Physics-Informed Sampling Scheme for Efficient Well Placement Optimization

Well placement optimization (WPO) is a crucial task in terms of oil and gas recovery and economics in the field development plan. WPO poses significant challenges due to the multitude of local optima, which demand massive computational cost for global search algorithms. To address this, many proxy models have been applied for replacing reservoir simulations in many cases. Among these, convolutional neural networks (CNN)-based proxy models utilizing streamline time of flight maps as input demonstrated excellent performances. Nevertheless, these models exhibit diminishing performances during optimization processes, so additional retraining processes are required for successful results. In this study, we propose an initial sampling scheme using physics-informed quality maps incorporating static and dynamic information. The quality maps combine drainage area with permeability to represent the quality of each reservoir grid. The proposed scheme provides better performance than other sampling schemes. We demonstrate that the proposed scheme provides efficient well placement optimization regardless of the number of samples without retraining.