Fast tracking of safe CO₂ trapping indices using machine learning for smarter reservoir management

Predicting trapping indices is crucial for assessing the viability of carbon dioxide (CO₂) storage sites, particularly in mitigating risks associated with cap-rock failure. Accurate prediction of trapping mechanisms—structural, residual, solubility, and mineral—is critical to ensuring effective CO₂ containment and minimizing environmental impacts. The traditional determination of trapping indices mostly relies on high-resolution and expensive numerical reservoir simulations, which considers the full physics occurring in subsurface. However, engineers often overlook an important aspect related to the safe storage of CO₂ in subsurface, which mainly accounts the risk of CO₂ leakage due to mechanisms like cap-rock failure. To bridge this gap and promote smarter reservoir management, this study develops machine learning (ML)-based predictive models to rapidly and accurately estimate effective CO₂ trapping indices, emphasizing the critical influence of cap-rock failure. We use traditional numerical reservoir simulations to generate a substantial amount of long-term (200 years) GCS simulation data, which considers reservoir heterogeneity, overburden zone/cap-rock layer/storage zone formation setup, as well as the coupled mechanisms of multiphase flow, geochemical reaction and geomechanics. We detect the cap-rock failure time based on the dynamic change of bottom-hole pressure from the monitoring injecting well, and the cap-rock failure time is further used as the safe CO₂ injection duration to calculate the effective trapping indices. Reservoir and well parameters were used as inputs, with effective trapping indices as outputs, to train ML models, including, fully connected deep neural networks (FCNN), decision trees (DT), K-nearest neighbors (KNN), random forests (RF), and gradient-boosting methods (AdaBoost and XGBoost). Among these, FCNN demonstrated superior performance, achieving an average prediction error below 5% and a coefficient of determination (R²) exceeding 0.99 across all trapping indices. This study highlights the potential of data-driven ML models to predict long-term safe trapping indices with high accuracy, significantly reducing computation time compared to conventional numerical simulations. These findings offer a practical solution for efficient GCS site assessment and risk management.