Machine learning-based UCS prediction from wireline logs with empirical pseudo-labeling: a case study of cross-well transferability and one-shot calibration in Western Australian basins

Unconfined compressive strength (UCS) is essential for wellbore stability and sand-production assessment, but laboratory measurements are sparse. This study develops a machine-learning framework for UCS prediction from wireline logs under data scarcity. As the only well with a formation-specific, laboratory-calibrated UCS relationship, Waitsia-3 provided the UCS–density-derived-porosity equation adopted without refitting; it reproduced three calibration measurements within 4.3–7.7% and generated 3,171 pseudo-UCS labels. Random forest (RF), support vector regression, and artificial neural network models were tuned by purged spatial cross-validation and compared under random and contiguous depth-based validation. The selected RF achieved a mean absolute error (MAE) of 7.80 MPa and an R 2 of 0.792 under random splitting, versus 21.06 MPa and 0.387 on the depth-based holdout, indicating optimistic performance when depth-adjacent rows occur in both partitions. SHapley Additive exPlanations (SHAP) attributed 70.3% of RF importance to compressional slowness and log-transformed deep resistivity, reflecting pore structure, fluid conditions, and rock-frame stiffness. Uncalibrated transfer to four wells produced MAEs of 43.18–72.70 MPa. A SHAP-weighted Kullback–Leibler score combining input divergence with RF reliance showed a strong but non-significant association with well-level mean absolute percentage error ( r = 0.913, p = 0.087, n = 4); simpler metrics performed comparably. Leave-one-out one-shot calibration reduced Harvey-3 MAE from 53.65 to 29.67 MPa but degraded the other wells. Novelty lies in integrating empirical pseudo-labeling, spatial validation, pre-application shift screening, and conditional calibration within one cross-well transfer workflow. The framework supports first-pass UCS profiling and targeted mechanical testing rather than stand-alone operational decisions.