Contrasting Machine Learning Regression Algorithms Used for the Estimation of Permeability from Well Log Data
Arabian Journal of Geosciences (2021) 14: 2070; https://doi.org/10.1007/s12517-021-08390-8
Posted: 10 Dec 2021
Date Written: September 26, 2021
The oil and gas industry has slowly shifted its focus to a more data science-driven interpretation approach from the last decade. The petrophysical data analysis using advanced statistical and machine learning methods has been widely accepted due to reducing uncertainties and predicting more accurate data trends than conventional methods. The same approach is reflected in permeability estimation, where regression models are used during the well log data interpretation. This becomes necessary because amassing permeability data by other means is economically unfavorable and time-consuming. The exploration and production giants like Equinor employ variations on elementary methods like simple linear regression (SLR) to establish regression models with accuracies up to 0.98 R2 scores. However, in recent years, various advanced machine learning algorithms have been developed that could be utilized in oil and gas data analysis and modeling with greater accuracy, coupled with thorough data cleaning and outlier removal. The current study demonstrates the application of modern machine learning algorithms to analyze drilling data from two wells of Equinor’s Volve field located at the Norwegian Continental Shelf (NCS) and compare the outcomes with conventional analysis methods. The core data analysis of wells F-15/9-19A and F-15 B&BT2 is used, and a relationship is established between permeability and data obtained from wireline logging (prominently used variables) where a relationship can be observed consistently are porosity (PHIF) and shale volume (VSH). The goodness of fit of the correlations is thus obtained by calculating the 'R2 score,' which gives an estimate about the accuracy of the regression models. Current study aims to compare the efficiency of four major regression algorithms, SLR, Lasso regression (LR), multiple linear regression (MLR), and support vector regression (SVR), in the estimation of Klinkenberg core corrected permeability (KLOGH) using porosity and shale volume. After performing a thorough cleaning and outlier filtering from the dataset, it was found that over 1000 iterations, SVR peaked when it came to R2 score value (SVR, 0.88), while MLR performed the best on average (MLR, 0.77).
Keywords: keywords machine learning, Klinkenberg core corrected permeability, regression, core data, R2 score
Suggested Citation: Suggested Citation