Using extreme gradient boosting with Optuna hyperparameter tuning for efficient lost circulation prediction

Lost circulation (LC) remains a significant challenge in drilling operations, leading to increased costs, non-productive time, and potential well integrity issues. This study focuses on predicting lost circulation during drilling operations using the XGBoost machine learning algorithm, coupled with hyperparameter tuning RPM via the Optuna framework. The dataset includes drilling parameters, mud properties, and geological features, incorporating noisy and outlier-prone real-world data. Three Optuna samplers—TPE (Tree-structured Parzen Estimator), CmaEs (Covariance Matrix Adaptation Evolution Strategy), and NSGAIII (Non-dominated Sorting Genetic Algorithm)—were evaluated for their effectiveness in optimizing the model. The TPE sampler achieved the highest lost circulation prediction performance, yielding a coefficient of determination R² of 82.27%, an adjusted R² of 81.40%, and a root mean squared error (RMSE) of 4.134. Feature importance analysis highlighted measured depth (MD), weight on bit (WOB), rotations per minute (RPM), and rate of penetration (ROP) as the primary predictors of lost circulation, underscoring the critical influence of geological and operational parameters. Each of these input features plays a significant role in predicting lost circulation: MD is crucial as it correlates with geological formations prone to mud loss; WOB influences the stress exerted on the formation, which can lead to fractures; RPM impacts the mechanical action of the drill string, contributing to wellbore instability; ROP reflects the drilling efficiency and the interaction between the bit and subsurface layers, which are critical for identifying lost-circulation zones. These features collectively enable the model to capture complex relationships and enhance its predictive performance. The findings provide actionable insights for improving drilling techniques, mitigating lost circulation risks, and enhancing operational efficiency. By retaining noisy and outlier data, the study aligns the predictive model with real-world complexities, demonstrating its robustness and practical relevance.