--> --> Machine learning facies identifications applied to the Late Devonian Duvernay Formation, Western Canadian Sedimentary Basin

AAPG Southwest Section Annual Convention

Datapages, Inc.Print this page

Machine learning facies identifications applied to the Late Devonian Duvernay Formation, Western Canadian Sedimentary Basin

Abstract

The Duvernay Formation is a major petroleum source rock in the Western Canadian Sedimentary Basin. With recent advances in drilling and completions technology, the Duvernay has transformed into a target for unconventional exploration and production. A refined understanding of the geologic controls on producibility will minimize exploration and production risks. The purpose of this study is to identify reservoir and non- reservoir facies associations (FA) from well logs using a machine learning model and compare the results with those from a more traditional semi-quantitative approach. Detailed descriptions of sedimentological characteristics from 42 wells were used to make FA. Two methods are used to identify FA from petrophysical data. The first method is a semi-quantitative approach that utilized bivariate statistical analysis to define well log cutoffs for FA classification. The second method used a decision-tree based machine learning (ML) model from the facies-classification competition proposed by Hall (2016) to identify FA. The ML model classified Duvernay FA with acceptable (>70%) accuracy with the best-performing identification achieving an accuracy of 72%. Most misclassifications were due to biases in the training data such as misclassification of FA with the fewest number of samples. This could be addressed using additional supervised learning and statistical methods. The semi- quantitative method was 76% accurate, and misclassifications were due to sedimentological similarities within basinal FA. This resulted in similar well-log cutoffs yielding nonunique choices. ML algorithms are superior to semi-quantitative approaches in their computational efficiency, reduction in user intervention and bias, and reproducibility; however, the semi-quantitative approach was more accurate. Additional conditioning of the petrophysical data is expected to improve the accuracy of the ML model such that it rivals or even eclipses the semi- quantitative approach. Hall, B., 2016, Facies classification using machine learning: The Leading Edge, v. 35, p. 906–909, doi:10.1190/tle35100906.1.