Low-Cost Surrogate Modeling of Compact Microstrip Circuits in Highly-Dimensional Parameters Spaces Using Variable-Fidelity Nested Co-Kriging

摘要

Full-wave electromagnetic (EM) analysis permits reliable but often computationally expensive evaluation of microwave devices. In particular, multiple simulations (e.g., entailed by numerical optimization procedures) may generate CPU costs that are unacceptable or at least significantly slow down the design cycles. This problem may be mitigated by the employment of fast surrogate models. Unfortunately, conventional techniques are not suitable for handling structures described by multiple geometry/material parameters and featuring highly nonlinear frequency characteristics. This paper proposes a novel modeling approach that incorporates variable-fidelity EM simulations into the recently reported nested kriging framework. A combination of model domain confinement due to nested kriging, and low-/high-fidelity EM data blending through co-kriging, enables the construction of reliable surrogates at a fraction of cost required by single-fidelity nested kriging. This is achieved without compromising the predictive power of the model. Our methodology is demonstrated using a miniaturized impedance matching transformer described by fifteen geometry parameters. Setting up the surrogate valid over wide ranges of the operating bandwidth (1.5 GHz to 3.5 GHz for the lower edge, and 4.5 GHz to 6.5 GHz for the upper edge) involves only about three hundred of (equivalent) high-fidelity EM simulations of the transformer circuit.