Feed-forward neural networks applied to the estimation of magnetic field distributions

The Finite Element and Finite Difference methods are both widely used in estimating magnetic field solutions. Both methods are based on refining an initial estimate oft a solution using an iterative process; unfortunately, this rarely contains knowledge of the most likely correct solution, which has the potential of reducing the iteration time. Feed-forward neural networks may provide the bridge to provide the initial estimate.This thesis reviews the basic framework of feed-forward neural networks, specifically Multi-Layered Perceptron (MLP) networks and Basis Function networks, which are, in turn, subdivided into Radial Basis Function (RBF) networks and Wavelet Basis Function (WBF) networks. Included are discussions on the latest innovations in the field of artificial intelligence, the methods adapting neural networks to estimate magnetic field solutions, input/output configurations, and neural network initialization and training. Finally, simulation results and a discussion are provided, including suggestions on possible improvements