Abstract
It is necessary to use sonic logs in gathering rock mechanical properties which are used to estimate in-situ stresses which are useful for fracture design and reservoir characterization. However, the high cost of its acquisition deters companies from getting complete well measurements and in some cases, we may have incomplete sonic log data due to tool damage, poor calibration, and other reasons. In other cases, the well of interest might only have a triple combo suite to determine target depths and left up to engineering judgment what a hydraulic fracture might look like.
This study shows a field application of machine learning (ML) algorithms using triple combo logs to predict elastic properties and finally estimate in-situ stress profile in the Natural Buttes gas field. We used a complete set of triple combo and dipole sonic logs to train K-Nearest Neighbors (KNN) regression and Bi-directional convolutional long short-term memory (Bi-ConvLSTM) models. A data preparation and transformation technique were applied to reduce heteroscedasticity by manipulating data distribution. Then, prediction-performances were compared to each other, and a selected ML model went through a validation to verify it effectively captures existing patterns in the well log data. The selected ML model was finally deployed into a GUI interface which is helpful for users to create predictions, visualize and analyze the generated results.
Comparative analysis of ML models shows that the KNN mitigates overprediction observed in the Bi-ConvLSTM. The KNN reduces training time while the Bi-ConvLSTM is computationally expensive as it considers well log data as an image. We also quantify uncertainty of KNN predictions through a heuristic approach to provide uncertainty thresholds as decision boundary for in-situ stress estimation. For calibration verification, we present one data set showing the predicted sonic velocities vs actual measured velocities as well as the predicted vs actual elastic and in-situ stress profile. The GUI can deploy trained models, take test data for prediction, visualize triple combo logs, predicted sonic logs, calculated elastic properties and in-situ stresses with variances for analysis. We conducted fracture modeling using a commercial numerical simulation tool for both the measured and synthetic cases and found the SRV was similar to each other.
This study demonstrates that the use of machine learning combined with well-established data transformation, uncertainty estimation, and visualization can reduce time and cost in extracting and analyzing a valuable information. Based on literature review, this is the first time that such a workflow has been shown up to the in-situ stresses portion of the workflow. We finally, perform hydraulic fracture simulation on the synthetic in-situ stress profile case and show equivalent results to the calibrated in-situ stress profile case.
