A Bayesian Inference/Prediction/Control Framework for Optimal Management of CO2 Sequestration

The focus of the proposed work is on integrating research developments in scientific computing, statistical analysis, and numerical analysis to provide a common platform for managing CO2 storage. Results from this work will be important to energy production in the US, an area of National interest.

Geological carbon storage faces two main challenges: the risk of inducing seismicity, and leakage of the injected CO2 into potable aquifers. The characterization of the injection site and continued monitoring of the CO2 migration as well as stress changes in the region of elevated pressure are therefore particularly important to maximize the amount of CO2 that can be stored, while ensuring the long term safety of storage sites. To address these challenges, the overall goal of the proposed research is to (1) integrate well pressure and, where available, surface deformation data into coupled poromechanics models by solving the inverse problem for unknown subsurface properties; (2) to quantify the uncertainty in the inversion for the subsurface properties, and (3) to use the resulting inferred poromechanics models together with their uncertainty to design optimal control strategies for well injection that optimize the amount of stored CO2 while controlling the risk of seismicity. It is essential that this poromechanics based inference/prediction/control framework takes into account uncertainties at every stage, since both the observational data and the models are uncertain. However, solving stochastic inverse/optimal control problems for large-scale PDE models, such as those of poromechanics, is intractable using current methods, which suffer from the "curse of dimensionality." Thus, it is proposed to overcome these barriers by developing scalable methods and algorithms that exploit the problem structure to reduce effective dimensionality. While the end application of CO2 storage is quite important in itself, the framework to be developed can be applicable to a broader set of science and engineering problems for which large-scale uncertain models must be inferred from large-scale uncertain data, and then used to solve optimal decision-making problems under uncertainty.