Neural Networks Applied to Electromagnetic Compatibility (EMC) Simulations

Data extrapolation in FDTD simulations using feedforward multi-layer Perceptron (MLP) showed promising results in a previous study. This work studies two different aspects of the problem: First is the learning aspect, including the effect of prior training with the same class of random signals, which is an attempt to find a general solution to the weight initialization problem in adaptive systems. The second aspect covers the steps to make the extrapolator fully adaptive, through optimization of the time step sensitivity and the input layer width of a sliding window extrapolator. Average mutual information is used as a performance measure in most of the work.