Variable-Fidelity Design Optimization of Antennas with Automated Model Selection

Computationally efficient design optimization of antenna structures can be realized using variable-fidelity EM simulations, where most of the operations are performed at the level of a coarse-discretization EM model (referred to as the lowfidelity model) with occasional reference to the high-fidelity one, for the sake of model correction and design verification. An important issue is an appropriate selection of the low-fidelity model in terms of its speed and accuracy. It is typically controlled by adjusting discretization density of the structure at hand. Using the model that is too coarse may lead to unreliable design and divergence of the optimization algorithm. If the model is too fine, the computational cost of the design process increases. So far, the model selection has been realized interactively through visual inspection of the model responses. In this paper, an automated procedure for low-fidelity model setup has been proposed based on statistical analysis of correlations between the models of various fidelities. Using an example of a dielectric resonator antenna we demonstrate that our approach allows for selecting the optimal discretization level of the lowfidelity model that results in a good compromise between the cost and reliability of the design process.