A neural network approach for the differentiation of numerical solutions of 3-D electromagnetic problems

An innovative approach employing a neural network (NN) is presented to compute accurately derivatives and differential operators (such as Laplacian, gradient, divergence, curl, etc.) of numerical solutions of three-dimensional electromagnetic problems. The adopted NN is a multilayer perceptron, whose training is performed off-line by using a class of suitably selected polynomial functions. The desired degree of accuracy can be chosen by the user by selecting the appropriate order of the training polynomials. The on-line utilization of the trained NN allows us to obtain accurate results with a negligible computational cost. Comparative examples of differentiation performed both on analytical functions and finite element solutions are given in order to illustrate the computational advantages.