A Physics-Informed Machine Learning Framework for Producer–Injector Well Configuration Optimisation: Surrogate Screening, Sensitivity Analysis, and Coupled Surface–Subsurface Validation
In reservoir and drilling engineering, maximizing the economic recovery of hydrocarbons requires a complex balancing act between injection strategies and production constraints. Traditional reservoir management often relies on manual iterative simulations that are computationally expensive and struggle to navigate the high-dimensional, non-linear landscape of fluid flow in porous media. This study addresses the inefficiencies of conventional optimization by proposing a machine learning-based framework for the optimal configuration of producing and injection wells. The study proposes and validates a Physics-Informed Machine Learning (PI-ML) framework that integrates three complementary components: a sensitivity analysis module applied to 500 waterflood simulations on the Egg Benchmark Model to identify the dominant configuration drivers and quantify the inherent predictive complexity of static features; a surrogate classification system using ensemble machine learning to rapidly screen well configurations as GOOD, MODERATE, or POOR performers; and a synthetic reservoir optimisation module using an IMPES finite-difference simulator calibrated to SPE-10 statistics, within which a physics-penalised reward function guides configuration selection. Sensitivity analysis of the 500-run Egg Model dataset revealed that no single static feature (well location, permeability, or spacing metric) achieves a Pearson correlation exceeding |r| = 0.10 with cumulative oil, confirming that well placement optimisation is fundamentally a high-dimensional, nonlinear system-level problem. This finding motivates the use of more sophisticated sequential decision frameworks. The surrogate Random Forest classifier achieved a cross-validated accuracy of 49.2% on a three-class problem (compared to 33.3% random baseline), with the Gradient Boosting classifier achieving the highest F1-Macro of 0.327. On the synthetic reservoir, the physics-informed optimised configuration achieved a Recovery Factor of 24.67% compared to 6.17% for the five-spot baseline which is a 300% relative improvement, with an NPV uplift of +529% attributed primarily to more efficient injector placement in high-permeability zones. When the model was benchmarked against published DRL frameworks, the proposed methodology is consistent with the 15–33% NPV improvements reported by Nasir and Durlofsky (2022) for comparable PPO-based systems, and the runtime reduction of 83% relative to brute-force simulation is competitive with state-of-the-art approaches. The framework provides a scalable, data-efficient pathway for field development engineers to rapidly screen, classify, and optimise well configurations without requiring full-physics simulation of every candidate.