Abstract

This work introduces an ensemble-based machine learning framework for mud loss prediction in carbonate reservoirs, addressing one of the major operational and economic challenges in drilling. The framework combines tree-based models with Bayesian hyperparameter optimization and conformal prediction to improve both accuracy and uncertainty quantification respectively. Among different approaches, the optimized Random Forest model achieved the highest prediction accuracy (R² ≈ 0.93, RMSE ≈ 23.6), reliably capturing the overall trend while showing some limitations in extreme mud loss cases. To enhance transparency of the model architecture, feature importance analysis (tree-based, permutation-based methods, and SHAP) and residual analysis were employed, bridging the gap between black-box prediction and explainable insights essential for operational decision-making. Uncertainty values through conformal prediction of Random Forest model provided reliable prediction intervals when comparing with Quartile regressor and Bayesian neural network operator, enabling improved prediction confidence in operational planning. Overall, the proposed framework demonstrates the potential of machine learning with uncertainty quantification as a robust tool for forecasting mud loss and mitigating lost circulation risk in drilling operations