Knowledge-Based Neural Network for Multiphysical Field Modeling
Ze Ye, Wei Shao, Xiao Ding, Bing‐Zhong Wang, Sheng Sun
Abstract
This article proposes an efficient knowledge-based neural network (KBNN) for parametric modeling of multiphysical fields. The input of the whole network is the multiphysical parameters, such as geometric variables, voltage, and temperature. The geometric variables with their corresponding electromagnetic (EM) responses are used to train a back-propagation (BP) artificial neural network (ANN) with two hidden layers based on the transfer function (TF). A BP-ANN with one hidden layer, in which the multiphysical parameters are the input and the geometric variables are the output, provides TF-ANN with prior knowledge. With the labeled sampling data from the multiphysical field simulation, the training of KBNN can be completed. KBNN can handle multiple non-geometric input parameters and it also has advantages for shape optimization. The validity of the proposed KBNN model is confirmed with two numerical examples of an iris waveguide bandpass filter and a tunable evanescent mode cavity filter.