SummaryThe initial hydrocarbon saturation has a major effect on field-development planning and resource estimation. However, the bases of the initial hydrocarbon saturation are indirect measurements from spatially distributed wells applying saturation-height modeling using uncertain parameters. Because of the multitude of parameters, applying assisted-matching methods requires trade-offs regarding the quality of objective functions used for the various observed data. Applying machine learning (ML) in a Bayesian framework helps overcome these challenges. In the present study, the methodology is used to derive posterior parameter distributions for saturation-height modeling honoring the petrophysical uncertainty in a field. The results are used for dynamic model initialization and will be applied for forecasting under uncertainty. To determine the dynamic numerical model initial hydrocarbon saturation, the saturation-height model (SHM) needs to be conditioned to the petrophysically interpreted logs. There were 2,500 geological realizations generated to cover the interpreted ranges of porosity, permeability, and saturations for 15 wells. For the SHM, 12 parameters and their ranges were introduced. Latin hypercube sampling was used to generate a training set for ML models using the random forest algorithm. The trained ML models were conditioned to the petrophysical log-derived saturation data. To ensure a fieldwide consistency of the dynamic numerical models, only parameter combinations honoring the interpreted saturation range for all wells were selected. The presented method allows for consistent initialization and for rejection of parameters that do not fit the observed data. In our case study, the most-significant observation concerns the posterior parameter-distribution ranges, which are narrowed down dramatically, such as the free-water-level (FWL) range, which is reduced from 645–670 m subsea level (mSS) to 656–668 mSS. Furthermore, the SHM parameters are proved independent; thus, the resulting posterior parameter ranges for the SHM can be used for conditioning production data to models and subsequent hydrocarbon-production forecasting. Additional observations can be made from the ML results, such as the correlation between wells; this allows for interpreting groups of wells that have a similar behavior, favor the same combinations, and potentially belong to the same compartment.
Read full abstract