Performing supervised methods to estimate facies as a quick prediction for carbonate facies in the Puttygut Field, Michigan Basin.

Abstract:

Manual interpretation of well log data can be time-consuming and sometimes subjective. This can be especially so when relating core-derived facies to log-derived electrofacies. Data analytics and machine learning are currently being explored as tools to quickly assess vast amounts of data. In this study, we use common supervised machine learning algorithms to accelerate Manual interpretation with an application to carbonate facies identification from well logs in the Puttygut field, Michigan Basin. The data available for this study includes a detailed core characterization describing ten carbonate facies. Well logs for several wells in the area, including the cored wells, were also available.

Quality control was performed on gamma-ray, porosity, and permeability well logs before the training stage by removing abnormal values. Normalization was applied to the remaining data. We then split the data into a training and a test set. Although ten facies were described on the core, we chose 6-class and 3-class estimation because of the presence of several underrepresented classes. Finally, using a host of supervised learning algorithms such as K-nearest neighbors, support vector machines and random forests, we constructed a classifier to identify facies from the well logs.

We compare and contrast the performance of all algorithms on this dataset. Although all models might misclassify portions of a reviewed well, K-nearest neighbors and support vector machines algorithms show better results than random forests identifying the real core facies.

We demonstrate that permeability data aids in the accuracy of the supervised classifier. Overall, classification methods such as K-nearest neighbors, support vector machines are good enough to allow a rough estimate for quick carbonate facies identification, which might be further adjusted with manual interpretation. The approach described here shows promising results for identification of facies from well logs, but the input of a geologist is always an essential part of the process.