Exascale Deep Learning for Scientific Inverse Problems