OPTIMIZATION OF ANN-BASED MODELS AND ITS EM CO-SIMULATION FOR PRINTED RF DEVICES

Abstract

Printed VO2 RF switch founds immense potential in RF reconfigurable applications. However, their generic electrical equivalent model is still intangible that can be further integrated in CAD tools and utilize for simulation, analysis and design of RF/microwave circuits and systems. The artificial neural network (ANN) has been gaining popularity in modeling various types of RF components. However, most of these works merely demonstrate the establishment of the ANN-based RF model in the MATLAB environment without involving significant optimization. Furthermore, the integration of such ANN-based RF models in the EM and circuit simulator as well as the co-simulation between the ANN-based model and conventional models have not been demonstrated or validated. Therefore, the earlier reported models are still one step removed from its real RF applications. In this work, by using the fully printed vanadium dioxide (VO2) RF switch as the modeling example, a systematic hyperparameter optimization process has been conducted. Compared to the nonoptimized ANN model, a dramatic improvement in the model's accuracy has been observed for the ANN model with fully optimized hyperparameters. A correlation coefficient of more than 99.2% for broad frequency range demonstrates the accuracy of the modeling technique. In addition, we have also integrated the Python-backed ANN-based model into Advanced Design System (ADS), where a reconfigurable T-resonator band stop filter is used as an example to demonstrate the co-simulation between the ANN-based model and the conventional lumped-based model.