S-parameter Modeling and Optimization using Deep Gaussian Processes

Abstract

In this work, a new methodology based on deep Gaussian processes (DGP) is proposed for the modeling and optimization of the S-parameter response of a microwave device. The DGP is used as a surrogate model to directly predict the magnitude (or phase) of the S-parameter as a function over the frequency and over the design parameters of the device. Subsequently an objective probability distribution is retrieved and maximized in a Bayesian optimization (BO) scheme. The new strategy overcomes the limitation of the standard Bayesian optimization that employs an objective function model: simple objective functions are easy to model but may lead to sub-optimal responses, while complicated objective functions may require more powerful and less efficient models. An adequate microwave example demonstrates the increased optimization accuracy of the proposed approach, comparing to standard BO.