Rapid Prediction and Optimization of Multi-Regime CO2-EOR and Geological Storage Using Compositional Simulation and Trustworthy Surrogate Models

Integrated CO2-enhanced oil recovery (CO2-EOR) and geological storage require simultaneous optimization of oil production and CO2 retention. This study developed a surrogate-assisted framework integrating CMG-GEM compositional simulation, machine learning, and reliability-based screening. Three regimes were evaluated: full-period WAG (0Y) and 5- or 10-year continuous CO2 pre-injection followed by WAG (5Y and 10Y). Based on 1500 simulation samples, LightGBM surrogate models were developed to predict 30-year cumulative oil production (Np) and net CO2 storage. The 5Y and 10Y regimes generally increased Np and net CO2 storage but reduced net CO2 storage efficiency, revealing a regime-level capacity–efficiency trade-off. SHapley Additive exPlanations (SHAP) analysis identified initial water saturation and injection intensity as key predictors of Np, whereas WAG fraction dominated net CO2 storage predictions. Base-scenario Group Holdout validation yielded R2 values of 0.905 and 0.977 for Np and net CO2 storage, respectively. Trusted-domain screening, uncertainty assessment, and engineering constraints, followed by Pareto analysis, reduced 30,000 candidate schemes to nine representative solutions for independent GEM verification, with mean absolute percentage errors of 1.88% and 4.87%, respectively. These findings demonstrate that surrogate-assisted screening can efficiently narrow the design space, while physics-based verification remains essential for reliable CO2-EOR and storage optimization.