Well-log depth alignment: a critical review and field-based comparison of classical and machine-learning methods

Abstract Accurate well-log depth matching is important because even small depth errors can affect fluid-contact interpretation, net-pay estimates, well-to-well correlation, and reservoir modelling. In practice, electric wireline logging (EWL) and logging-while-drilling (LWD) logs often differ in depth because of cable and drillstring stretch, friction, thermal effects, borehole conditions, and drilling operations. This paper critically reviews well-log depth-alignment methods, including cross-correlation, dynamic time warping (DTW), continuous regularised optimisation, feature-based approaches, and machine-learning methods such as supervised, unsupervised, self-supervised, and reinforcement-learning models. Beyond summarising the literature, the paper evaluates method families against physical and operational requirements, including depth-order preservation, control of unrealistic local deformation, mapping smoothness, and separation of true depth error from tool-response mismatch. The review is supported by an illustrative field-data comparison on one EWL–LWD interval from the Norwegian Continental Shelf. Five representative methods—cross-correlation, band-constrained DTW, continuous regularised optimisation, supervised convolutional neural network (CNN) alignment, and unsupervised structural-boundary alignment—were evaluated using controlled synthetic depth perturbations imposed on the LWD data. The experiments included a common 10 m shift, shifts from 2.5 to 10 m, window-length sensitivity, and a smoothly depth-varying error over approximately 1.2 km of field data. For the common 10 m experiment, DTW produced the highest mean alignment-channel correlation gain (0.8901) and the lowest shift mean absolute error (MAE) (1.0648 m), but also the largest maximum strain (3.50), indicating stronger local deformation. Continuous optimisation produced the highest cross-channel correlation gain (0.8108) with substantially lower maximum strain (1.0584), while the unsupervised method achieved a correlation gain of 0.8804 and a shift MAE of 1.2136 m with relatively low deformation. Across imposed shifts of 2.5–10 m, positional errors remained approximately metre scale, and in the depth-varying experiment all five methods achieved mean shift MAEs below 1 m. The results show that improved curve similarity alone is insufficient to establish reliable depth correction: positional accuracy, cross-channel consistency, and physical plausibility of the inferred mapping must be considered together. The main contribution of this review is therefore a unified critical assessment of classical and machine-learning alignment methods combined with a controlled field-data comparison using direct depth-error and mapping-quality metrics, providing practical guidance for method selection and future validation.