Reducing the Risk of Gas Leaks into the Ocean Floor Induced by Offshore Production Well Failure in the Gulf of Mexico

In the Gulf of Mexico, there are tens of thousands of wells, between producing, shut-in, and abandoned ones. Considering the high number of wells having a risk of leaking gas into the surrounding formations, which may result in gas broaching at the seafloor, a deep understanding of the fate and transport of gas released from damaged wells is of special relevance for hazard assessment and prevention in offshore petroleum operations. This thesis explores a novel strategy to reduce the risk and impact of contaminant releases in the Gulf of Mexico by analyzing the applicability of machine learning technology as a tool to forecast the information regarding a possible broaching in a loss of containment scenario of an offshore well. In this thesis, a conceptual 3-D model of an area representative of the Gulf of Mexico was used to simulate the geosystem behavior for different scenarios of subsurface containment failure. A total of 20 heterogeneous permeability fields were initially used in the modeling of gas broaching. For each of the 20 heterogeneous permeability fields, this work tested 13 different locations where a hypothetical vertical well would be placed, and then a leakage point due to failure would lead to gas escaping from containment. The TOUGH+HYDRATE code for the simulation of system behavior in hydrate-bearing media was selected to be used in this work. For each simulation, we compiled the broaching day and location, the hydrate mass generated in the system until broaching day and at the end of the simulation, and the total released CH4 in gas phase in the system until broaching day and at the end of the simulation. The data generated from the different scenarios of well containment failure were compiled to be used as input data for training the Artificial Neural Network (ANN) models. We trained six networks to be data-driven models for the prediction of the different outputs. The data-driven models to predict the gas broaching time and location were created as functions of the cartesian coordinates of the leakage point and the permeability of all blocks in the vertical column above the leakage. The data-driven models to predict the hydrate mass and gas phase CH4 volume on broaching day and at the end of the simulation were created as functions of the gas broaching time and gas broaching location. From the six ANN models created and trained, four of them were able to find a strong or very strong correlation between the input and output features during training and when tested with the full dataset. We used a new permeability model to generate a new set of input and output parameters for measuring the network’s generalization. The new dataset was applied to the chosen ANN models predicting broaching location, broaching time, hydrate mass generated in the system at the end of the simulation, and the gas phase CH4 volume in the system at the end of the simulation. When presented with the validation dataset, the outstanding performance of these ANN models indicated their reliability in the prediction of the broaching parameters, and the significant generalization capability of the models.