We use convolutional neural networks (CNN) to predict scattering geometry from the fields outside of the scatterer. While this problem is nonunique, we show that by training on specific datasets, the CNN learns the underlying structure of the scatterers. I.e., if there is prior knowledge of the expected structure or form of the scatterers, this can be used to obtain a much more accurate solution to the inverse scattering problem. We show that our method faithfully recovers the original geometry for highly specific classes of structures, while the more conventional method falls victim to the nonuniqueness and fails to recover plausible-looking geometries.
|