Seismic Acoustic Impedance Estimation by Learning from Sparse Wells via Deep Neural Networks

Summary Seismic acoustic impedance is one of the most important properties closely related to the subsurface geology, and thus robust acoustic impedance estimation from seismic data is an essential process in subsurface mapping and reservoir interpretation. For compensating the limited bandwidth in seismic data, one feasible approach is to integrate 3D seismic volume with 1D wells that are usually sparsely distributed within a seismic survey, and such integration aims at finding the optimal non-linear mapping function between them. Most of the existing mapping methods, particularly these powered by machine learning, are performed in 1D and/or require down-sampling of well logs to the seismic scale, which run of the risk of limiting the estimation valid only around the training wells and fail to provide consistent prediction throughout the entire seismic survey. We present a semi-supervised learning workflow for estimating the acoustic impedance over a given seismic survey by learning from a small number of sparsely-distributed wells via two deep neural networks. Applications to the synthetic SEAM dataset of a complex salt intrusion demonstrates its capability in reliable seismic and well integration, particularly in the zones of poor seismic signals due to the presence of geologic complexities, such as saltbodies.