A Study of Conceptual Model Uncertainty in Large-Scale CO2 Storage Simulation

[1] In this study, multiscale permeability upscaling is combined with a sensitivity study of model boundary condition to identify an optimal heterogeneity resolution in developing a reservoir model to represent a deep saline aquifer in CO2 storage simulation. A three‐ dimensional, fully heterogeneous reservoir model is built for a deep saline aquifer in western Wyoming, where each grid cell is identified by multiple material tags. On the basis of these tags, permeability upscaling is conducted to create three increasingly simpler site models, a facies model, a layered model, and a formation model. Accuracy of upscaling is evaluated first, before CO2 simulation is conducted in all models. Since at the injection site, uncertainty exists in the nature of the reservoir compartment, end‐member boundary conditions are evaluated, whereby brine production is introduced to control formation fluid pressure. The effect of conceptual model uncertainty on model prediction is then assessed for each boundary condition. Results suggest that for the spatial and temporal scales considered, without brine production, optimal complexity of the upscaled model depends on the prediction metric of interest; the facies model is the most accurate for capturing plume shape, fluid pressure, and CO2 mass profiles, while the formation model is adequate for pressure prediction. The layered model is not accurate for predicting most of the performance metrics. Moreover, boundary condition impacts fluid pressure and the amount of CO2 that can be injected. For the boundary conditions tested, brine production can modulate fluid pressure, affect the direction of mobile gas flow, and influence the accuracy of the upscaled models. In particular, the importance of detailed geologic resolution is weakened when viscous force is strengthened in relation to gravity force. When brine production is active, variability of the predictions by the upscaled models becomes smaller and the predictions are more accurate, suggesting a subtle but important interplay between heterogeneity resolution, fluid driving forces, and model predictions.