Machine Learning for Facies Distribution of Large Carbonate Reservoir Models- A Case Study

Abstract

Multibillion barrels oil in-place carbonate reservoirs have their unique static and dynamic modelling challenges due to the nature of the reservoir with both vertical and lateral heterogeneities. The complex geological processes which took place both during and after the deposition, results in the heterogeneity, which are reflected in the reservoir characterization of these large-scale carbonate reservoirs. Capturing the geological facies variability in the reservoir description is thus one of the critical elements to ensure the model's validity, robustness, and forecasting ability. This case study exemplifies the use of a machine learning approach to tackle this subsurface complexity within a multidisciplinary integrated study to construct a field scale reservoir model for a large carbonate reservoir. The carbonate field has recently acquired additional core data in several newly drilled wells. These cores have been described by sedimentologists to define reservoir depositional facies and lithofacies. This geological description has been used by a machine learning algorithm to train the conventional triple combo logs to recognize the reservoir facies. The training of the facies definitions at the cored wells were also conditioned to the sequence stratigraphic correlation framework of the reservoir. Later these geological facies have been propagated using logs to more than 80 un-cored wells to provide facies predictions within a geological context. The result from the machine learning algorithm gives an excellent replication rate on the cored wells. It is also robust on the un-cored wells throughout the field. This robustness of the facies definitions has been verified using production / injection log survey (PLT / ILT), core CT-Scan and core descriptions. Firstly, for the cored wells, the production / injection zones identified by PLT surveys clearly correspond to the best reservoir facies. Secondly, in the un-cored wells, the best facies predicted by the machine learning algorithm correspond to the production / injection zones interpreted from the production and injection logging surveys (PLT survey) The predicted geological facies in both cored and un-cored wells and seismic inversion trends were used to condition the 3D distribution of the facies in the reservoir model. The use of machine learning for facies prediction has also helped to validate the underlying geological concept of an older good quality reservoir interval in certain areas of the field, which were not adequately sampled from the existing core data. In the future, the machine learning based reservoir models will be used to identify new infill locations where "best producing" facies are likely to be present.