Prediction of CO2 storage efficiency and its uncertainty using deep-convolutional GANs and pore network modelling

The overall capacity of CO2 storage is controlled by the morphological and flow characteristics of the pore space. Pore-scale imaging and modelling is widely used to predict storage efficiency, but its industrial application is limited due to time constraints and cost associated with high-resolution imaging techniques and computational resources required for data processing and flow simulations. In this study, a deep convolutional generative adversarial network (DCGAN) approach was introduced using Bentheimer sandstone tomographic images to quantify geological CO2 storage efficiency and its uncertainty. DCGAN was applied to generate an ensemble of realizations of three-dimensional porous media (1000 images with 5123 voxels in this study) and an enlargement factor of 83 in volume compared with the dimensions of the training data (643 voxels) was achieved. Then these images were fed into a network extraction and flow modelling simulator to determine macroscopic properties and their distribution. The average flow and geometric properties of the generated networks matched those of the imaging dataset, with a wider standard deviation and total range, showing that the realizations captured the full range of variability in pore structure. The predicted relative permeability and capillary pressure also matched experimental measurements in the literature. With a maximum capillary pressure of 7.0 kPa (representing initial injection into a storage formation), we found that the residual CO2 saturation is 0.354 ± 0.014. We further explored the minimum number of digital repeats needed to reproduce the statistics of geometric and flow properties from the 1000 cases. Overall, this work proposes a pore-scale simulation strategy coupling with deep learning to predict and assess storage efficiency and its uncertainty which can be used to guide the design of geological CO2 storage systems.