Modified Newton Optimization Algorithm for Fast-convergent Parameter Update in Microwave Filter Tuning

Due to its high sensitivity to the geometric variables, the tuning and optimization of microwave filters has consistently presented as a time-consuming and challenging task. Several optimization algorithms have been proposed in recent decades, but they either rely on initial good value, exhibiting low efficiency, or involve complex operations. To address this issue, this article proposes a novel reduced-order model based modified Newton optimization algorithm for the tuning of microwave filters. In the reduced-order model, each element of the coupling matrix associated with the S-parameters of the microwave filter, is represented as a second-order multi-dimensional function of the geometric variables. The reduced-order model can replace the time-consuming electromagnetic (EM) simulations for calculating the cost function during optimization process, and thus substantially reduce the number of iterations. Subsequently, utilizing the reduced-order model, the optimization process can be implemented as searching for a point on a multivariate surface through Newton optimization which incorporates a penalty term for the input offset to guarantee the convergence. Unlike neural networks, the approach does not necessitate the collection of training data, leading to a significant enhancement in optimization efficiency. Two numerical optimization examples are given to demonstrate the effectiveness of the reduced-order model based modified Newton optimization algorithm. The simulation results obtained from TF-eMag — an electromagnetic simulation software, confirm the accuracy of the reduced-order model and validate the effectiveness of the modified Newton optimization algorithm.