Compact MOR-Based Neuro-TF Parametric Modeling Incorporating In-Band Approximation and Out-of-Band Conversion Technique

This article proposes a compact model order reduction (MOR)-based neuro-transfer function (neuro-TF) parametric modeling method incorporating in-band approximation and out-of-band conversion techniques for passive microwave components. This method uses complex exponent-pole–zero TF to represent the EM response of microwave devices. The TF based on pole-zero gain format from MOR is used to represent the EM response at first. Then, the poles/zeros are classified to divide the TF into the in-band pole-zero function and the out-of-band complex exponential function. The frequency linear approximation (FLA) is used to guide the classification of poles/zeros and the sorting of in-band poles/zeros. The conversion of complex exponential functions is used to solve the mismatch and nonlinear issue of out-of-band poles and zeros. This method avoids the complex mismatch issue of all poles and zeros. Neural networks are employed to capture the relationship between geometrical parameters and the coefficients of complex exponential functions, poles, and zeros. In comparison to traditional MOR-based neuro-TF parametric modeling techniques, the proposed approach demonstrates superior accuracy and robustness, particularly when dealing with substantial variations in geometrical parameters. The effectiveness of this method is confirmed through three examples involving microwave filters.