3D Convolutional Neural Networks for Fault Interpretation

Summary Faults in 3D seismic volumes are identified with a machine learning approach. For this task, a 3D convolutional neural network (3DCNN) is designed to produce a heat map of fault locations in a given seismic volume. Then, the heat map yields fault picks that can be used for building fault planes and surfaces. Using 3D convolutional filters, as opposed to 2D, allows the algorithm to extract a feature set for classification that merges information from all dimensions available. This ability translates in better prediction accuracy. The 3DCNN incorporates batch normalization and dropout layers to accelerate convergence and minimize over-fitting. Initial results on a 3D field data cube show a 88 % accuracy in predicting fault locations.