Inverse scattering in one-dimensional random media using deep learning

Abstract

The inverse scattering problem is in general ill-posed and highly nonlinear. The aim of this thesis is to develop a fast algorithm that provides solutions to such inverse scattering problems in compactly supported one-dimensional random media. A promising candidate for this nonlinear task is Deep Learning, which showed great success in the recent past. The methodology of this approach is to train an Artificial Neural Network for a stochastic class of samples on numerically generated data. Inverse scattering is then performed by means of a simple forward-pass through the Artificial Neural Network. It is shown that in cases where the inverse scattering problem has a unique solution and where the scattering is not too strong, an Artificial Neural Network is able to solve the inverse scattering problem more efficiently than preexisting methods.