A DVANCED M ACHINE L EARNING T OOLS FOR O IL W ELL L OG A NALYSIS AND I NTERPRETATION

Abstract

The interpretation of well log data is crucial for accurate reservoir characterization. Despite their importance, well logs often suffer from a high rate of missing values, a recurring issue that undermines the reliability of key petrophysical estimates such as porosity, clay volume, and water saturation. This thesis proposes a machine learning workflow applied to a multi-well dataset from Sonatrach (over 25,000 measurements), aimed at harmonizing heterogeneous logs and systematically filling in data gaps. A comprehensive data analysis phase was conducted on the Sonatrach dataset, includ- ing the study of distributions, detection of missing value patterns, and evaluation of multi- variate relationships, essential steps to support meaningful imputations. Then, several im- putation methods were compared, ranging from simple statistical techniques to advanced machine learning and deep learning models. Each approach was evaluated both in terms of imputation accuracy and its impact on the prediction of petrophysical parameters. Random forests and WGAIN stand out, delivering superior performance over traditional methods. A major contribution of this work lies in the integration of sequential context: the use of sliding-window statistics along depth enables better capture of local geological trends and enhances model stability, especially for WGAIN. Additionally, expanding the training set from three to five wells further increases robustness when generalizing to unseen wells. This work provides Sonatrach with a robust and interpretable framework that com- bines geological expertise with artificial intelligence, offering strong potential for broader application across the oil and gas industry.