Fault Detection in Seismic Data Using Graph Attention Network

Patitapaban Palo, Aurobinda Routray, Sanjai Kumar Singh

Abstract: This paper presents a natural attention-based approach for automated fault detection in seismic data. Fault analysis in seismic data is important for drilling and exploration in the oil and natural gas industries. The seismic fault is a perceptual phenomenon, and manual fault detection is still practiced in various industries. The convolutional neural network (CNN) is the most commonly used method in the newly conducted research for automated fault detection. However, our paper uses a graph attention network (GAT) based approach. We first extract 2D patches centered around the points of concern. Next, we present these extracted patches in the graph domain using the k-nearest neighbor graph. The graph representation of patches is connectional in the graph domain based on seismic amplitude similarity. Then, we apply GAT to classify the faults. Both the training and testing sets contain both synthetic and real data. The proposed methodology gives good accuracy when applied to field data.