FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation

Delineating faults from seismic images is a key step for seismic structural interpretation, reservoir characterization, and well placement. In conventional methods, faults are considered as seismic reflection discontinuities and are detected by calculating attributes that estimate reflection continuities or discontinuities. We consider fault detection as a binary image segmentation problem of labeling a 3D seismic image with ones on faults and zeros elsewhere. We have performed an efficient image-to-image fault segmentation using a supervised fully convolutional neural network. To train the network, we automatically create 200 3D synthetic seismic images and corresponding binary fault labeling images, which are shown to be sufficient to train a good fault segmentation network. Because a binary fault image is highly imbalanced between zeros (nonfault) and ones (fault), we use a class-balanced binary cross-entropy loss function to adjust the imbalance so that the network is not trained or converged to predict only zeros. After training with only the synthetic data sets, the network automatically learns to calculate rich and proper features that are important for fault detection. Multiple field examples indicate that the neural network (trained by only synthetic data sets) can predict faults from 3D seismic images much more accurately and efficiently than conventional methods. With a TITAN Xp GPU, the training processing takes approximately 2 h and predicting faults in a 128 x 128 x 128 seismic volume takes only milliseconds.