Predicting the severity of drilling lost circulation using just-in-time learning

Lost circulation is a major drilling complication that can increase nonproductive time, operational costs, and well-control risks. Conventional global machine-learning models may provide limited performance when geological and operational conditions vary substantially among wells. This study investigates whether query-specific local sample reconstruction using Just-in-Time Learning can improve lost-circulation-severity prediction. A field dataset containing 58,525 drilling records from 17 wells and four severity categories was analyzed. The wells were divided into training and testing subsets by well identity to prevent records from the same well from appearing in both subsets. Random Forest-based Recursive Feature Elimination was used to identify the principal input variables. For each testing sample, a hybrid similarity measure combining cosine similarity and Euclidean distance was used to retrieve relevant historical records and construct local Random Forest, XGBoost, and Deep Belief Network models. The Just-in-Time Learning-based models consistently achieved higher weighted F1-scores and multi-class Matthews correlation coefficients than their corresponding global models. These results indicate that query-specific local sample reconstruction can reduce the adverse influence of between-well data heterogeneity and improve severity classification under the adopted well-wise holdout setting. The novelty of this study lies in introducing query-dependent local model construction into multi-class lost-circulation-severity prediction and evaluating its applicability across three different learning algorithms.