Electromagnetic inverse scattering using neural networks

A computer simulation study was carried out on the application of neural networks to a specific electromagnetic (EMF) inverse scattering problem. Small dielectric spheres (radius=1 mm) ranging in relative permittivity from 2 to 100 were individually exposed to a transverse electromagnetic (TEM) plane wave. The spheres´ centers were located in the x-y plane, with 0⩽x,y⩽1 cm. A sensor grid with 25 “point” sensors sampled the EMF scattered by each sphere. The EMF data was evaluated at a frequency of 1 MHz, using a dipole approximation. A principal components analysis was applied to the EMF data to derive a feature vector with 4 components, which were input to a feedforward neural network with one 15 unit hidden layer and 3 outputs, for training. The 3 desired outputs were the x and y coordinates of the sphere center and the relative dielectric permittivity of the sphere, respectively. For an independent EMF (testing) data set, the resulting, root mean-square (RMS) errors were quite small, being 0.03 mm, 0.05 mm, and 1, for the x- and y-sphere center and sphere permittivity estimates, respectively. The results of this study demonstrate the neural networks may be useful in the solution of inverse scattering problems