelity EM models belongs to one of the most promising approaches to expedited simulation-driven antenna design. In SBO, the optimization burden is shifted to a coarsely discretized EM representation of the structure, whereas the high-fidelity model evaluations are carried out only to acquire data necessary for correction of the simplified model or verification of the obtained results. An important issue of variable-fidelity SBO schemes is determination of appropriate low-fidelity model, understood as finding an acceptable compromise between its evaluation time and accuracy. On one hand, unreliable responses of the model that is too coarse may lead to a failure (e.g. divergence) of the optimization algorithm. On the other hand, numerical cost of the design process significantly grows with discretization density of the computational model. Hence, adjustment of the latter should be done carefully so as to control the design cost. Normally, the trade-off between the cost and reliability of the low-fidelity model is determined manually based on the designer’s experience, mostly by checking for discrepancies between the responses for various discretization levels. In this work, the mesh setup is determined automatically through statistical analysis of correlations between the models of various fidelities. The method allows for selecting the discretization level of the low-fidelity model so that it provides acceptable trade-off between the speed and reliability of the design process. The proposed technique is verified using two structures: a dielectric resonator antenna and a slot-ring-coupled patch antenna.