loading page

Visualization of the sequestered carbon-dioxide plume in the subsurface using unsupervised learning
  • Keyla Gonzalez,
  • Siddharth Misra
Keyla Gonzalez
Texas A&M University, Texas A&M University

Corresponding Author:[email protected]

Author Profile
Siddharth Misra
Texas A&M University, Texas A&M University
Author Profile

Abstract

Subsurface sequestration of carbon dioxide (CO2) requires long-term monitoring of the injected CO2 plume to prevent CO2 leakage along the wellbore or across the caprock. Accurate knowledge of the location and movement of the injected CO2 is crucial for risk management at a geological CO2-storage complex. Conventional methods for locating/assessing the injected CO2 plume in the subsurface assume a geophysical model, which is specific and may not be applicable to all types of CO2-injection reservoirs and scenarios. We developed an unsupervised-learning-based visualization of the subsurface CO2 plume that adapts and scales based on the data without requiring an assumption of the geophysical model. The data-processing workflow was applied to the cross-well tomography data from the SECARB Cranfield carbon geo-sequestration project. A multi-level clustering approach was developed to account for data imbalance due to the absence of CO2 in the large portion of the imaged reservoir. The first level of clustering differentiated CO2-bearing regions from the non-CO2 bearing regions and achieved a silhouette score of 0.85, a Calinski-Harabasz index of 160666, and a Davies-Bouldin index of 0.43, which are indicative of high quality, reliable clustering. The second level of clustering further differentiated the CO2-bearing regions into regions containing low, medium, and high CO2 content. Overall, the multi-level clustering achieved a silhouette score, Calinski-Harabasz index, and Davies-Bouldin index of 0.74, 59656, and 0.32, which confirm the high quality and reliability of the newly proposed unsupervised-learning-based visualization. Three distinct clustering techniques, namely k-means, mean-shift, and agglomerative, generated similar visualizations. In terms of the adjusted Rand index, the similarity of clusters identified by the three distinct clustering techniques is around 0.98, which indicates the robustness of the cluster labels assigned to various regions of the CO2-injection reservoir. Further, we find certain geophysical signatures, such as Fourier transform and wavelet transform, to be highly relevant and informative indicators of the spatial distribution of CO2 content.