Deep learning joint inversion of seismic and electromagnetic data for salt reconstruction
Yen Sun, Bertrand Denel, Norman Daril, Lory Evano, Paul Williamson, Mauricio Araya‐Polo
Abstract
Depth imaging projects dedicated to hydrocarbon exploration or field development rely heavily on velocity model building. When salt bodies are present, their accurate delineation is crucial to ensure the quality of seismic images, especially for sub-salt targets. We investigate a supervised deep learning (DL) approach which predicts salt geometry by using seismic and electromagnetic data simultaneously. Different network architectures were designed to incorporate these distinct data types and tested to assess which best improved the prediction accuracy compared to either seismic or electromagnetic alone. The networks were trained using synthetic data created from pseudo-randomlygenerated velocity and resistivity models. After training, the network was tested with synthetic data from models representative of real Gulf of Mexico salt geometries. The Intersection over Union (IoU) metric was used to compare the predictions from integrated multiple geophysical data and single domain data. The results demonstrated that the accuracy of the reconstructed salt geometry can be enhanced by combining seismic and electromagnetic data. While these results were obtained for 2D synthetic data only, and many significant challenges will need to be overcome to apply the method to real data, this study may shed light on the potential of machine learning approaches to Multiphysics joint inversion. Presentation Date: Tuesday, October 13, 2020 Session Start Time: 8:30 AM Presentation Time: 9:20 AM Location: 360A Presentation Type: Oral