Assisted Fault Interpretation by Multi-scale Dilated Convolutional Neural Network

Abstract

Summary Assisted fault interpretation leveraging machine learning techniques has become a promising way to identify faults in seismic. In geophysical exploration, faults are often considered as a sealing surface which traps hydrocarbons and forms reservoir zones. Thus, correctly identifying fault locations is critical. Fault identification can be treated as a semantic segmentation issue where we classify each seismic pixel into one of a given set of categories, such as fault or non-fault. To be successful we need to combine pixel-level accuracy with global-level feature identification. In this abstract, we propose a novel deep learning network with multi-scale dilated convolution to identify fault locations. It is based on adaptions of a convolutional neural network architecture which has been used for image classification and semantic segmentation. The motivation is that dilated convolution supports exponentially expanding receptive fields without losing resolution or coverage. We implemented multiple dilated convolution layers with variable dilation rates to systematically aggregate multi-scale seismic information. Several tests are shown and demonstrate the improvement of identification accuracy with higher resolution.