Deep learning model of the geochemical impacts of carbon dioxide and brine leakage into overlying aquifers at geologic storage sites

One of the risks associated with geologic carbon sequestration is the potential for a containment loss event within the storage complex resulting in fluid leakage into overlying aquifers. Evaluating the impact to aquifers due to leaked carbon dioxide and brine requires estimating leakage rates through legacy wells or faults, establishing groundwater quality impact thresholds, and simulating potential leakage scenarios. Developing a site-specific reactive transport model for overlying aquifers can be time-consuming and challenging during the early characterization phase of a carbon storage project due to lack of data. To reduce the time required and uncertainty with the development of a new aquifer model at a site we have quantified the sensitivity of hydraulic, geochemical and leak parameters for aquifer impact predictions. This is used to develop a generic aquifer impact model for uncertainty quantification, with a simplified geochemical reaction network that captures relevant groundwater quality parameters and accurately predicts the extent of impacted groundwater.

We employ a conditional generative adversarial network developed for image-to-image translation. The generator is a U-Net based architecture, and the discriminator is a convolutional classifier, which only penalizes structure at the scale of small patches of the model grid. The U-Net is an encoder-decoder with skip connections between mirrored layers in the encoder and decoder stacks. Our network differs from previous implementations in that it utilizes 3D convolutional and deconvolutional layers, rather than 2D. The model is trained using the results of a suite of reactive transport simulations. The generator must learn to fool the discriminator, while also reproducing the reactive transport model results. The resulting surrogate model runs in less than a second, rather than hours required for the reactive transport model. This approach allows for rapidly estimating leakage impacts, quantifying parameter sensitivity, and capturing uncertainty without expending project resources on highly unconstrained data.