Toward generalized models for machine-learning-assisted salt interpretation in the Gulf of Mexico

Interpreting salt bodies in the Gulf of Mexico (GoM) can be complex due to various factors affecting the accuracy of automated techniques. Variability of salt structures, seismic acquisition parameters, and imaging algorithms can impact the resulting seismic image. These differences can result in variations in seismic resolution and texture, making it challenging to develop automated interpretation techniques that are accurate and reliable for identifying salt bodies in the GoM. However, using seismic images with similar acquisition parameters and processing methods minimizes these differences and makes machine-learning (ML) models applicable. Utilizing nine data sets from the eastern GoM, a nine-fold cross-validation technique was applied to measure the generalization performance of the ML model. This method involves using one data set as the test set and the remaining eight for training and repeating the process for all subsets. We further applied an ensemble of the nine models to predict salt on a new unseen survey in Green Canyon. The study aimed to illustrate how salt variability and morphology in the GoM can impact the ability of the ML algorithm to predict salt bodies on unseen data.