Relative performance of support vector machine, decision trees, and random forest classifiers for predicting production success in US unconventional shale plays

Unconventional shale reservoirs have revolutionized the energy industry. However, the prediction of production based on reservoir geology characterization has largely focused on sweet spot definition rather than on over-arching production trends across multiple plays. This study uses machine learning (ML) techniques to analyze the relationships between well log data and production success within seven North American shale plays. Three ML algorithms were evaluated: stochastic gradient descent kernel trained support-vector machine (SGD-SVM), decision tree (DT), and random forest (RF) classifier. Accuracy of predictions using the SGD-SVM and DT classifiers did not exceed 55%. A fine-tuned RF classifier is the most successful method at predicting well success based on normalized initial production, with an accuracy of 97%. To achieve this result, the RF is trained on the following input features: average play thickness, pore pressure, TVD, and resource concentration. The main factors impacting performance of our algorithm when trying to predict success in unconventional plays are previous understanding of heterogeneities in individual formations, and consistency of data availability across multiple wells. Despite challenges, ML and the RF method in particular show promising applications in the unconventional petroleum industry as a means to streamline production and data collection.