Geoscientist-driven machine learning-assisted fault interpretation: A case study from Atlanta Field to demonstrate the impact of fault labeling on the fault prediction cube

Abstract The energy industry is undergoing a digital transformation, leveraging cloud capabilities and artificial intelligence technology to overcome the challenges of conventional geoscience workflows. Machine learning (ML) methodologies have been developed and applied to various geophysical and geologic workflows, such as seismic processing, imaging, seismic interpretation, and petrophysical analysis. The traditional seismic interpretation approach requires manual interpretation of faults line by line at a fixed increment. Consequently, depending on the geologic complexity and size of the area of interest, the traditional approach can be a time-consuming and repetitive task. We focus on applying ML to seismic data for fault interpretation using a 2D convolutional neural network methodology to accelerate the fault interpretation process. This approach has the potential to significantly reduce the time and effort required to interpret faults, depending on the seismic data quality and geologic complexity, which is demonstrated by our study in the Atlanta Field, a postsalt asset in Brazil's northern Santos Basin. The basin has high structural complexity due to the salt tectonics, which required additional labeling interpretation to capture different fault patterns and train the algorithm to predict faults within the seismic volume. Additionally, the study provides the impact of the initial fault interpretation label on the ML result, such as in the fault extension and fault dip. Overall, the ML approach reduced the total effort from two weeks to four days, representing a time improvement of 60% from interpretation to structural framework workflow.