Estimating electrical resistivity from logging data for oil wells using machine learning

Abstract

Formation resistivity is crucial for calculating water saturation, which, in turn, is used to estimate the stock-tank oil initially in place. However, obtaining a complete resistivity log can be challenging due to high costs, equipment failure, or data loss. To overcome this issue, this study introduces novel machine learning models that can be used to predict the electrical resistivity of oil wells, using conventional well logs. The analysis utilized gamma-ray (GR), delta time compressional logs (DTC), sonic shear log (DSTM), neutron porosity, and bulk density. The study utilized a dataset of 3529 logging data points from horizontal oil carbonate wells which were used to develop different machine learning models using random forest (RF) and decision tree (DT) algorithms. The obtained results showed that both models can predict electrical resistivity with high accuracy, over 0.94 for training and testing data. Comparing the models based on accuracy and consistency revealed that the RF model had a slight advantage over the DT model. Based on the data analysis, it was found that the formation resistivity is more significantly impacted by GR logs compared to DTC logs. These new ML models offer a low-cost and practical alternative to estimate well resistivity in oil wells, providing valuable information for geophysical and geological interpretation.

Oil and Gas Production Forecasting Using Decision Trees, Random Forst, and XGBoost

“…It takes the average of results from multiple DTR to arrive at the final forecast. The prediction outcome is determined by computing the mean of the outputs generated by each tree [28]. It was first introduced in 2001 by University of California, Berkeley professor Leo Breiman.…”

Evaluation of the Possibilities Validation of Interval Velocity Models Using Non-Seismic Data and Its Impact on Geological Interpretation of PreSDM Results

“…In practice, the determination of correlations between resistivity and seismic wave velocity is possible for genetically defined, relatively homogenous rock formations but, unfortunately, it often is rather ambiguous. Modern statistical methods are quite helpful [4] together with machine learning methodologies [10][11][12][13]. In the project reported below, the authors applied the modern method of integrated inversion of independent geophysical data, which is still under development.…”