Predicting the CO₂ propagation in geological formations from sparsely available well data

The storage of CO₂ in geological formations is dependent on many uncertainties and poses as a challenge for the accurate description of the fluid flow pattern in the porous media where the carbon is stored. Conversely, accurate monitoring of the plume evolution is required for safe long-term operations, which is traditionally carried through the numerical simulation of the multiphase flow. These simulations require solving large non-linear systems of equations, thus precluding real-time monitoring with such tools, in which we dynamically anticipate and/or mitigate the risks involved with the CO₂ storage. In this work, we propose the adaptation of continuous conditional generative adversarial networks (CCGAN) for a data-driven model of geologic CO₂ storage and plume motion. The proposed model works in a sparse setting, meaning that it maps the sparsely available input data from three wells to the CO₂ saturation over the whole domain. The obtained results show that our model enables fast prediction of the CO2 plume with reasonable accuracy, by conferring a substantial reduction in the computational cost when compared to traditional numerical simulations.