On incorporating finite impulse response neural network with finite difference time domain method for simulating electromagnetic problems

The finite difference time domain method (FDTD) is a very powerful numerical method to solve electromagnetic (EM) problems. It is very flexible to simulate the problems which have very complex boundaries. It is well known that FDTD method requires long computation time for solving the resonant or high-Q passive structures. The reason for this is because the algorithm is based on the leap-frog formula. For EM modeling is very important to speed up the simulation. The finite impulse response neural network (FIR NN) is applied as a nonlinear predictor to predict time series signal for speeding up the FDTD simulations. The FIR NN is trained by temporal backpropagation learning algorithm. A waveguide filter is used as an example and simulated by the FDTD method. It demonstrates that a short segment of an FDTD data is used to train the predictor, and the predictor can predict later information very well. The total least square (TLS) method is used as a predictor as well. By comparing the predicted error, it is shown that FIR neural network gives better prediction than that of the TLS.