Optimization-based robustness enhancement of compact microwave component designs with response feature regression surrogates

摘要

The ability to evaluate the effects of fabrication tolerances and other types of uncertainties is a critical part of microwave design process. Improving the immunity of the device to parameter deviations is equally important, especially when the performance specifications are stringent and can barely be met even assuming a perfect manufacturing process. In the case of modern miniaturized microwave components of complex topologies, it is of paramount importance to carry out tolerance-aware design at the highest available accuracy level (i.e., with the use of full-wave electromagnetic (EM) simulations). Although reliable, EM-driven tolerance-aware design is extremely costly if conventional techniques are to be applied (e.g., Monte Carlo simulation). To overcome this setback, this paper proposes a simple and computationally efficient algorithm for robustness enhancement of compact microwave component designs. The objective is to increase the allowed deviations of geometry parameter values (described using the coefficients of an underlying probability distributions, e.g., the variance) so that the prescribed performance specifications are still fulfilled. The presented approach incorporates knowledge-based surrogate models, constructed using the characteristic points (response features) of EM-simulated system outputs, and utilized for low-cost prediction of the fabrication yield. The parameter vector of the microwave circuit of interest is adjusted within the trust-region (TR) framework to identify the maximum levels of deviations still ensuring 100-percent yield. The employment of TR also permits the adaptive control of design relocation and ensures convergence of the optimization process. Numerical verification of the presented methodology is carried out using three miniaturized microstrip circuits, including two equal-split couplers and a wideband filter. The major finding is that incorporating knowledge-based feature surrogates allows for achieving a significant improvement of the acceptable input tolerance levels (nearly two fold on the average) at a remarkably low cost of few dozen EM simulations