Continuous conditional generative adversarial networks for data-driven modelling of geologic CO2 storage and plume evolution

Storing CO2 in geological formations requires accurate monitoring of the plume evolution for safe long-term operations, and traditionally, carbon storage practices rely on the numerical simulation of the multiphase flow and plume movement. Such simulations may be impractical for rapid, real-time monitoring, in which the risks involved are dynamically anticipated and/or mitigated, as they require the numerical solution of large non-linear systems of equations that also change over time. In this work, we investigate the application of continuous conditional generative adversarial networks (CCGAN) for predicting the CO2 plume evolution in cases in which limited data is available. We adapt the CCGAN for mapping the sparsely available input data from three wells to the spatially distributed CO2 saturation over the whole domain. Our main focus in this work is the accuracy and performance of the CCGAN when only sparse data is available. As such, we limit our investigation to the 2-D conceptual model of an aquifer and employ a simple immiscible and isothermal two-phase flow model. Our results show that the CCGAN can predict the CO2 plume from sparse data with at least 90% of accuracy while allowing a substantial reduction in the computational cost when compared to traditional methods based on numerical simulations, with a speed-up of no less than 700-folds. Future works may include the extension to 3-D reservoir models with irregular geometry.