Training an Inverse Operator on Shingling Seismic Data based on cGANs

We focus on a specific inverse problem, which is to work out the near-surface velocity model from shingling data that may be caused by reversed velocity sequences. The traditional inversion method first arrival traveltime tomography cannot handle such type of seismic data. A conditional generative adversarial network has the similar goal with an inverse problem. That is to generate model from given data, trying not to be distinguished from the true model given in training process. The network learns the relationship between data and model, especially the direction from data to model, and it can be seen as an inverse operator. We rebuild our network based on a well used general image-to-image cGANs framework, and create our datasets of shingling data through wave-equation based finite difference simulation method. After training 2D synthetic pairs of shingling seismic data and velocity model, we acquire a 100 …