A semi-supervised learning framework for seismic acoustic impedance estimation

Abstract

Summary For compensating the limited bandwidth in seismic data, one reliable approach for robust acoustic impedance estimation is to integrate 3D seismic data with 1D well logs by building an optimal non-linear mapping function between them. However, most of the existing mapping methods, including these by machine learning, are performed in 1D that utilizes only the single seismic trace corresponding to a well. Therefore, their performance is restricted within a small zone around the wells, while consistent prediction cannot be obtained throughout the entire seismic survey. In addition is the down-sampling of high-resolution well logs to the seismic scale, which fails to fully utilize the information available in the wells. For resolving both limitations, this work presents a semi-supervised learning framework of two components: (1) seismic feature self-learning and (2) seismic-well integration, each of which is formulated as a deep convolutional neural network. The performance of the proposed framework is evaluated through an application to the synthetic SEAM dataset. The good match between the machine prediction and the earth model demonstrates the capability of the proposed semi-supervised learning in reliable seismic and well integration, particularly in the zones of poor seismic signals due to the presence of geologic complexities.