Informed reduction of the geological data on rock material interfaces in subsurface CO2 storage reservoirs

Recent developments in subsurface data capturing technologies have presented the opportunity of representing high-resolution information in field scale reservoir models of CO2 geo-sequestration sites. While the wealth of geological data is critical for the accurate prediction of CO2 flow and trapping in the subsurface, it often presents computational challenges. This is especially true for the representation of rock, or lithological, interfaces which are small-scale features in reservoirs and significantly control CO2 migration and trapping. It is important to reduce this information to ensure that numerical simulations are completed within a reasonable time. However, care must be taken to ensure that none of the key lithological interface associations are lost during data reduction. This study presents a machine learning based approach for expressing the amount of rock interfaces characteristic of sedimentary CO2 storage reservoirs in terms of a reduced number of principal components which capture the necessary geological information. The workflow is applied to a high-resolution reservoir model of the Paaratte Formation, Otway Basin, Australia, which is coastal to shallow marine siliciclastic reservoir. The in-formation in the reservoir model was captured using 922 k rock interfaces. However, the algorithm predicts that the same information could be represented using only 7 principal lithological interface associations. The out-comes are validated using multiphase flow simulations which show the uniqueness of these principal components in terms of their impact on CO2 flow and trapping.

Automated summary

Recent developments in subsurface data capturing technologies have provided the opportunity to represent high-resolution information in reservoir models of CO2 geo-sequestration sites. However, representing the small-scale rock interfaces poses computational challenges and data reduction must be done carefully to avoid loss of key lithological interface associations. This study presents a machine learning based approach to express rock interfaces in terms of a reduced number of principal components for accurate CO2 flow and trapping prediction. The results are validated using multiphase flow simulations.