Sustained casing pressure (SCP), prevention of cross-flow between intervals of formation, and reliable cement zonal isolation depend on attaining integrity of the well over time and eliminating cross-flow between intervals. The standard cement evaluation devices (CET) like ultrasonic, acoustic logs have high fidelity but tend to be few, costly to purchase or become inaccessible in high angle, deepwater or slimhole completions. The paper builds a machine-learning (ML) model to forecast the quality of cement zonal isolation with mixed data; rig-site telemetry, mud logging parameters, pumping schedules, and a few cement evaluation logs. The gradient boosting and sequence learning networks were combined in a hybrid and trained on multi-well data sets with full and partial cement logs. The findings indicate that the ML workflow predicts the quality of isolation with an accuracy of 92% and a mean absolute error of 0.07 and an AUC of 0.89, and on the other hand, the isolated empirical correlations and physics-based rules perform worse. The model has been able to generalize on wells having no cement logs and has allowed the proactive definition of risky intervals and the use of optimal cementing plans. The results indicate that combining various telemetry signals with scarce evaluation data can be highly useful in predicting cement quality and lessening the utilization of expensive wireline cement logs.
Okoroma et al. (Thu,) studied this question.