ABSTRACTIn oil and gas exploration, it is vital to acquire information about the bottom hole conditions. This is done in the field using wireline logging. The sonic log is one of the most prolific logs as it assists in porosity determination, cement evaluation and identification of lithology and gas‐bearing intervals. However, sonic logging tools are not always a part of the wireline logging arrangement. Still, there are sections where the logging data are missing, and in some cases, these are dependent upon old tools. The tool is incapable of recording shear wave transit times. This study explores the possibility of substituting the sonic log using machine learning and artificial intelligence techniques. These techniques can also predict the sonic log data in sections where these are missing or unreliable. Artificial neural networks, decision tree regression, random forest regression, support vector regression and extreme gradient boosting are the most popular tools available at our disposal for making these estimations. This study has compared these different techniques for their effectiveness and accuracy in making sonic transit time predictions based on other available well logs. The obtained results suggest that despite all the attention on artificial neural networks, eXtreme gradient boosting and the random forest regression outperform it for the given purpose. In the case of missing shear transit time data, random forest regression made predictions with a root mean squared error of 1.03 × 10–4 and a root mean squared error of 0.97 while eXtreme gradient boosting regression did so with a root mean squared error of 1.36 × 10–4 and the same regression coefficient (0.97). When no sonic data were available, random forest regression estimated shear transit time with a root mean squared error of 6.41×10–4 and a regression coefficient of 0.95, and compressional transit time with a root mean squared error of 9.06×10–5 and a root mean squared error of 0.94. The root mean squared error of shear transit time and compressional transit time predictions made using eXtreme gradient boosting were found to be the same as those of random forest regression. The root mean squared error, however, was observed to be slightly less for the compressional transit time predictions and somewhat more for the shear transit time predictions. As data analysis, in general, is a better method for estimation than the use of empirical correlations, these machine learning‐based predictions can serve as powerful tools in the oil and gas exploration industry.
Read full abstract