Identification and Characterization of Faults Using Deep Learning

The identification and characterization of faults is an important process that provides necessary knowledge from the subsurface in geological and geophysical research. 3D seismic surveys are commonly utilized for the task of exploring the structural framework of the subsurface, as they enable the view of entire large structures in 3D. In seismic interpretation manual interpretation of faults is a tedious and complicated process, additionally the results are prone to human error. Another approach for interpreting faults on seismic data is to use attributes. Attributes that highlight discontinuity on seismic data have been used to detect faults. Although, these methods are not capable to evolve independently, thus constantly rely on the interpreters knowledge. Recently, Machine Learning (ML) techniques in general and Convolutional Neural Networks (CNN) as part of Deep Neural Networks (DNN) have been used to detect and image faults on seismic data with the aim of making the process more automated. CNN networks learn and evolve from manually annotated or labeled fault interpretations. In this study I have applied supervised CNN to image faults through binary segmentation, where faults are detected pixel-wise as ones and other background as zeros. The task was solved on 3D seismic surveys collected from three separate locations along the Norwegian Continental Shelf and the Efficient UNET and Light UNET CNN architectures were utilized to perform the task. Additionally, techniques such as data augmentation (geometric transformations) and hyperparameter adjustments were applied to improve the learning process and performance of the deep learning algorithms. The application of data augmentation to the training and testing data, generally led to improvement in the performance of CNN on fault predictions. Although the magnitude of improvement was varying with respect to the different surveys. The initial CNN fault prediction improvement mainly relied on the quality, and size of faults present in the 3D seismic volume. Further, improvement was achieved by the adjustment of certain hyperparameters affecting the training and testing process of the CNN. Regardless, little to no improvement was noticed particularly on one seismic volume containing high levels of noise. The characterization of fault geometries and the width of fault damage zone utilizing the best performing CNN was successfully conducted. Additionally, fault frequency plots and cumulative fault frequency p