ABSTRACT: This study investigates machine learning (ML) techniques to enhance geomechanical characterization for potential geological carbon storage (GCS) in the Uinta Basin. The Mancos Shale Sealing Zone (MSZ) and the Dakota Sandstone Reservoir Zone (DRZ) were of specific interest. We compared the performance of several ML models and found Gaussian Process Regression (GPR) achieving the highest accuracy in predicting compressional acoustic slowness (DT), a crucial parameter for estimating key geomechanical properties. Using GPR-derived DT, we estimated Young's modulus, Poisson's ratio, shear modulus, and bulk modulus. Our results demonstrate a strong correlation (R2 0.84 in the MSZ) between geomechanical properties calculated using predicted and measured sonic logs. The estimated parameters for MSZ, benefiting from a larger dataset (n = 24,960 data points), exhibited higher prediction accuracy. The DRZ, characterized by interbedded sandstone and shale with limited data (n = 1,632 data points), generated less accuracy (R2 0.79 in the DRZ). These results emphasize the importance of data availability for effective ML model training and the need for more comprehensive datasets to improve predictions in complex geological settings. This study showcases the potential of ML to provide cost-effective and scalable solutions for reservoir characterization, ultimately contributing to improved risk assessment and decision-making for CO2 storage projects.
Bakelli et al. (Sun,) studied this question.