Breakthrough pressure (BP) is a key parameter for evaluating the sealing capacity of shale caprocks, whereas direct laboratory measurements are time-consuming and costly, limiting their use in continuous regional assessment. This study develops a conventional-log-based workflow for BP prediction in the Quan-4 Member (K1q4) caprock underlying the Qingshankou shale oil interval in the Gulong Sag, Songliao Basin. Four routinely available logs—gamma ray (GR), acoustic interval transit time (AC), bulk density (DEN), and compensated neutron log (CNL)—were integrated with core-measured BP data. A GR-AC multiple-regression baseline and five machine-learning algorithms, including stochastic gradient descent (SGD), extremely randomized trees (ERT), Random Forest (RF), extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost), were compared under a unified workflow. The training set was used for normalization, model fitting, grid search, and internal five-fold cross-validation, whereas the held-out test set and external prediction wells were kept separate for performance evaluation. The results show that BP generally increases with GR and DEN and decreases with AC and CNL, indicating that clay content, compaction, and pore connectivity jointly control the logging response of caprock sealing capacity. Among the evaluated models, AdaBoost achieved the best overall performance, with RMSE, MAE, and R2 values of 1.33 MPa, 0.97 MPa, and 0.89 on the held-out test set, and 1.19 MPa, 0.92 MPa, and 0.94 in external prediction wells. Train–test diagnostics, learning curves, and SHAP analysis indicate stable performance and physically plausible feature contributions within the present dataset. The proposed workflow can therefore provide a practical supplement to laboratory BP measurements for caprock evaluation in the study area, although broader application still requires well-level independent validation and explicit prediction-uncertainty quantification.
Wu et al. (Fri,) studied this question.