Delineating Faults in the Soda Lake Geothermal Field Using Machine Learning

Abstract

Accurate fault detection on seismic images is one of the most important and challenging tasks in the field of automatic seismic interpretation. Conventional human-picking and semi-human-intervened fault detection approaches are being replaced by fully automatic methods thanks to the development of machine learning. We develop a novel machine learning-based fault detection approach using a multiscale connection-fusion U-shaped convolutional neural network (MCFU for short). The most important characteristics of our MCFU method is that it uses skip connections to connect feature maps of different spatial resolutions, and uses a fusion operation to generate the final fault map. This architecture enables our MCFU to fully use spatial information embedded in the feature maps. We demonstrate through seismic migration images from the Soda Lake geothermal field that our MCFU produces cleaner, more interpretable fault maps for complex seismic images compared with the ant-tracking method that is widely used in industry, and with the conventional U-shaped convolutional neural network, thus leading to potentially improved geologic interpretability of detected faults.