Field Development Plan Optimization Under Uncertainty Using Physics-Informed Neural Networks and Generative AI: A Real-Field Case Study in North America

Field Development Plan (FDP) optimization is a notoriously challenging, high-dimensional problem, often involving an immense number of variables and complex operational constraints. Traditional optimization approaches, which rely on iterative reservoir simulations, are computationally expensive and frequently yield sub-optimal solutions. This paper introduces a novel, integrated workflow for FDP optimization under uncertainty, harnessing the capabilities of Generative AI (Gen AI) and advanced physical surrogate models to address these challenges. The proposed methodology comprises three core components. First, a Generative AI model is employed to recommend realistic well locations, rigorously accounting for all operational constraints and practical considerations. Second, physical surrogate models are developed to emulate the behavior of high-fidelity reservoir simulators. These models are trained using physical reservoir flow concepts, enabling them to predict production rates with remarkable speed—delivering results in milliseconds—while preserving the same accuracy as a high-fidelity simulator. Third, a metaheuristic optimizer is integrated with the physical surrogates to efficiently navigate the high-dimensional optimization landscape and identify the optimal FDP configuration. This approach overcomes key limitations of conventional deep learning methods, which often require extensive datasets and may fail to honor physical laws. By embedding physical constraints directly into the surrogate model architecture, the physical surrogates ensure both speed and accuracy, bridging the gap between physics-based simulation and data-driven modeling. The Gen AI component is among the first to generate realistic well trajectories that comply with operational restrictions, uncovering new opportunities that traditional workflows might overlook. Furthermore, the use of metaheuristic optimization enables the efficient solution of high-dimensional problems within practical timeframes. Applied to a real-field complex North America reservoir which comprised 10 completed wells distributed across multiple zones within four distinct reservoir units, the workflow achieved a ~24% increase in expected net present value (NPV) and improved mean oil production across uncertainty scenarios, while maintaining low surrogate prediction error (~3%) even in extrapolation. Alternative uncertainty weightings revealed distinct optimal well placement patterns with only modest NPV reductions, demonstrating robustness and adaptability.