Parameterized Modeling Incorporating MOR-Based Rational Transfer Functions With Neural Networks for Microwave Components

Abstract

Neuro-transfer function (neuro-TF) methods are deemed as powerful tools in modeling the electromagnetic (EM) behavior of microwave passive components. Existing neuro-TF methods either endure the issue of “order-changing” or the issue of mismatch of poles/zeros, both calling for specific algorithms to process the data of the transfer function coefficients as a part of model development. This letter proposes a novel neuro-TF method to eliminate the two issues simultaneously by using combined neural networks and model-order reduction (MOR)-based rational transfer functions. In the proposed method, the coefficients of the transfer function are computed by the MOR technique instead of vector fitting. The use of MOR allows the order of the transfer function to remain constant in different regions in the design parameter space, thereby avoiding the issue of “order-changing.” Additionally, the coefficients of the rational transfer function resultant from MOR are naturally sorted according to the order of frequency, which eliminates the issue of mismatch of poles/zeros and subsequently improves modeling accuracy. Compared with existing neuro-TF methods, the proposed method achieves better modeling accuracy for the same geometrical variations. Two microwave examples are utilized to demonstrate the advantages of the proposed method.