Residual deep neural network and response features for inverse modeling of microwave passives

Abstract

Electromagnetic (EM) solvers are widely used in microwave design for their accuracy, but are computationally expensive. Surrogate models offer a remedy, yet face challenges such as the curse of dimensionality. This work introduces a novel inverse modeling approach using deep learning and carefully selected response features to predict circuit geometry from desired operating parameters such as center frequency and power split ratio. Data samples for model training are collected through a performance-driven sequential process. Setting the model at the feature level implicitly reduces dimensionality, as the feature set is smaller than the geometry parameter set. This allows reliable regressors to be built with far fewer training samples, greatly lowering computational cost. Additional savings are achieved through explicit dimensionality reduction, carried out using global sensitivity analysis. To set up an inverse surrogate, we employ a customized residual deep residual neural network (ResNet), which is particularly well-suited for capturing dependencies between features and geometry parameters. This is due to ResNet’s resilience to a certain degree of parametric redundancy in the circuit structure. The inverse model produces parameter vectors sufficiently close to the optimal ones, thus only minor refinement is necessary. Also, our model is shown to be more accurate than state-of-the-art forward and inverse metamodels. The proposed ResNet embedded in the inverse modeling regime and combined with dimensionality-reduced sequential sampling are the main artificial intelligence-related contributions of this study. Engineering applications encapsulate the successful employment of our framework for rapid global circuit optimization, as exemplified using three microwave passives.