Neural network EM-Field based modeling for 3d substructure in finite element method

Most existing neural network (NN) approaches for modeling EM problems are based on training with the S-parameter at the external input-output ports of the EM structures. In this paper we present a new method to describe EM behavior at the internal interface between decomposed 3D substructures. We train NNs to represent the EM behavior of the substructure as seen from the interface. This approach allows EM coupling effect between substructures to be better represented. The method is developed in the finite element method (FEM) environment. EM transfer function matrix is formulated to produce training data to allow the NN to learn the coupling between EM field variables at various locations across the entire interface of the 3D substructure. A new formulation allowing trained NN models to be connected with FEM equations of other substructures for efficient simulation of the overall EM structure is proposed. Examples of waveguide circuit simulations show that the proposed method provides better accuracy and efficiency over the conventional neural network method.