Seismic characterization of CO2 storage driven by time-lapse images of a prior injection using the artificial neural network

To optimize geologic CO2 storage and ensure its safety, it is necessary to demonstrate conformance between the reservoir simulations and geophysical monitoring such as time-lapse (TL) seismic. This process, known as history matching, often relies on subjective judgment and intuition of a reservoir modeling team because a direct examination of the multitude of plausible geologic scenarios is prohibitively expensive. The artificial neural network (ANN) aims to reconstruct the observed plume based on a set of seismic attribute maps. Via a randomized test, the trained model then provides an estimate of the importance of each attribute according to the attribute’s contribution to the accuracy of the plume prediction. This same test is also used to identify specific geologic controls for each part of the CO2 plume. The developed ANN is then used to forecast a CO2 plume that will likely arise from a future injection into the same formation 700 m away from the previous injection. The predicted map of the probability of the occurrence of CO2 after the future injection looks reasonable and agrees with existing reservoir simulations. At the same time, the neural network predicts some potential risks (e.g., across the fault migration) that were not considered in the fluid flow simulations. Although the neural network cannot fully replace high-fidelity fluid flow simulations, it can highlight geologic and petrophysical scenarios that should be simulated. Hence, our workflow may significantly improve the efficiency and accuracy of manual history matching.