Physics Embedded Neural Network (PENN) Architecture for Solving Maxwell’s Equations towards Accelerated Microwave Modeling

In this paper we demonstrate a neural network architecture that is useful in providing a solution of wave equation through the optimization of the neural network. The developed algorithm is based on integrating the physics equations directly into the neural network and therefore provides an effective, datafree method of solution of the underlying equations. The proposed neural network exhibits solution time similar to that of classical finite element method (FEM) solver. Additionally, by employing transfer learning, neural network achieves even faster solution for similar new equation. As compared to the direct neural network optimization, the transfer learning solution is up to 7.1 times faster for the cases considered.