Multi-Output Variable-Fidelity Surrogate Modeling for Microwave Components Design

This paper proposes a Gaussian processes-based modeling technique for handling multi-output (frequency-dependent vector-valued) microwave systems, in which variable-fidelity data is available. This approach assumes that each frequency point has its own mean and covariance, which are independent of the ones of other frequencies, and a unique correlation matrix describes the full frequency range. Using these assumptions can significantly reduce the number of matrix calculations required by the maximum likelihood estimation of the data. The notable advantage of this proposed approach is the capability of modeling systems with multi-output responses without requiring any pre-conditioning of the data while keeping a high accuracy, which is especially useful in the microwave design. The proposed methodology is demonstrated for a parameterized PCB connector design and compared against usual surrogate modeling approaches.