Identification of Vuggy Zones in Carbonate Reservoirs From Wireline Logs Using Machine Learning Techniques

Abstract

Vugs are irregular cavities inside rocks, formed by dissolution processes that may result in higher permeability zones. Vugs are identified through the analysis of image logs and cores. These datasets are generally sparse because they are expensive to acquire. Vugs are not readily identified with the common triple combo logging suite. We seek to develop decision rules to correlate triple combo logs with the presence or absence of vuggy zones as determined from image logs and cores.

Image Logs from six wells in the Appalachian Basin were analyzed for the presence of vugs and translated into a binary vuggy zone indicator log. Multiple machine learning models were trained to predict the indicator based on logged values for gamma ray, neutron porosity, photo electric, and bulk density.

Performance was assessed using well-level cross-validation. Each well's data was held out of the dataset, a model was trained using data from the other five wells, and the model was used to predict the vuggy zone indicator for the held-out well. The support vector machine (SVM) model was the top performer with a 78% correct identification rate. The proportion of entries in the held-out wells that were correctly predicted as either Vug or No-Vug ranged from 71% to 91%.

Note that many techniques, including SVM, result in predictive models that do not have a simple closed-form representation. A recursive partitioning tree analysis is also presented, which correlates the logs and vuggy zone indicator in a way that is easier to interpret and visualize.

AAPG Datapages/Search and Discovery Article #90218 © 2015 Eastern Section Meeting, Indianapolis, Indiana, September 20-22, 2015