A hybrid simulation-machine learning proxy model for waterflood design optimization in the Bahariya Formation

Abstract This study presents a novel hybrid methodology combining machine learning (ML) with conventional reservoir simulation to optimize waterflooding in the geologically complex Bahariya Formation, Western Desert, Egypt. The research addresses the critical need for accurate oil recovery efficiency (RF%) prediction by developing and deploying robust data-driven models. Linear regression quantified the impact of nine key parameters across three injection patterns. A pivotal finding from numerical simulation was the ranking of Ultimate Oil Recovery (UOR) for each pattern under identical conditions: Peripheral flooding achieved the highest recovery at 44.7%, followed by Staggered Line Drive (SLD) at 39.4%, and the 5-Spot pattern at 33.7%. Our ML models demonstrated exceptional predictive accuracy, with R² scores of 0.974, 0.972, and 0.953 for the respective patterns, and a correspondingly low RMSE range of 0.0057–0.0085. Permutation importance analysis quantified the dominant influence of residual oil saturation (Sor), accounting for 38–42% of predictive power. Crucially, the models revealed distinct, pattern-dependent control parameters: injection rate (WINJ) showed markedly higher sensitivity in the Peripheral pattern (23% contribution), while API gravity was the second most important feature for the 5-Spot pattern (18% contribution). The findings provide a validated, efficient framework for rapid waterflooding scenario screening and optimization. This work highlights the substantial potential of hybrid AI-numerical approaches to enhance decision making and challenges conventional assumptions about pattern selection, demonstrating that the optimal pattern is profoundly dependent on specific reservoir characteristics. Field engineers can apply these insights to optimize injection strategies during reservoir development planning, potentially increasing recovery factors while reducing reliance on time-intensive simulation runs.