A Data-Driven Model for Bound Fluid Prediction in Carbonate Reservoirs

Abstract

AbstractThe objective of this paper is to develop a machine learning workflow in order to replicate the Nuclear Magnetic Resonance (NMR) bound fluid volumes utilizing triple-combo data. The developed model is to overcome the logging operations challenges and to enhance the petrophysical formation evaluation in carbonate reservoirs and most importantly a proper completion design to avoid free water bearing zones (Dodge et al 1998).The presented data-driven model is based on artificial intelligence (AI) and data mining techniques. In order to select the required data input, many logging parameters were investigated and several iterations with different sets of parameters were attempted to identify those with the most impact on bound fluid volume prediction. Henceforth, the iteration with the lowest root mean square error (RMSE) is to be selected. The model was trained, calibrated and validated to accurately predict the NMR bound fluid volume. The main purpose of this study is to draw attention to the capabilities of the artificial intelligence techniques to provide logging responses with excellent match compared to the measured logs in the event that operational complexities or budget constraints arise.Several wells with excellent data quality were selected from the same field to build and test the RF (Random Forest) model. 56% of the wells were used for training and the remaining 44% were kept for blind tests. During the modeling process, several iterations were attempted to identify what was really driving the BFV prediction, the best model was chosen based on the lowest RMSE (Root Mean Square Error). The model statistics indicated the combination of resistivity, density, neutron, GR, produced the best model with lowest RMSE = 0.78 pu (porosity units). The model was later deployed to the remaining blind test wells, the results were of excellent accuracy when compared to the measured log data with error difference way below the acceptable industry standard of 1 pu. With accurate BFV prediction, free water zones can now be identified and avoided during completion. This methodology will help operators reduce risk, maximize production without straining their budget to acquire NMR logs in every horizontal development well, or even attempt formation tester samples unnecessarily.Implementing this new machine learning methodology will help in increasing the confidence in petrophysical analysis and reservoir management decisions, as well as optimizing the logging runs. This results in cost savings without compromising the quality of the formation evaluation, in addition to minimizing the carbon footprint associated with unnecessary operations.