A reproducible public benchmark and uncertainty-guided review framework for measurement-while-drilling analytics

Measurement while drilling (MWD) signals entangle ground response with rig control, tooling, circulation, and operator actions. This study establishes a public-data protocol for state recognition, depth-ordered rate-of-penetration (ROP) prediction, and uncertainty-based review prioritisation. Two complementary tasks are evaluated: blast-round-level classification using 48 released statistics from 15 Norwegian hard-rock tunnels, and depth-ordered sequential prediction using 258 observations from one borehole. On the author-provided classification split, balanced LightGBM reached 0.873 accuracy, 0.850 balanced accuracy, and 0.848 Macro-F1, compared with 0.048 Macro-F1 for the majority baseline and 0.083 for the label-permutation control. Under leave-one-tunnel-out (LOTO) evaluation, the model exceeded the training-majority baseline in 12 of 15 tunnels, with mean Macro-F1 0.387 and a tunnel-cluster bootstrap 95% CI of 0.264–0.516. Leakage-free nested rolling-origin evaluation identified temporal persistence as a strong ROP reference: engineered and raw LightGBM obtained RMSE 22.141 and 21.378, respectively, while one-step persistence obtained 11.556. A confidence threshold selected only on training validation data retained 89.3% of the untouched author test set at 90.5% accuracy. Split-conformal sets achieved 90.4% marginal coverage at an 11.8% review rate under the author split, and the cross-tunnel diagnostic conservatively routed all shifted predictions to review. The results provide a reproducible benchmark, a leakage-free comparison framework, and an uncertainty-aware pathway for multi-project drilled-shaft validation.