Improved prediction of shale gas productivity in the Marcellus shale using geostatistically generated well-log data and ensemble machine learning

Abstract

This study proposes the application of geostatistically generated well-log data to predict well productivity in Marcellus shale reservoirs using ensemble machine learning (ESM). ESM was voting weighted by four individual machine learning methods: random forest (RF), gradient boosting, extreme gradient boosting, and light gradient boosting. The ESM model was trained based on 9921 wells that are qualified as training data. The well-log data had to be predicted using ordinary kriging for all 9921 wells because there were only 18 wells with well-log data for gamma-ray, resistivity, and density. Three cases were compared to analyze the influence of well-log data on machine learning performance. Case 1 comprised only basic well information, such as well location and fracturing-related parameters. Case 2 included the data from Case 1, and the well-log data of the closest well among the 18 wells were assigned to 9921 wells. In Case 3, to data of Case 1, the kriging-generated well-log data were added. In conclusion, Case 3 yielded the best performance among the machine learning algorithms, the highest coefficients of determination, and the lowest mean square errors. According to the RF models, the feature importance of the well-log data in Case 3 was double that in Case 2. The geostatistical method allowed us to make good use of the well-log data despite the restricted number of wells with the well-log compared to the enormous scale of the Marcellus shale. Moreover, basic well information is affordable compared to costly data such as permeability and porosity, even before the production start point. Thus, it can be applied to guide decision-makers when determining a drilling target. Furthermore, owing to its versatility in data affordability, the proposed method could be expanded to other shale basins, such as Barnett, Eagle Ford, and Scoop.