Borehole Resistivity Measurement Modeling Using Machine-Learning Techniques
- Yankai Xu (Google) | Keli Sun (Schlumberger) | Hui Xie (Schlumberger) | Xiaoyan Zhong (Schlumberger) | Ettore Mirto (Schlumberger) | Yao Feng (Schlumberger) | Xiaobo Hong (Schlumberger)
- Document ID
- Society of Petrophysicists and Well-Log Analysts
- Publication Date
- December 2018
- Document Type
- Journal Paper
- 778 - 785
- 2018. Society of Petrophysicists & Well Log Analysts
- 5 in the last 30 days
- 178 since 2007
- Show more detail
Resistivity measurement while drilling is important for real-time drilling and measurement (D&M) business. It helps to identify variations of rock and is used more specifically to indicate reservoir carbonates and/or water zones. The measurements are also sensitive to range of factors, including borehole condition, mud resistivity, tool eccentricity, invasion and therefore it provides an opportunity to invert for these parameters. In fact, borehole condition, mud resistivity and invasion can be used as important indicators for drilling safety, and possibly used as inputs for well completion. Traditionally, an inversion of the parameters requires a tremendous number of forward-computing iterations, therefore, fast forward modeling is essential to the success of a real-time drilling application. Resistivity-modeling algorithms that accurately account for mud effects and tool eccentricity are currently rather slow. Speed can be addressed by using a look-up table of precomputed partial results for factors that affect resistivity measurements, such as formation properties, mud resistivity, borehole size and tool eccentricity. In the formation evaluation phase, the traditional inversion algorithm obtains a modeled measurement through table look-up and interpolation instead of running a forward-modeling process. However, interpolation accuracy is compromised by the fact that the look-up table grid has to be coarse in order to maintain a manageable size.
Fast forward modeling can be formulated as a function approximation problem. The most effective and popular method to solve function approximation problems is machine learning. There are many existing machine-learning techniques, including neural network, gradient-boost-regression tree, Bayesian regression, SVM regression, and Gaussian process regression. With the same look-up table grid, we can train a machine-learning model to predict modeling measurements. The trained model requires much less memory than the look-up table and has a much higher accuracy compared to interpolation. Performance-wise, the training phase of machine learning consumes substantial computation resources, but it is done offline ahead of the actual drilling execution. The prediction phase is as fast as table look-up and interpolation. Overall, the machine-learning- based method is very suitable for fast resistivity-measurement modeling.
|File Size||9 MB||Number of Pages||8|