Passivity enforcement for passive component modeling subject to variations of geometrical parameters using neural networks

A novel passivity enforcement technique for passive component modeling subject to variations of geometrical parameters is proposed using combined neural networks and rational functions. A constrained neural network training process to enforce passivity of Y-parameters is introduced. Eigenvalues of Hamiltonian matrix for parametric model at many geometrical samples are used simultaneously as constraints for neural network training. Furthermore, a new passivity conditioning parameter e is proposed to guide the training process. Once trained, the parametric model can provide accurate, fast and passive behavior of passive components for various values of geometrical variables within the model training range. A parametric modeling example of an interdigital capacitor is presented to demonstrate the validity of the proposed technique.