Prediction of Gas Hydrate Formation at Blake Ridge Using Machine Learning and Probabilistic Reservoir Simulation

Abstract

Abstract Methane hydrates are solid structures containing methane inside of a water lattice that form under low temperature and relatively high pressure. Appropriate hydrate‐forming conditions exist along continental shelves or are associated with permafrost. Hydrates have garnered scientific interest via their potential as a source of natural gas and their role in the global carbon cycle. While methane hydrates have been collected in multiple diverse geographic settings, their quantities and distribution in sediments remain poorly constrained due to sparse relevant data. Using statistical and machine learning approaches, we have developed a workflow to probabilistically predict methane hydrate occurrence from local microbial methane sourcing. This approach utilizes machine‐learned global maps produced by the Global Predictive Seabed Model (GPSM) as inputs for the statistical sampling software, Dakota, and multiphase reservoir simulation software, PFLOTRAN. Dakota performs Latin hypercube sampling of the GPSM‐predicted values and uncertainties to generate unique sets of input parameters for 1‐D PFLOTRAN simulations of gas hydrate and free gas formation resulting from methanogenesis to steady state. We ran 100 1‐D simulations spanning a kilometer in depth at 5,297 locations near Blake Ridge. Masses of hydrate and free gas formed at each location were determined by integrating the predicted saturation profiles. Elevated hydrate formation is predicted to occur at depths >500 meters below sea level at this location, and is strongly associated with high seafloor total organic carbon values. We produce representative maps of expected hydrate occurrence for the study area based on multiple realizations that can be validated against geophysical observations.

Keywords

• acoustic velocity
• free gas
• gas hydrate
• geospatial machine learning
• seafloor mapping