Predicting horizons for salt body models using machine learning from neighboring seismic surveys: A case study from the northern Gulf of Mexico

Abstract

The northern Gulf of Mexico has an abundance of seismic volumes, interpreted horizons, and their associated Earth models from different imaging projects. Every imaging and exploration project requires the interpretation of horizons to construct models for these newly acquired surveys in frontier exploration areas. Considerable time and manual effort are required to provide horizons that are input into the earth model building process. The quality of these horizons determines the resulting quality of the seismic volume. In this case study, we present a novel approach of using machine learning (ML) technology on northern Gulf of Mexico datasets to predict the position of the regional top and base salt interfaces in the process of building earth models for seismic imaging. This approach uses interpretations and seismic volumes from a neighboring survey with a similar geologic setting and comparable sediment and salt velocity floods as a training dataset to predict the position of the top and base of salt interface on a proximal seismic dataset. This ML capability provided substantial efficiency improvement by removing the manual time required for manual interpretation of newly acquired seismic volumes, resulting in the acceleration of model construction for seismic imaging. Furthermore, with ML assistance, the generated top and base horizons provide fine details along the model boundaries and consistently track amplitude events. The ML predicted top of salt (TOS) and base of salt (BOS) covered up to 85% and 75% of the entire survey, respectively. Increased efficiency due to ML led to improvements in turnaround time; this allows the geoscientists more time to enhance the finer points of the more complex geology, thereby yielding an enhanced structural representation of the subsurface leading to more accurate earth models.