Date of Award
December 2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Petroleum Engineering
First Advisor
Kegang Ling
Abstract
Understanding subsurface properties is fundamental for optimizing hydrocarbon recovery and ensuring the long-term integrity of carbon capture and storage (CCS) operations. Conventional methods such as empirical correlations, advanced logging, and laboratory measurements, though widely used, are costly, time-consuming, and often limited in accuracy within formations. This study develops an ML framework to enhance the prediction of petrophysical and geomechanical properties and experimental investigation to evaluate time-dependent alterations in petrophysical and geomechanical rock properties under CO₂-brine exposure.In the first phase, six ML algorithms Decision Tree, Random Forest, Extra Trees, XGBoost, LightGBM, and K-Nearest Neighbor were applied to predict key petrophysical parameters (porosity, permeability, and water saturation) in the Bakken formation using well log and core analysis data. Ensemble-based models, particularly Extra Trees, delivered superior predictive accuracy, achieving R² values of 0.98, 0.89, and 0.86 for water saturation, porosity, and permeability, respectively.
The second phase used a hybrid ML framework to predict geomechanical properties in the Williston Basin. Four optimized algorithms Random Forest, Extra Trees, XGBoost, and LightGBM were trained on well log data to estimate shear (DTSM) and compressional (DTCO) wave transit times, along with derived parameters including Poisson’s ratio, Young’s modulus, and shear modulus. Extra Trees achieved the highest accuracy, with DTSM and DTCO predictions yielding R² values of 0.97. Compared to Castagna’s empirical correlation, ML-based predictions provided closer alignment with measured logs, underscoring the practical value of ML workflows as cost-effective and reliable alternatives to laboratory-based methods in geomechanical property estimation.
The third phase experimentally investigated the effects of supercritical CO2-brine exposure on Madison carbonate rocks under simulated reservoir conditions (140°F and 4,000 psi). Samples exposed for 10 days exhibited increased porosity and enlarged pore structures due to calcite dissolution, while those exposed for 30 days showed reduced porosity linked to secondary mineral precipitation (dolomite, quartz, halite, magnesium oxide). SEM and XRD confirmed dissolution–precipitation processes, while NMR revealed time-dependent pore evolution. Mechanical testing indicated increased Young’s modulus and Poisson’s ratio, suggesting matrix stiffening through mineral re-cementation. These results highlight a dual-phase mechanism of initial weakening followed by structural strengthening, with important implications for long-term CO2 storage security.
Collectively, this dissertation demonstrates that machine learning offers a powerful and cost-effective framework for predicting subsurface properties, while experimental insights into CO₂-brine-rock interactions provide critical understanding of subsurface propoerties evolution under storage conditions. The novelty of this work lies in bridging data-driven prediction with experimental validation to deliver an integrated approach that advances unconventional reservoir characterization and supports safer, more predictable CCS implementation.
Recommended Citation
Nkok, Luc Yvan, "Integrated Machine Learning For Predicting Reservoir Properties And Experimental Approach For Evaluating Co2-Brine-Rock Interactions In Carbonate Formations" (2025). Theses and Dissertations. 8239.
https://commons.und.edu/theses/8239