Dynamic Calibration of Saturation in Reservoir Simulation Initialization

Abstract Accurate initialization of water saturation is a critical step in reservoir simulation, particularly in heterogeneous carbonate reservoirs where dynamic parameters such as capillary pressure – water saturation functions are often burdened with high uncertainty. This study introduces a novel, data-driven workflow to calibrate dynamic saturation initialization by optimizing these uncertain parameters using a machine learning (ML)-based proxy model. The workflow aligns simulation-initialized saturation with well-log-derived saturation through systematic variation of capillary pressure endpoints, water-oil contact, and rock typing scenarios, aiming to improve initialization quality and simulation performance. The proposed approach was implemented on an open-source carbonate reservoir model composed of 144 geological realizations, grouped into three distinct rock typing scenarios. In this study, three representative realizations corresponding to the 3, 5, and 7 rock type configurations were selected to evaluate which rock typing scheme best represents the reservoir conditions. For each rock typing scenario, 200 simulation cases were generated by varying capillary pressure endpoints and water-oil contact depths. The saturation mismatch between simulation output and well-log-derived saturation was computed for each case. Machine learning regression models were trained using these outputs and inputs to predict saturation mismatch. Multiple ML algorithms were evaluated, and random forest regression provided the best performance. Using the trained model, the dynamic parameters were then optimized to minimize the predicted saturation mismatch, enabling more accurate and efficient initialization. Two calibration approaches were investigated: Method 1, which predicts and optimizes the average saturation mismatch per simulation case, and Method 2, which predicts and optimizes saturation mismatch at each depth point. The results indicate that Method 2 provides superior calibration performance, with the five-rock-type model yielding the lowest saturation mismatch after optimization. The random forest model prediction of this model achieved a high predictive performance, with an R² score of 0.933 in estimating saturation mismatch across the simulation dataset. The optimization process led to a reduction in average saturation mismatch by more than 50%, improving from 0.34 to 0.17. As a result, the calibrated initialization contributed to more stable numerical performance in the reservoir simulator, reducing simulation runtime by approximately 52%. Furthermore, the improved saturation distribution facilitated faster and more reliable history matching. This case study demonstrates that integrating machine learning into the dynamic initialization process can significantly reduce uncertainty, improve model representativeness, and increase efficiency across the reservoir modeling workflow. This study presents a novel integration of machine learning and dynamic parameter optimization to calibrate reservoir simulation initialization. It provides a systematic, automated alternative to manual trial-and-error, reducing effort while improving model quality and stability. Additionally, the approach provides insight into which rock typing configuration best represents the reservoir, thereby enhancing uncertainty quantification and supporting more informed decision-making in reservoir development planning.