English abstract
The sonic log is an essential petrophysical log and it is used in many petroleum
applications (seismic and geomechanically applications). Since sonic logs are hard to
acquire and expensive to run in all boreholes, oil companies conduct them in a few
wells. Thus, several petrophysical and geophysical workflows incorporate sonic log
synthetization. Empirical correlations or statistical methods are conventional methods to
estimate the sonic value. Despite the advantages of simple correlations for quick
estimations, these traditional approaches are not as powerful as modern-day machine
learning techniques.
This study presents combining different Machine learning and Feature Selection
algorithms that were applied to predict highly accurate synthetic sonic logs from basic
petrophysical logs. Machine learning algorithms have been implemented for this
purpose, including Multi-Regression (ML), Artificial Neural Network (ANN), Support
Vector Machine (SVM), and Random Forest (RF). This study is implemented with data
from seven wells in North Oman's tight gas sandstone. To highlight which of the logs
will be good predictors of sonic logs, feature selection approaches were carried out.
Approximately 70% of the data was used for model training, while the rest was used for
testing and validation. Analyzing the logs of the measured wells serves as a means of
determining the efficiency of the built models. This implies that compared to other
algorithms, the Hybrid Random Forest algorithm with a backward elimination feature
selection approach can be an efficient way of predicting the sonic logs and yielding a
very high correlation coefficient of more than 0.95 %.
For geomechanical applications, this method will provide enhanced sonic prediction and
is especially important for prospective wells where logs of sonic suits do not need to be
acquired which ultimately leads to effective cost-effective solution.
