MOR-Based Neuro-TF

Neuro-transfer function (neuro-TF) approaches are useful instruments for electromagnetic (EM) parameterized modeling of microwave passive components. A vector fitting technique is commonly used to obtain the TF coefficients from the EM responses utilized in conventional neuro-TF approaches. However, the use of this method can introduce an “order-changing” issue, where the TF orders vary across different regions of the design parameter space. Recently, an alternative neuro-TF method has been proposed, which employs a model-order reduction (MOR)-based TF. Instead of utilizing a vector fitting process, the MOR process acquires the poles/zeros (rational coefficients) during full-wave EM simulations. The MOR technique ensures a consistent TF order across different geometric parameter values. This article provides an extensive overview of the most recent developments in MOR-based neuro-TF techniques. In particular, the approach is examined in two formats: the rational format and the pole/zero format. The former technique provides a solution to the problem of “order-changing” in conventional vector fitting-based neuro-TF, but it still has problems with poles/zeros being mismatched. Thus, the latter format, an EM sensitivity-based approach, has been developed to guarantee the continuity of the poles/zeros across different geometric training samples to address the mismatch problem.