CarbonWatch: Rock Physics-Based Machine Learning Solution for End-to-End Monitoring of CO2 Storage Sites

CO2 capture and storage is a key technology for the mitigation of greenhouse gas emissions. Finding suitable geological storage sites, demonstrating effective CO2 containment, and establishing long-term monitoring of storage reservoirs are key to the safe storage of CO2. The risk of CO2 leakage can be reduced through real-time monitoring and early prediction of possible leaks. Early prediction of leaks involves identifying the onset of a leak and immediately visualizing the migration pathways. The traditional seismic imaging method has some disadvantages: (i) seismic imaging data is not always available, and it is expensive to conduct seismic acquisition and imaging; (ii) the application at larger depths is limited due to the attenuation of sufficiently high frequencies; and (iii) given the expense of seismic acquisition, processing, and analysis, it may be important to record repeat seismic surveys only at critical times. In this proposed research work, we plan to use the large amount of multiscale and multimodal data available in EDX for various formations to create machine learning-aided rock physics models. The aim is to use the data in the EDX platform to develop an end-to-end mapping between the rock and seismic properties. Based on the change in rock properties, the CO2 plume migration path can be visualized. Key parameters from these models will also be used to reduce the uncertainty in the ultimate reservoir storage cost and capacity by integration with reduced order models of dynamic CO2 injection. Rock physics modeling (RPM) development is driven by data availability: X-Ray Diffraction, PVT, and core measurements to provide rock and fluid properties, well logs to calibrate the model, and real-time seismic data to validate RPM-derived synthetic seismic. Therefore, this study is only possible because of the wide variety of data available in EDX for various storage formations. Data mining, therefore, will be an integral part of our proposed project. The biggest hurdle in RPM is the many uncertainties present in the model. While rock physics modeling forms the skeleton of what connects rock and seismic properties, supervised machine learning models become an essential component to reduce the uncertainties between the properties’ mapping and improve the accuracy of the models. Thus, we advocate a physics-based machine learning approach to RPM. The proposed CarbonWatch technology has huge benefits in both the economy and technology, and in minimizing an adverse environmental impact.