Machine Learning for Rapid Evaluation of CO2 Storage Efficiency in the Weyburn Carbonate Reservoir

Reliable screening of geological CO2 storage sites requires methods that relate reservoir heterogeneity and injection history to screening indicators without running a full-physics simulation for every candidate scenario. This study presents a machine learning workflow for solubility trapping index (STI)-based screening in the Weyburn Oilfield, Canada, using 77 well records and six geological and operational descriptors: porosity, permeability, rock density, reservoir thickness, injection depth, and injection duration. The residual trapping index (RTI) is retained as a descriptive component of the analytical trapping index framework, while quantitative model comparison focuses on STI. Both indices are treated as well-level analytical screening quantities rather than direct physical measurements. Six regression algorithms were compared under random and spatial block cross-validation. In the displayed random-split comparison, XGBoost achieved the highest R2 (0.81), but spatial block validation reduced the aggregated R2 to 0.61, indicating that random-split performance was optimistic due to spatial dependence among nearby wells. Distance-buffer analysis further showed that prediction accuracy declined as training-test separation increased. Feature importance was evaluated using repeated permutation importance with uncertainty intervals and permutation-test p-values, supplemented by SHAP analysis to quantify the direction and magnitude of feature contributions. Injection depth and permeability were the most influential predictors in the fitted model, although the ranking is interpreted as an association within the analytical index framework rather than a causal conclusion. The workflow is proposed as a conditional, data-supported screening tool that requires reproducible spatial validation, documented target-index inputs, and further validation against high-fidelity simulation before quantitative storage conclusions are drawn.