Recent Advances in Novel Training Approaches for Microwave Parametric Modeling Using Padé via Lanczos and EM Sensitivities

This paper provides the recent advance in training approachs for microwave parametric modeling of passive components with changes of geometrical variables combining electromagnetic (EM) sensitivities and matrix Padé via Lanczos (MPVL). In this paper, the pole-zero-gain transfer function is used as the surrogate model of EM response of microwave components versus frequency. The relationships between geometrical parameters and the gain/zeros/poles are learned directly by artifical neural networks (ANN). MPVL algorithm is performed to compute/recompute the zeros/poles corresponding to different geometric parameters to generate training data. After recomputations, the indices of the zeros/poles not always have right correspondences with original indices, leading to an unstable prediction of the zeros/poles when geometrical parameters have a new change. A novel sensitivity-based zero/pole-matching algorithm is used to achieve the right correspondences between the zeros/poles at different geometrical variables. This method uses the EM sensitivities as an aid information for the moving direction of the zeros/poles, to match the zeros/poles and geometrical parameters that changes each time. Using the matched zeros/poles to train the neural network has reliable and fast predictions for large geometrical variations, which increases the robustness and accuracy of the surrogate model. Compared with those existing algorithms, this method can train a more accurate model in applications involving large geometrical variations.