Characterizing CO2 Sequestration Dynamics in Complex Geological Formations: Surrogate-Based Uncertainty Quantification Approach

Non-Gaussian characteristics are common in geological formations formed in complex depositional environments with distinct facies boundaries and highly contrasting hydrological properties. These non-Gaussian features pose significant challenges for characterizing CO 2 behaviors during geological carbon sequestration (GCS) projects. Because traditional physics-based numerical simulators often demand extensive computational resources for uncertainty quantification, this study develops a spatiotemporal surrogate model to expedite the process. The surrogate model is constructed upon a spatiotemporal convolutional-recurrent architecture, serving as a mapping operator that relates static non-Gaussian geological characteristics (permeability and porosity), and variable injection well configurations to dynamic evolution of CO 2 pressure and saturation. Subsequently, the proposed surrogate model substitutes the numerical simulator to facilitate uncertainty quantification through the iterative ensemble smoother (IES) algorithm. Results indicated that the proposed surrogate model effectively handles non-Gaussian geological characteristics and variable injection well configurations. It accurately replicated the performance of the MATLAB Reservoir Simulation Toolbox (MRST) simulator, achieving a coefficient of determination ( R 2 ) above 0.96 and a Pearson correlation coefficient (CC) above 0.98, and improving computational efficiency by a factor of 3,145 in forward modeling. Furthermore, integrating the surrogate model with the IES approach significantly reduced geological uncertainty. This research highlights the potential of combining deep learning techniques with data assimilation models for efficient forward modeling and uncertainty quantification in GCS projects.