Time-Domain Computational Electromagnetics with Machine Intelligence
The objective of this research is to investigate new computational methodologies that utilize large databases obtained from simulations and measurements in order to develop predictive physical models. In the talk, we will present our recent progress on the emulation of transient electrodynamic physics using deep neural network (DNN). The specific neural network being considered for this study is a recurrent neural network (RNN) and convolutional autoencoder, which recursively update its state (hidden layer) to generate consecutive time steps of wave propagation. Through the training of a sufficiently large number of simple problems, a new model problem can be solve rapidly with a high probability of correct response.