Application of neural networks to the inverse light scattering problem for spheres

A new approach suitable for solving inverse problems in multiangle light scattering is presented. The method takes advantage of multidimensional function approximation capability of radial basis function neural networks. An algorithm for training the networks is described in detail. It is shown that the radius and refractive index of homogeneous spheres can be recovered accurately and quickly, with maximum relative errors of the order of 10(-3) and mean errors as low as 10(-5). The influence of the angular range of available scattering data on the loss of information and inversion accuracy is investigated, and it is shown that more than two thirds of input data can be removed before substantial degradation of accuracy occurs.