Numerical Dispersion Compensation for FDTD via Deep Learning

A hybrid deep neural network consisting of convolutional layers and LSTM structures is developed for compensating for the numerical dispersion error in FDTD. The proposed network is designed for the wideband computation of S-parameters of planar microwave circuits at coarse meshes. Without dispersion compensation, the accuracy of these computations is severely compromised. We show that deep learning can restore the accuracy to the level obtained by dense mesh simulations.