Machine Learning Enhanced Interpretation of Wellbore Data for Underground Storage

ABSTRACT: The Illinois Basin has an excellent history of pilot- and demonstration-scale CO2 injection projects, positioning it as a promising potential hub for carbon capture and storage. Since the presence of fractures increases risk of an induced seismic response and CO2 leakage, it is essential that potential injection sites are properly characterized before project deployment. Accurately measured distribution of seismic waves velocities is essential for prediction of fracture locations and monitoring of the injected CO2. Site investigation for the pilot-scale CO2 storage projects yielded abundant high-quality data, especially in terms of geophysical wellbore measurements. This information is utilized to create a dataset specific to the Illinois Basin aimed at assistance during the wellbore and seismic data interpretation, which (as any inverse problem) heavily relies on physical constraints to limit the range of potential solutions. The developed machine learning algorithms allow predicting P- and S-wave velocities based on mineralogical composition from ELAN log with accuracy superior to conservative effective media approaches. Three deployed algorithms include random forest, gradient boosting, and neural network – all of them yielded significantly more accurate estimations of seismic velocities compared to the sonic log. Additionally, this technique reconstructs complex patterns in variations of velocities associated with heterogeneities of geologic formations. It opens ways for further enhancement of datasets to evaluate involved hydromechanical parameters (e.g., poroelastic properties and permeability) to refine the velocity models and make identification of subsurface features and monitoring of the fluid migration processes more accurate. 1. INTRODUCTION Successful implementation of geoenergy projects requires knowledge of the hydromechanical, stratigraphic, and lithological features of the participating geological formations. The access to the sites can be gained during the well boring (1D data), as well as via the use of geophysical methods (3D data), such as seismic, gravimetric, and magnetic surveys (Kearey et al., 2002; Manivannan, 2018). These techniques measure potential, resistivity, porosity, formation factor, temperature, fluid saturation, and sonic data obtained from various wellbore logs (Asquith et al., 2004), as well as mineralogy, lithology, and hydrocarbon content (Pelton et al., 1978). The datasets can contain thousands of records for each of the parameters as a function of depth, and combined, represent huge arrays of data to be interpreted. Machine learning (ML) techniques have been shown to serve as a powerful algorithm for enhancing the datasets and capturing intricate nonlinear relationships within them (Bressan et al., 2020; Glubokovskikh et al., 2020). The overall effort is in augmenting subsurface characterization efforts and enhancing understanding of the geological conditions.