Training Neural Operator Model and Solving Inverse Problems with AH FDTD Simulation Data

This paper explores a data driven machine learning neural network, the neural operator network, as an alternative to traditional simulation methods for solving the inverse problem of the TMz field in a two-dimensional non-uniform conductivity distribution. The neural operator network learns the mapping between the non-uniform conductivity distribution and the output Ez field distribution, the training data obtained by AH FDTD method. After training the neural operator network model, it can predict the conductivity distribution with reasonable accuracy given an Ez field. Furthermore, a comparison was made among four neural operator network models, with the GNO model demonstrating the highest accuracy. The LNO model required the least amount of training time.