Shallow Gas and Hydrate Hazard Prediction in the Gulf of Mexico Using a Probabilistic Machine Learning-Reservoir Simulation Workflow

Abstract

AbstractIn shallow marine sediments, gas accumulations and hydrates are significant geohazards to subsea infrastructure, drilling, and production. Therefore, predicting their occurrence is crucial to ensure offshore drilling safety and submarine infrastructure security. In this study, we generate predictions with uncertainty estimates with the goal of providing ageohazard assessment before any shallow hazard surveys are performed.We used a geospatial machine learning model to predict total organic carbon (TOC) and porosity at the seafloor in the northern Gulf of Mexico, and model sedimentation and consolidation in one dimension (1D) with microbial methanogenesis. Our model assumed that seafloor organic carbon is the source material for shallow hydrocarbon occurrences. The machine learning model outputs and uncertainties were sampled statistically to generate a suite of seafloor property realizations that were fed into our 1D model. Our predictions illustrate that gas and hydrate are more likely to be present along the shelf where seafloor TOC values are high, and are less likely to be present in deepwater areas (>500 m water depth) where seafloor TOC values are low. The results show that shallow water depth, lower sedimentation rate, and higher seafloor TOC are correlated with higher predicted gas saturations and shallower gas accumulation.Deepwater areas with significant reported oil production, such as AlaminosCanyon Block 857 (Great White), Green Canyon Block 640 (Tahiti/Caesar/Tonga), and Garden Banks Block 215 (Baldpate/Conger), are less likely to have shallow gas hazards. Any seafloor seeps identified in areas of high drilling activity likely originate from deep reservoirs, not from shallow gas accumulation.This work provides granular predictions of shallow geohazards on a basin scale and offers a holistic approach to identifying shallow hazards using big data and machine learning techniques. Leveraging geospatial machine learning models improves the predictions of subsurface hydrocarbons, despite sparse sampling of seafloor properties, and can be made pre-drill and without additional observations like sediment samples. This method can complement and augment existing hazard survey techniques.