A NEURAL-NETWORK APPROACH FOR THE SOLUTION OF ELECTRIC AND MAGNETIC INVERSE PROBLEMS

Multilayer neural networks, trained via the back-propagation rule, are proved to provide an efficient means for solving electric and/or magn etic inverse problems. The underlying model of the system is learned b y the network by means of a dataset defining the relationship between input and output parameters. The merits of the method are illustrated at the light of three example cases. The first two samples deal with i nverse electrostatic problems which are relevant for nondestructive te sting applications. In a first problem, a boss on an earthed plane is identified on the basis of the map of potential produced by a point ch arge. In the second problem, the geometric parameters of an ellipsoid carrying an electric charge are identified. In both cases, database of simulated measurements has been generated thanks to the available ana lytical solutions. As a sample magnetic inverse problem, the identific ation of a circular plasma in a tokamak device from external flux meas urements is carried out. The results achieved show that the method her e proposed is promising for technically meaningful applications.