Perpustakaan Digital ITB

Optimal placement of infill wells is vital for increasing oil recovery, but evaluating each possible location is computationally expensive with numeric reservoir simulations. In this study, a machine learning surrogate for infill-well optimization of the PUNQ-S3 reservoir is developed, which relates the values of the SOI, DTOF, flow capacity, hydrocarbon pore volume, inter-well spacing, and the grid coordinates to the dFOPT, utilizing the XGBoost algorithm. Cross-validation based on the k-fold and leave-one-out methods is applied to validate the model. Uncertainty in predictions is assessed using CQR, which allows risk-aware optimization through GA and NSGA-II algorithms. In such cases, an evaluation function is based on a surrogate, which is improved using the UCB active learning approach under a limited number of simulations. Thus, the initial 50 points obtained with the help of LHS are increased to 62 points with UCB acquisition. Blind exhaustive testing against the simulator results in the R² values between 0.596 and 0.721 and Pearson's r values between 0.787 and 0.855, while CQR achieves approximately 80% coverage of uncertainty intervals. Compared to full-scale simulator validation, risk-aware optimization finds near-optimum points: Infill-1 at rank 5 (99.3% of the absolute optimum) and Infill-2 at rank 7 (96.3% of the absolute optimum). Further study is required to confirm the methodology, particularly for larger and heterogeneous fields.