A hybrid data-centric machine learning for lost circulation prediction using imbalanced drilling and geological datasets

Lost circulation remains one of the most critical operational challenges in drilling engineering, yet predictive modeling in real-field conditions is often constrained by severely imbalanced and limited datasets. While previous studies have primarily relied on single imbalance-handling techniques or large datasets, this study introduces a sequential, hybrid data-centric machine learning framework specifically designed for small, multiclass, and highly skewed drilling datasets. A dataset of 258 operational records from the Buzergan oil field in southern Iraq, including drilling parameters, mud properties, and geological indicators, were used to predict lost circulation severity. Unlike conventional approaches, this work systematically integrates synthetic sampling techniques, ensemble learning (bagging and stacking), and cost-sensitive class-weight optimization within a staged workflow, where each enhancement step builds upon the previous one. Importantly, class weights were computed from the original imbalanced distribution and transferred to synthetically augmented datasets, preserving the true imbalance structure while improving minority-class sensitivity. Hyperparameter tuning was conducted using Optuna with Tree-structured Parzen Estimator (TPE), and model evaluation emphasized the Matthews Correlation Coefficient (MCC) to ensure robust assessment under multiclass imbalance. The results demonstrate that the proposed hybrid framework significantly outperforms standalone oversampling or ensemble strategies. The CatBoost model combined with the Synthetic Minority Oversampling Technique (SMOTE) and optimized custom class weights achieved the highest classification performance (Accuracy: 0.8846; Weighted F1: 0.8651; MCC: 0.8236; weighted Area Under the Curve, AUC-W: 0.9154), while a stacking ensemble exhibited superior ranking capability (AUC-W: 0.934). The primary novelty of this study lies in (1) treating class weight as a tunable hyperparameter rather than a fixed correction factor, (2) sequentially combining synthetic data generation with imbalance-aware weighting, and (3) validating the framework under strict stratified cross-validation and clean hold-out testing to prevent data leakage. This structured hybrid methodology offers a template applicable to other small-sample, imbalanced classification problems in drilling; validating the trained models themselves on other fields remains future work.