Machine Learning Applications for Predicting Fault Reactivation Risks in CO2 Sequestration Sites

Abstract

Geological CO₂ storage is a critical component of net-zero emissions strategies, yet the injection of CO₂ into deep formations poses geomechanical risks including fault reactivation, induced seismicity, CO₂ leakage, and surface deformation, which must be systematically assessed before large-scale deployment. Traditional coupled geomechanical simulations are computationally prohibitive for multi-site screening, creating a need for faster, data-driven alternatives. This thesis develops and evaluates a physics-informed machine learning framework for predicting composite geomechanical risk at geological CO₂ storage sites. A multi-site database of 74 sites across 21 countries was compiled from peer-reviewed literature, comprising 31 geological, geomechanical, fault-structural, and operational input features. A novel four-class composite risk label (No Risk, Low, Medium, High) was constructed by integrating four failure-mode indicators through a data-driven ROC-AUC weighting scheme and operational vulnerability multipliers. Eight supervised classification algorithms: Random Forest, Gradient Boosting, Extra Trees, Logistic Regression, SVM, K-Nearest Neighbors, Decision Tree, and AdaBoost were trained and compared under identical stratified cross-validation conditions with GridSearchCV hyperparameter optimization. Random Forest achieved the highest macro-ROC-AUC of 0.809, substantially above the random-classifier baseline of 0.5, and correctly classified 68.9% of sites with 94.7% No Risk recall, supporting the framework as a first-pass screening tool. Permutation importance analysis revealed that Fault Proximity Class accounts for 93.6% of cumulative feature importance, directly contradicting the initial hypothesis that injection pressure would be the dominant control. Sensitivity analysis confirmed this: transitioning sites from far-field (greater than 10 km) to near-field (less than 5 km) fault proximity produced a 3.8-fold increase in combined elevated risk probability, while doubling injection pressure shifted high-risk probability by only 0.7 percentage points. Empirical operational thresholds were derived for all four failure modes, with the 2.7 MPa gap between the fault reactivation threshold (24.7 MPa) and the seismicity threshold (27.4 MPa) suggesting a potential pressure interval where aseismic slip may precede seismic rupture, qualitatively consistent with field observations. The framework provides a reproducible, computationally efficient screening methodology that complements site-specific physics-based simulation and delivers transparent feature rankings applicable to regulatory and site-selection decision-making.