An intelligent prediction of CO2 leakage in underground geological storage sites

Geological carbon storage (GCS) has emerged as a critical technology for mitigating anthropogenic CO 2 emissions, offering the potential to sequester billions of tonnes of CO 2 annually in deep subsurface formations. However, ensuring long-term containment integrity remains a significant challenge due to the complex risk of CO 2 leakage through subsurface pathways, including faults, fractures, and poorly sealed wells. This study develops a data-driven framework for predicting CO 2 leakage probability across five major underground storage types: salt caverns, saline aquifers, depleted oil and gas reservoirs, coal seams, and basalts. A comprehensive dataset of 467 data points encompassing 14 key geological, operational, and geochemical parameters was compiled from prior experimental and field studies. Five machine learning algorithms (AdaBoost, CatBoost, XGBoost, Decision Tree, and Random Forest) were systematically trained, optimized using Optuna hyperparameter tuning, and evaluated using RMSE, MAE, MAPE, R 2 , and SHAP-based interpretability. CatBoost consistently outperformed all other models, achieving the lowest RMSE (0.394), MAE (0.168), MAPE (0.445%), and highest R 2 (0.9997), with minimal training-cross-validation gap. Random Forest ranked second (R 2 =0.9996), followed by XGBoost (R 2 =0.9991). Learning curves confirmed CatBoost’s superior convergence, while Decision Tree exhibited pronounced overfitting. SHAP analysis revealed a critical finding: tree-based models rely almost exclusively on permeability, whereas CatBoost distributes importance more realistically across permeability, porosity, temperature, pressure, and injection rate, capturing the multivariate nature of CO 2 leakage mechanisms. Cross-site validation showed CatBoost produces predictions closest to input data averages across all storage types. The framework reduces assessment time from days of numerical simulation to minutes, offering a reliable, interpretable, and computationally efficient tool for site screening, operational monitoring, and risk management. The findings enhance understanding of parameter importance in controlling leakage risk across diverse storage environments, supporting informed decision-making for site selection, regulatory compliance, and adaptive injection strategies, thereby advancing the safe and scalable deployment of GCS for climate mitigation.