Abstract
Wellbore wall collapse under complex geological conditions presents a significant challenge in oil well drilling, increasing repair costs and operational downtime. This study proposes a machine learning–based approach to predict wellbore stability, developing a robust model utilizing geomechanical rock properties, drilling parameters, and geological data, with a binary target variable (1 for stable, 0 for unstable wells). A dataset of 5,000 records, including 200 collapse cases, was preprocessed – removing duplicates and missing values, handling outliers, normalizing numerical features, and encoding categorical variables – before being split into 80% training and 20% testing subsets. Gradient boosting (XGBoost) and random forest (Scikit-learn) were applied for binary classification, with hyperparameters optimized via GridSearchCV gradient boosting outperformed random forest, achieving 93% accuracy, 89% recall, and 91% F1-score, compared to 91%, 87%, and 89%, respectively. The study recommends integrating the gradient boosting model into a real-time monitoring system, analyzing sensor data every 10 minutes to provide recommendations (e.g. increasing mud density or reducing drilling speed), potentially reducing collapses by 25%, cutting repair costs and downtime by 10%, and enhancing drilling efficiency. This research underscores machine learning potential to improve wellbore stability prediction, delivering significant economic and operational benefits to the oil and gas industry.