Variable-fidelity optimization of microwave filters using co-kriging and trust regions

In this paper, a variable-fidelity optimization methodology for simulation-driven design optimization of filters is presented. We exploit electromagnetic (EM) simulations of different accuracy. Densely sampled but cheap low-fidelity EM data is utilized to create a fast kriging interpolation model (the surrogate), subsequently used to find an optimum design of the high-fidelity EM model of the filter under consideration. The high-fidelity data accumulated during the optimization process is combined with the existing surrogate using the co-kriging technique. This allows us to improve the surrogate model accuracy while approaching the optimum. The convergence of the algorithm is ensured by embedding it into the trust region framework that adaptively adjusts the search radius based on the quality of the predictions made by the co-kriging model. Three filter design cases are given for demonstration and verification purposes.