Neural Network Learning Techniques Comparison for a Multiphysics Second Order Low-Pass Filter

The design process of high frequency (HF) structures has been a major challenge in recent years, with the increasing demand for better performance and more complex designs. This process is even worse when considering multiphysics simulation, such as thermal and mechanical physics combined with the electromagnetic response of the circuit. The traditional approach of trial and error can be highly time consuming and prone to error while the use of direct optimization techniques can also be computationally expensive. To speed up the design process, artificial neural networks (ANN) have been extensively used to develop fast and accurate surrogate models of high-frequency structures. In this work, we compare the performance of three ANN learning techniques (Levenberg-Marquardt, Bayesian, and Resilient) for modeling a multi-physics second order low-pass filter implemented in COMSOL. Results demonstrate that the Bayesian ANN exhibits the best performance for efficiently and accurately model the multi-physical response of HF structures.