USING MACHINE-LEARNING FOR DEPOSITIONAL FACIES PREDICTION IN A COMPLEX CARBONATE RESERVOIR

摘要

This paper describes an innovative new workflow to identify depositional facies in a complex carbonate reservoir from logs, using multiple clustering methods combined in machine learning. The workflow was developed using 250m of described core plus 26 different logs covering a combined total of 5km oil-bearing heterogeneous lacustrine pre-salt carbonate reservoir section from five wells, Deepwater Brazil.The challenge to identify the original depositional facies in carbonate reservoirs is often unresolved, as the same depositional facies can be affected by different diagenetic processes over time, resulting in Reservoir Rock Types (RRTs) with different petrophysical properties. These pre-salt carbonates are suitable for developing this technique, as depositional facies are still visually distinct in core, independent of present-day RRT. The inter-relationships between depositional facies and petrophysical logs are not straight-forward and conventional training or supervised methods are usually not successful for depositional facies prediction. Using conventional techniques (e.g. Support Vector Machines, Random Forest, Artificial Neural Networks, Naïve Bayes, k-nearest neighbor, etc.) we found that depositional facies cannot be predicted with more than a 40% accuracy in the pre-salt carbonate data set.The new workflow developed here for depositional facies prediction combines data preparation, supervised and unsupervised learning methods and probabilistic algorithms. The workflow consists of four steps:1)Data preparation (normalization, depth-matching,facies border removal, missing data and outliershandling);2)Simultaneous application of several supervisedlearning methods using the full suite of available logs toidentify a) which training method clusters best againstcore depositional facies and b) to define the optimumselected log combination;3)Application of an unsupervised clustering methodequivalent to the best supervised method (2a) on theselected best set of input logs (2b) to define petrophysicalclusters;4)Comparison of the petrophysical clusters with the coredepositional facies, to find rules on frequency,stratigraphic distributions and likely patterns and toassign the clusters to a particular depositional facies.Two models for depositional facies prediction were run: Model 1 (3 main depositional facies) and Model 2 (5 depositional facies). In both cases the Naïve Bayes classifier scored best for the supervised clustering step and identified the best logs to select for Step 2. In Step 3 Naïve Bayes algorithm run in unsupervised mode identified seventeen petrophysical clusters.After Step 4, assigning petrophysical clusters to the depositional facies, performance tests for predicting the depositional facies were increased by 35% (Model 1) and 15% (Model 2) to 68% and 55% respectively.