Reconstructing Seismic Images and Creating Pseudo-3D Volumes: a Machine Learning Approach

We describe how machine learning can be used to solve a range of interpolation problems regularly encountered in post-stack seismic data. In all experiments we train a U-net type of deep learning model on examples extracted in good-quality data areas. The 2D or 3D input images utilized for our training set are manipulated such that the inputs exhibit the same kind of problems as observed in the areas with poor, or missing data. We repair these poor data areas by applying the trained model. We show examples of interpolating missing traces, decreasing the bin-size (quadrupling the number of traces), and replacing a bad data patch in an undershoot area. We also demonstrate that these trained models are generic interpolators that can be reused without retraining to solve similar problems in other data sets. Finally, we use the same technology to create pseudo-3D volumes from 2D data. We present two workflows: a direct transformation approach and a transformation that takes place in the flattened domain. The latter approach is more demanding as it involves interpretation followed by flattening and unflattening the data.