Abstract Shear‑wave velocity (Vs) is vital for seismic reservoir characterization, AVO/AVA analysis, rock‑physics modeling and geomechanics, yet complete Vs logs are often unavailable. In this study, we present a comprehensive comparison of 36 methods to estimate Vs. The methods used for this comparison include empirical relations, linear and regularized regressors, power‑ and polynomial‑based transforms, support‑vector, tree and gradient‑boosting algorithms, as well as ensemble frameworks that we applied to data from a clastic-dominated well in Australia’s Cooper–Eromanga Basin. This workflow assumes minimal measured Vs coverage (ultra-sparse to minimally sparse), aiming to reconstruct missing intervals rather than generate complete logs without supervision. To simulate real life operational conditions, Vs values were deliberately removed from randomly selected intervals (training subset) while a depth‑balanced quality‑control (QC) subset retaining measured Vs was reserved for validation. All models were run with their default hyper‑parameters to reflect “out‑of‑the‑box” performance and were ranked with multiple error metrics: coefficient of determination (R 2 ), mean absolute error (MAE), root mean square error (RMSE), and mean square error (MSE), along with depth-dependent residual diagnostics. Extra Trees — an ensemble method using many decision trees where split thresholds are chosen randomly rather than optimally — yielded the highest accuracy in this well (R 2 = 0.921; MAE = 82 m/s) for Dandy-001, but applying the same top‑ten models to two different wells (Casimir‑001 and Bagheera East‑001) revealed markedly different model rankings among the three clasticdominated wells (in the same basin). These differences highlight variability in method workability across lithological and data‑quality scenarios. Castagna (Geophysics 50(4):571-581, 1985) and simple interpolation anchored the lower‑bound performance (R 2 ≈ 0.87 and 0.63, respectively), while several default machine‑learning configurations (e.g., Gaussian‑Process, SGD) were unstable. Overall, although no single technique dominates in every case, a best-fitting method that most closely reflects the measured data can be identified for each dataset. However, this requires performing a thorough, scenario-specific QC to evaluate potential bias, variance, and situational reliability.
Karaca et al. (Mon,) studied this question.