Machine Learning–Based Surrogate Modeling and Screening Framework for CO2-Enhanced Oil Recovery and Permanent Carbon Storage in Mature Niger Delta Reservoirs

Abstract Mature Niger Delta reservoirs offer a compelling opportunity for integrated CO2-enhanced oil recovery and permanent geological carbon storage, but their heterogeneity, structural complexity, and limited data availability make conventional simulation-based screening workflows difficult to apply at scale. This study develops and validates a machine learning surrogate modeling and screening framework aimed at enabling rapid, data-driven evaluation of reservoir storage suitability and petrophysical properties, as a necessary precursor to full multi-objective CO2-EOR optimization. Well log data from two wells of the Sleipner CO2 Storage Project were selected as the analog dataset based on shared clastic depositional characteristics, demonstrated storage efficacy, and public data availability. Processing yielded 33,890 samples across key petrophysical measurements. A feature engineering workflow then derived geologically grounded variables including theoretical CO2 storage capacity, shale volume-based lithology classification, Kozeny-Carman permeability proxies, Reservoir Quality Index, Flow Zone Indicator, and a composite Caprock Integrity Score, together encoding injectivity, storage potential, and containment risk across all depth intervals.Three supervised regression models were trained to predict formation porosity as the primary surrogate target: linear regression, Random Forest, and XGBoost. All porosity-derived features were excluded from model inputs to prevent data leakage. XGBoost delivered the strongest results, achieving a test R2 of 0.9752 and a mean absolute error of 0.0028, a 67% reduction over the linear baseline, which confirms that significant nonlinear petrophysical relationships exist in the data and that the ensemble approach captures them effectively. Bulk density and neutron porosity ranked as the most influential predictors, consistent with their near-perfect physical correlations with formation porosity (r = -0.98). Lithological analysis identified clean sands at roughly 45% of the dataset and high-integrity caprock intervals at 18%, providing adequate geological diversity for surrogate generalization. Once trained, the surrogate models predict porosity and estimate storage capacity across thousands of geological realizations in milliseconds rather than simulation-scale hours, supporting uncertainty-aware screening in data-limited settings. The framework is designed as the foundational data-driven layer for subsequent coupling with dynamic reservoir simulation and evolutionary optimization, working toward full multi-objective CO2-EOR and storage optimization across the Niger Delta.