The process of well-log correlation requires significant time and expertise from the interpreter, is often subjective and can be a bottleneck to many subsurface characterization workflows. Algorithmic approaches to well-to-well correlation suffer from the inherent heterogeneity of geophysical measurements in the wellbore, both from a geologic and data-quality perspective. We demonstrate a rigorous and repeatable method for well-log correlation by deploying a correlation tool that leverages a machine learning model for pattern matching between well logs and programmed stratigraphic correlation techniques. A supervised-learning approach was used to train a novel deep convolutional neural network (CNN) architecture using over five million data samples, which were derived from thousands of well logs and expert interpreted correlations. To ensure that a robust pattern-matching model was trained, well logs from several US onshore basins with various tectonic regimes and environments of deposition were used to construct training and validation datasets. The result is a universal model for pattern matching of wireline measurements that can incorporate multiple geophysical-log signals as input data and can be deployed at scale without the need for retraining. Overall, the pattern-matching model was able to achieve a level of accuracy of 96.6% and classification area-under-the curve (AUC) of 0.954 on a separate validation dataset. The universal deep CNN is one component of the correlation tool. Algorithmic three-dimensional search logic was constructed around the deep CNN model which determines the optimal correlation and marker propagation pathway. Rules-based criteria have also been applied to the model output ensuring conformance to stratigraphic principles including preserving stratigraphic order and honoring present-day structural trends. We present several examples to highlight the strengths and weaknesses of this machine-learning-based approach to well-log correlation which can be used to efficiently generate high-density datasets for regional exploration, development mapping and reservoir characterization exercises.

This content is only available via PDF.
You can access this article if you purchase or spend a download.