Identification of thermally mature total organic carbon-rich layers in shale formations using an effective machine-learning approach

We have developed a support vector machine (SVM) method that relies on core-measured data as well as gamma-ray, deep resistivity, sonic, and density wireline well-log data in identifying thermally mature total organic carbon (TOC)-rich layers at depth intervals with missing geochemical data in unconventional resource plays. We first test the SVM method using the Duvernay Shale Formation data. The SVM method successfully classifies the TOC data set into TOC-rich and TOC-poor classes and the [Formula: see text] data set into thermally mature and thermally immature classes when the optimal features are selected. To further test the SVM approach, we generate depth-separated training and test data sets from a well in the Duvernay Shale Formation and successfully use the approach to identify thermally mature TOC-rich intervals. We also examine the successful cross basin application of the SVM approach in predicting TOC using data from the Barnett and Duvernay Shale Formations as the training and test data sets, respectively.